xref: /freebsd/sys/powerpc/aim/moea64_native.c (revision c5fda9bac0325eb8c5b447717862d279006f318f)
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 int moea64_crop_tlbie;
138 
139 static __inline void
140 TLBIE(uint64_t vpn) {
141 #ifndef __powerpc64__
142 	register_t vpn_hi, vpn_lo;
143 	register_t msr;
144 	register_t scratch, intr;
145 #endif
146 
147 	static volatile u_int tlbie_lock = 0;
148 
149 	vpn <<= ADDR_PIDX_SHFT;
150 
151 	/* Hobo spinlock: we need stronger guarantees than mutexes provide */
152 	while (!atomic_cmpset_int(&tlbie_lock, 0, 1));
153 	isync(); /* Flush instruction queue once lock acquired */
154 
155 	if (moea64_crop_tlbie)
156 		vpn &= ~(0xffffULL << 48);
157 
158 #ifdef __powerpc64__
159 	__asm __volatile("tlbie %0" :: "r"(vpn) : "memory");
160 	__asm __volatile("eieio; tlbsync; ptesync" ::: "memory");
161 #else
162 	vpn_hi = (uint32_t)(vpn >> 32);
163 	vpn_lo = (uint32_t)vpn;
164 
165 	intr = intr_disable();
166 	__asm __volatile("\
167 	    mfmsr %0; \
168 	    mr %1, %0; \
169 	    insrdi %1,%5,1,0; \
170 	    mtmsrd %1; isync; \
171 	    \
172 	    sld %1,%2,%4; \
173 	    or %1,%1,%3; \
174 	    tlbie %1; \
175 	    \
176 	    mtmsrd %0; isync; \
177 	    eieio; \
178 	    tlbsync; \
179 	    ptesync;"
180 	: "=r"(msr), "=r"(scratch) : "r"(vpn_hi), "r"(vpn_lo), "r"(32), "r"(1)
181 	    : "memory");
182 	intr_restore(intr);
183 #endif
184 
185 	/* No barriers or special ops -- taken care of by ptesync above */
186 	tlbie_lock = 0;
187 }
188 
189 #define DISABLE_TRANS(msr)	msr = mfmsr(); mtmsr(msr & ~PSL_DR)
190 #define ENABLE_TRANS(msr)	mtmsr(msr)
191 
192 /*
193  * PTEG data.
194  */
195 static volatile struct pate *moea64_part_table;
196 static volatile struct lpte *moea64_pteg_table;
197 static struct rwlock moea64_eviction_lock;
198 
199 /*
200  * PTE calls.
201  */
202 static int	moea64_pte_insert_native(mmu_t, struct pvo_entry *);
203 static int64_t	moea64_pte_synch_native(mmu_t, struct pvo_entry *);
204 static int64_t	moea64_pte_clear_native(mmu_t, struct pvo_entry *, uint64_t);
205 static int64_t	moea64_pte_replace_native(mmu_t, struct pvo_entry *, int);
206 static int64_t	moea64_pte_unset_native(mmu_t mmu, struct pvo_entry *);
207 
208 /*
209  * Utility routines.
210  */
211 static void	moea64_bootstrap_native(mmu_t mmup,
212 		    vm_offset_t kernelstart, vm_offset_t kernelend);
213 static void	moea64_cpu_bootstrap_native(mmu_t, int ap);
214 static void	tlbia(void);
215 
216 static mmu_method_t moea64_native_methods[] = {
217 	/* Internal interfaces */
218 	MMUMETHOD(mmu_bootstrap,	moea64_bootstrap_native),
219 	MMUMETHOD(mmu_cpu_bootstrap,	moea64_cpu_bootstrap_native),
220 
221 	MMUMETHOD(moea64_pte_synch,	moea64_pte_synch_native),
222 	MMUMETHOD(moea64_pte_clear,	moea64_pte_clear_native),
223 	MMUMETHOD(moea64_pte_unset,	moea64_pte_unset_native),
224 	MMUMETHOD(moea64_pte_replace,	moea64_pte_replace_native),
225 	MMUMETHOD(moea64_pte_insert,	moea64_pte_insert_native),
226 
227 	{ 0, 0 }
228 };
229 
230 MMU_DEF_INHERIT(oea64_mmu_native, MMU_TYPE_G5, moea64_native_methods,
231     0, oea64_mmu);
232 
233 static int64_t
234 moea64_pte_synch_native(mmu_t mmu, struct pvo_entry *pvo)
235 {
236 	volatile struct lpte *pt = moea64_pteg_table + pvo->pvo_pte.slot;
237 	struct lpte properpt;
238 	uint64_t ptelo;
239 
240 	PMAP_LOCK_ASSERT(pvo->pvo_pmap, MA_OWNED);
241 
242 	moea64_pte_from_pvo(pvo, &properpt);
243 
244 	rw_rlock(&moea64_eviction_lock);
245 	if ((be64toh(pt->pte_hi) & LPTE_AVPN_MASK) !=
246 	    (properpt.pte_hi & LPTE_AVPN_MASK)) {
247 		/* Evicted */
248 		rw_runlock(&moea64_eviction_lock);
249 		return (-1);
250 	}
251 
252 	PTESYNC();
253 	ptelo = be64toh(pt->pte_lo);
254 
255 	rw_runlock(&moea64_eviction_lock);
256 
257 	return (ptelo & (LPTE_REF | LPTE_CHG));
258 }
259 
260 static int64_t
261 moea64_pte_clear_native(mmu_t mmu, struct pvo_entry *pvo, uint64_t ptebit)
262 {
263 	volatile struct lpte *pt = moea64_pteg_table + pvo->pvo_pte.slot;
264 	struct lpte properpt;
265 	uint64_t ptelo;
266 
267 	PMAP_LOCK_ASSERT(pvo->pvo_pmap, MA_OWNED);
268 
269 	moea64_pte_from_pvo(pvo, &properpt);
270 
271 	rw_rlock(&moea64_eviction_lock);
272 	if ((be64toh(pt->pte_hi) & LPTE_AVPN_MASK) !=
273 	    (properpt.pte_hi & LPTE_AVPN_MASK)) {
274 		/* Evicted */
275 		rw_runlock(&moea64_eviction_lock);
276 		return (-1);
277 	}
278 
279 	if (ptebit == LPTE_REF) {
280 		/* See "Resetting the Reference Bit" in arch manual */
281 		PTESYNC();
282 		/* 2-step here safe: precision is not guaranteed */
283 		ptelo = be64toh(pt->pte_lo);
284 
285 		/* One-byte store to avoid touching the C bit */
286 		((volatile uint8_t *)(&pt->pte_lo))[6] =
287 #if BYTE_ORDER == BIG_ENDIAN
288 		    ((uint8_t *)(&properpt.pte_lo))[6];
289 #else
290 		    ((uint8_t *)(&properpt.pte_lo))[1];
291 #endif
292 		rw_runlock(&moea64_eviction_lock);
293 
294 		critical_enter();
295 		TLBIE(pvo->pvo_vpn);
296 		critical_exit();
297 	} else {
298 		rw_runlock(&moea64_eviction_lock);
299 		ptelo = moea64_pte_unset_native(mmu, pvo);
300 		moea64_pte_insert_native(mmu, pvo);
301 	}
302 
303 	return (ptelo & (LPTE_REF | LPTE_CHG));
304 }
305 
306 static int64_t
307 moea64_pte_unset_native(mmu_t mmu, struct pvo_entry *pvo)
308 {
309 	volatile struct lpte *pt = moea64_pteg_table + pvo->pvo_pte.slot;
310 	struct lpte properpt;
311 	uint64_t ptelo;
312 
313 	moea64_pte_from_pvo(pvo, &properpt);
314 
315 	rw_rlock(&moea64_eviction_lock);
316 	if ((be64toh(pt->pte_hi & LPTE_AVPN_MASK)) !=
317 	    (properpt.pte_hi & LPTE_AVPN_MASK)) {
318 		/* Evicted */
319 		moea64_pte_overflow--;
320 		rw_runlock(&moea64_eviction_lock);
321 		return (-1);
322 	}
323 
324 	/*
325 	 * Invalidate the pte, briefly locking it to collect RC bits. No
326 	 * atomics needed since this is protected against eviction by the lock.
327 	 */
328 	isync();
329 	critical_enter();
330 	pt->pte_hi = be64toh((pt->pte_hi & ~LPTE_VALID) | LPTE_LOCKED);
331 	PTESYNC();
332 	TLBIE(pvo->pvo_vpn);
333 	ptelo = be64toh(pt->pte_lo);
334 	*((volatile int32_t *)(&pt->pte_hi) + 1) = 0; /* Release lock */
335 	critical_exit();
336 	rw_runlock(&moea64_eviction_lock);
337 
338 	/* Keep statistics */
339 	moea64_pte_valid--;
340 
341 	return (ptelo & (LPTE_CHG | LPTE_REF));
342 }
343 
344 static int64_t
345 moea64_pte_replace_native(mmu_t mmu, struct pvo_entry *pvo, int flags)
346 {
347 	volatile struct lpte *pt = moea64_pteg_table + pvo->pvo_pte.slot;
348 	struct lpte properpt;
349 	int64_t ptelo;
350 
351 	if (flags == 0) {
352 		/* Just some software bits changing. */
353 		moea64_pte_from_pvo(pvo, &properpt);
354 
355 		rw_rlock(&moea64_eviction_lock);
356 		if ((be64toh(pt->pte_hi) & LPTE_AVPN_MASK) !=
357 		    (properpt.pte_hi & LPTE_AVPN_MASK)) {
358 			rw_runlock(&moea64_eviction_lock);
359 			return (-1);
360 		}
361 		pt->pte_hi = htobe64(properpt.pte_hi);
362 		ptelo = be64toh(pt->pte_lo);
363 		rw_runlock(&moea64_eviction_lock);
364 	} else {
365 		/* Otherwise, need reinsertion and deletion */
366 		ptelo = moea64_pte_unset_native(mmu, pvo);
367 		moea64_pte_insert_native(mmu, pvo);
368 	}
369 
370 	return (ptelo);
371 }
372 
373 static void
374 moea64_cpu_bootstrap_native(mmu_t mmup, int ap)
375 {
376 	int i = 0;
377 	#ifdef __powerpc64__
378 	struct slb *slb = PCPU_GET(aim.slb);
379 	register_t seg0;
380 	#endif
381 
382 	/*
383 	 * Initialize segment registers and MMU
384 	 */
385 
386 	mtmsr(mfmsr() & ~PSL_DR & ~PSL_IR);
387 
388 	switch (mfpvr() >> 16) {
389 	case IBMPOWER9:
390 		mtspr(SPR_HID0, mfspr(SPR_HID0) & ~HID0_RADIX);
391 		break;
392 	}
393 
394 	/*
395 	 * Install kernel SLB entries
396 	 */
397 
398 	#ifdef __powerpc64__
399 		__asm __volatile ("slbia");
400 		__asm __volatile ("slbmfee %0,%1; slbie %0;" : "=r"(seg0) :
401 		    "r"(0));
402 
403 		for (i = 0; i < n_slbs; i++) {
404 			if (!(slb[i].slbe & SLBE_VALID))
405 				continue;
406 
407 			__asm __volatile ("slbmte %0, %1" ::
408 			    "r"(slb[i].slbv), "r"(slb[i].slbe));
409 		}
410 	#else
411 		for (i = 0; i < 16; i++)
412 			mtsrin(i << ADDR_SR_SHFT, kernel_pmap->pm_sr[i]);
413 	#endif
414 
415 	/*
416 	 * Install page table
417 	 */
418 
419 	if (cpu_features2 & PPC_FEATURE2_ARCH_3_00) {
420 		mtspr(SPR_PTCR,
421 		    ((uintptr_t)moea64_part_table & ~DMAP_BASE_ADDRESS) |
422 		     flsl((PART_SIZE >> 12) - 1));
423 	} else {
424 		__asm __volatile ("ptesync; mtsdr1 %0; isync"
425 		    :: "r"(((uintptr_t)moea64_pteg_table & ~DMAP_BASE_ADDRESS)
426 			     | (uintptr_t)(flsl(moea64_pteg_mask >> 11))));
427 	}
428 	tlbia();
429 }
430 
431 static void
432 moea64_bootstrap_native(mmu_t mmup, vm_offset_t kernelstart,
433     vm_offset_t kernelend)
434 {
435 	vm_size_t	size;
436 	vm_offset_t	off;
437 	vm_paddr_t	pa;
438 	register_t	msr;
439 
440 	moea64_early_bootstrap(mmup, kernelstart, kernelend);
441 
442 	switch (mfpvr() >> 16) {
443 	case IBMPOWER4:
444 	case IBMPOWER4PLUS:
445 	case IBM970:
446 	case IBM970FX:
447 	case IBM970GX:
448 	case IBM970MP:
449 	    	moea64_crop_tlbie = true;
450 	}
451 	/*
452 	 * Allocate PTEG table.
453 	 */
454 
455 	size = moea64_pteg_count * sizeof(struct lpteg);
456 	CTR2(KTR_PMAP, "moea64_bootstrap: %d PTEGs, %d bytes",
457 	    moea64_pteg_count, size);
458 	rw_init(&moea64_eviction_lock, "pte eviction");
459 
460 	/*
461 	 * We now need to allocate memory. This memory, to be allocated,
462 	 * has to reside in a page table. The page table we are about to
463 	 * allocate. We don't have BAT. So drop to data real mode for a minute
464 	 * as a measure of last resort. We do this a couple times.
465 	 */
466 
467 	if (cpu_features2 & PPC_FEATURE2_ARCH_3_00) {
468 		moea64_part_table =
469 		    (struct pate *)moea64_bootstrap_alloc(PART_SIZE, PART_SIZE);
470 		if (hw_direct_map)
471 			moea64_part_table = (struct pate *)PHYS_TO_DMAP(
472 			    (vm_offset_t)moea64_part_table);
473 	}
474 	/*
475 	 * PTEG table must be aligned on a 256k boundary, but can be placed
476 	 * anywhere with that alignment on POWER ISA 3+ systems. On earlier
477 	 * systems, offset addition is done by the CPU with bitwise OR rather
478 	 * than addition, so the table must also be aligned on a boundary of
479 	 * its own size. Pick the larger of the two, which works on all
480 	 * systems.
481 	 */
482 	moea64_pteg_table = (struct lpte *)moea64_bootstrap_alloc(size,
483 	    MAX(256*1024, size));
484 	if (hw_direct_map)
485 		moea64_pteg_table =
486 		    (struct lpte *)PHYS_TO_DMAP((vm_offset_t)moea64_pteg_table);
487 	DISABLE_TRANS(msr);
488 	if (cpu_features2 & PPC_FEATURE2_ARCH_3_00) {
489 		bzero(__DEVOLATILE(void *, moea64_part_table), PART_SIZE);
490 		moea64_part_table[0].pagetab =
491 		    ((uintptr_t)moea64_pteg_table & ~DMAP_BASE_ADDRESS) |
492 		    (uintptr_t)(flsl((moea64_pteg_count - 1) >> 11));
493 	}
494 	bzero(__DEVOLATILE(void *, moea64_pteg_table), moea64_pteg_count *
495 	    sizeof(struct lpteg));
496 	ENABLE_TRANS(msr);
497 
498 	CTR1(KTR_PMAP, "moea64_bootstrap: PTEG table at %p", moea64_pteg_table);
499 
500 	moea64_mid_bootstrap(mmup, kernelstart, kernelend);
501 
502 	/*
503 	 * Add a mapping for the page table itself if there is no direct map.
504 	 */
505 	if (!hw_direct_map) {
506 		size = moea64_pteg_count * sizeof(struct lpteg);
507 		off = (vm_offset_t)(moea64_pteg_table);
508 		DISABLE_TRANS(msr);
509 		for (pa = off; pa < off + size; pa += PAGE_SIZE)
510 			pmap_kenter(pa, pa);
511 		ENABLE_TRANS(msr);
512 	}
513 
514 	/* Bring up virtual memory */
515 	moea64_late_bootstrap(mmup, kernelstart, kernelend);
516 }
517 
518 static void
519 tlbia(void)
520 {
521 	vm_offset_t i;
522 	#ifndef __powerpc64__
523 	register_t msr, scratch;
524 	#endif
525 
526 	i = 0xc00; /* IS = 11 */
527 	switch (mfpvr() >> 16) {
528 	case IBM970:
529 	case IBM970FX:
530 	case IBM970MP:
531 	case IBM970GX:
532 	case IBMPOWER4:
533 	case IBMPOWER4PLUS:
534 	case IBMPOWER5:
535 	case IBMPOWER5PLUS:
536 		i = 0; /* IS not supported */
537 		break;
538 	}
539 
540 	TLBSYNC();
541 
542 	for (; i < 0x200000; i += 0x00001000) {
543 		#ifdef __powerpc64__
544 		__asm __volatile("tlbiel %0" :: "r"(i));
545 		#else
546 		__asm __volatile("\
547 		    mfmsr %0; \
548 		    mr %1, %0; \
549 		    insrdi %1,%3,1,0; \
550 		    mtmsrd %1; \
551 		    isync; \
552 		    \
553 		    tlbiel %2; \
554 		    \
555 		    mtmsrd %0; \
556 		    isync;"
557 		: "=r"(msr), "=r"(scratch) : "r"(i), "r"(1));
558 		#endif
559 	}
560 
561 	EIEIO();
562 	TLBSYNC();
563 }
564 
565 static int
566 atomic_pte_lock(volatile struct lpte *pte, uint64_t bitmask, uint64_t *oldhi)
567 {
568 	int	ret;
569 	uint32_t oldhihalf;
570 
571 	/*
572 	 * Note: in principle, if just the locked bit were set here, we
573 	 * could avoid needing the eviction lock. However, eviction occurs
574 	 * so rarely that it isn't worth bothering about in practice.
575 	 */
576 
577 	__asm __volatile (
578 		"1:\tlwarx %1, 0, %3\n\t"	/* load old value */
579 		"and. %0,%1,%4\n\t"		/* check if any bits set */
580 		"bne 2f\n\t"			/* exit if any set */
581 		"stwcx. %5, 0, %3\n\t"      	/* attempt to store */
582 		"bne- 1b\n\t"			/* spin if failed */
583 		"li %0, 1\n\t"			/* success - retval = 1 */
584 		"b 3f\n\t"			/* we've succeeded */
585 		"2:\n\t"
586 		"stwcx. %1, 0, %3\n\t"       	/* clear reservation (74xx) */
587 		"li %0, 0\n\t"			/* failure - retval = 0 */
588 		"3:\n\t"
589 		: "=&r" (ret), "=&r"(oldhihalf), "=m" (pte->pte_hi)
590 		: "r" ((volatile char *)&pte->pte_hi + 4),
591 		  "r" ((uint32_t)bitmask), "r" ((uint32_t)LPTE_LOCKED),
592 		  "m" (pte->pte_hi)
593 		: "cr0", "cr1", "cr2", "memory");
594 
595 	*oldhi = (pte->pte_hi & 0xffffffff00000000ULL) | oldhihalf;
596 
597 	return (ret);
598 }
599 
600 static uintptr_t
601 moea64_insert_to_pteg_native(struct lpte *pvo_pt, uintptr_t slotbase,
602     uint64_t mask)
603 {
604 	volatile struct lpte *pt;
605 	uint64_t oldptehi, va;
606 	uintptr_t k;
607 	int i, j;
608 
609 	/* Start at a random slot */
610 	i = mftb() % 8;
611 	for (j = 0; j < 8; j++) {
612 		k = slotbase + (i + j) % 8;
613 		pt = &moea64_pteg_table[k];
614 		/* Invalidate and seize lock only if no bits in mask set */
615 		if (atomic_pte_lock(pt, mask, &oldptehi)) /* Lock obtained */
616 			break;
617 	}
618 
619 	if (j == 8)
620 		return (-1);
621 
622 	if (oldptehi & LPTE_VALID) {
623 		KASSERT(!(oldptehi & LPTE_WIRED), ("Unmapped wired entry"));
624 		/*
625 		 * Need to invalidate old entry completely: see
626 		 * "Modifying a Page Table Entry". Need to reconstruct
627 		 * the virtual address for the outgoing entry to do that.
628 		 */
629 		if (oldptehi & LPTE_BIG)
630 			va = oldptehi >> moea64_large_page_shift;
631 		else
632 			va = oldptehi >> ADDR_PIDX_SHFT;
633 		if (oldptehi & LPTE_HID)
634 			va = (((k >> 3) ^ moea64_pteg_mask) ^ va) &
635 			    VSID_HASH_MASK;
636 		else
637 			va = ((k >> 3) ^ va) & VSID_HASH_MASK;
638 		va |= (oldptehi & LPTE_AVPN_MASK) <<
639 		    (ADDR_API_SHFT64 - ADDR_PIDX_SHFT);
640 		PTESYNC();
641 		TLBIE(va);
642 		moea64_pte_valid--;
643 		moea64_pte_overflow++;
644 	}
645 
646 	/*
647 	 * Update the PTE as per "Adding a Page Table Entry". Lock is released
648 	 * by setting the high doubleworld.
649 	 */
650 	pt->pte_lo = htobe64(pvo_pt->pte_lo);
651 	EIEIO();
652 	pt->pte_hi = htobe64(pvo_pt->pte_hi);
653 	PTESYNC();
654 
655 	/* Keep statistics */
656 	moea64_pte_valid++;
657 
658 	return (k);
659 }
660 
661 static int
662 moea64_pte_insert_native(mmu_t mmu, struct pvo_entry *pvo)
663 {
664 	struct lpte insertpt;
665 	uintptr_t slot;
666 
667 	/* Initialize PTE */
668 	moea64_pte_from_pvo(pvo, &insertpt);
669 
670 	/* Make sure further insertion is locked out during evictions */
671 	rw_rlock(&moea64_eviction_lock);
672 
673 	/*
674 	 * First try primary hash.
675 	 */
676 	pvo->pvo_pte.slot &= ~7ULL; /* Base slot address */
677 	slot = moea64_insert_to_pteg_native(&insertpt, pvo->pvo_pte.slot,
678 	    LPTE_VALID | LPTE_WIRED | LPTE_LOCKED);
679 	if (slot != -1) {
680 		rw_runlock(&moea64_eviction_lock);
681 		pvo->pvo_pte.slot = slot;
682 		return (0);
683 	}
684 
685 	/*
686 	 * Now try secondary hash.
687 	 */
688 	pvo->pvo_vaddr ^= PVO_HID;
689 	insertpt.pte_hi ^= LPTE_HID;
690 	pvo->pvo_pte.slot ^= (moea64_pteg_mask << 3);
691 	slot = moea64_insert_to_pteg_native(&insertpt, pvo->pvo_pte.slot,
692 	    LPTE_VALID | LPTE_WIRED | LPTE_LOCKED);
693 	if (slot != -1) {
694 		rw_runlock(&moea64_eviction_lock);
695 		pvo->pvo_pte.slot = slot;
696 		return (0);
697 	}
698 
699 	/*
700 	 * Out of luck. Find a PTE to sacrifice.
701 	 */
702 
703 	/* Lock out all insertions for a bit */
704 	if (!rw_try_upgrade(&moea64_eviction_lock)) {
705 		rw_runlock(&moea64_eviction_lock);
706 		rw_wlock(&moea64_eviction_lock);
707 	}
708 
709 	slot = moea64_insert_to_pteg_native(&insertpt, pvo->pvo_pte.slot,
710 	    LPTE_WIRED | LPTE_LOCKED);
711 	if (slot != -1) {
712 		rw_wunlock(&moea64_eviction_lock);
713 		pvo->pvo_pte.slot = slot;
714 		return (0);
715 	}
716 
717 	/* Try other hash table. Now we're getting desperate... */
718 	pvo->pvo_vaddr ^= PVO_HID;
719 	insertpt.pte_hi ^= LPTE_HID;
720 	pvo->pvo_pte.slot ^= (moea64_pteg_mask << 3);
721 	slot = moea64_insert_to_pteg_native(&insertpt, pvo->pvo_pte.slot,
722 	    LPTE_WIRED | LPTE_LOCKED);
723 	if (slot != -1) {
724 		rw_wunlock(&moea64_eviction_lock);
725 		pvo->pvo_pte.slot = slot;
726 		return (0);
727 	}
728 
729 	/* No freeable slots in either PTEG? We're hosed. */
730 	rw_wunlock(&moea64_eviction_lock);
731 	panic("moea64_pte_insert: overflow");
732 	return (-1);
733 }
734 
735