xref: /freebsd/sys/powerpc/aim/moea64_native.c (revision ff490346fd519501e9263988959ffb583013ccdb)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD AND BSD-4-Clause
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 
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 /* POWER9 only permits a 64k partition table size. */
133 #define	PART_SIZE	0x10000
134 
135 /* Actual page sizes (to be used with tlbie, when L=0) */
136 #define	AP_4K		0x00
137 #define	AP_16M		0x80
138 
139 #define	LPTE_KERNEL_VSID_BIT	(KERNEL_VSID_BIT << \
140 				(16 - (ADDR_API_SHFT64 - ADDR_PIDX_SHFT)))
141 
142 /* Abbreviated Virtual Address Page - high bits */
143 #define	LPTE_AVA_PGNHI_MASK	0x0000000000000F80ULL
144 #define	LPTE_AVA_PGNHI_SHIFT	7
145 
146 /* Effective Address Page - low bits */
147 #define	EA_PAGELO_MASK		0x7ffULL
148 #define	EA_PAGELO_SHIFT		11
149 
150 static bool moea64_crop_tlbie;
151 static bool moea64_need_lock;
152 
153 /*
154  * The tlbie instruction has two forms: an old one used by PowerISA
155  * 2.03 and prior, and a newer one used by PowerISA 2.06 and later.
156  * We need to support both.
157  */
158 static __inline void
159 TLBIE(uint64_t vpn, uint64_t oldptehi)
160 {
161 #ifndef __powerpc64__
162 	register_t vpn_hi, vpn_lo;
163 	register_t msr;
164 	register_t scratch, intr;
165 #endif
166 
167 	static volatile u_int tlbie_lock = 0;
168 	bool need_lock = moea64_need_lock;
169 
170 	vpn <<= ADDR_PIDX_SHFT;
171 
172 	/* Hobo spinlock: we need stronger guarantees than mutexes provide */
173 	if (need_lock) {
174 		while (!atomic_cmpset_int(&tlbie_lock, 0, 1));
175 		isync(); /* Flush instruction queue once lock acquired */
176 
177 		if (moea64_crop_tlbie) {
178 			vpn &= ~(0xffffULL << 48);
179 #ifdef __powerpc64__
180 			if ((oldptehi & LPTE_BIG) != 0)
181 				__asm __volatile("tlbie %0, 1" :: "r"(vpn) :
182 				    "memory");
183 			else
184 				__asm __volatile("tlbie %0, 0" :: "r"(vpn) :
185 				    "memory");
186 			__asm __volatile("eieio; tlbsync; ptesync" :::
187 			    "memory");
188 			goto done;
189 #endif
190 		}
191 	}
192 
193 #ifdef __powerpc64__
194 	/*
195 	 * If this page has LPTE_BIG set and is from userspace, then
196 	 * it must be a superpage with 4KB base/16MB actual page size.
197 	 */
198 	if ((oldptehi & LPTE_BIG) != 0 &&
199 	    (oldptehi & LPTE_KERNEL_VSID_BIT) == 0)
200 		vpn |= AP_16M;
201 
202 	/*
203 	 * Explicitly clobber r0.  The tlbie instruction has two forms: an old
204 	 * one used by PowerISA 2.03 and prior, and a newer one used by PowerISA
205 	 * 2.06 (maybe 2.05?) and later.  We need to support both, and it just
206 	 * so happens that since we use 4k pages we can simply zero out r0, and
207 	 * clobber it, and the assembler will interpret the single-operand form
208 	 * of tlbie as having RB set, and everything else as 0.  The RS operand
209 	 * in the newer form is in the same position as the L(page size) bit of
210 	 * the old form, so a slong as RS is 0, we're good on both sides.
211 	 */
212 	__asm __volatile("li 0, 0 \n tlbie %0, 0" :: "r"(vpn) : "r0", "memory");
213 	__asm __volatile("eieio; tlbsync; ptesync" ::: "memory");
214 done:
215 
216 #else
217 	vpn_hi = (uint32_t)(vpn >> 32);
218 	vpn_lo = (uint32_t)vpn;
219 
220 	intr = intr_disable();
221 	__asm __volatile("\
222 	    mfmsr %0; \
223 	    mr %1, %0; \
224 	    insrdi %1,%5,1,0; \
225 	    mtmsrd %1; isync; \
226 	    \
227 	    sld %1,%2,%4; \
228 	    or %1,%1,%3; \
229 	    tlbie %1; \
230 	    \
231 	    mtmsrd %0; isync; \
232 	    eieio; \
233 	    tlbsync; \
234 	    ptesync;"
235 	: "=r"(msr), "=r"(scratch) : "r"(vpn_hi), "r"(vpn_lo), "r"(32), "r"(1)
236 	    : "memory");
237 	intr_restore(intr);
238 #endif
239 
240 	/* No barriers or special ops -- taken care of by ptesync above */
241 	if (need_lock)
242 		tlbie_lock = 0;
243 }
244 
245 #define DISABLE_TRANS(msr)	msr = mfmsr(); mtmsr(msr & ~PSL_DR)
246 #define ENABLE_TRANS(msr)	mtmsr(msr)
247 
248 /*
249  * PTEG data.
250  */
251 static volatile struct lpte *moea64_pteg_table;
252 static struct rwlock moea64_eviction_lock;
253 
254 static volatile struct pate *moea64_part_table;
255 
256 /*
257  * Dump function.
258  */
259 static void	*moea64_dump_pmap_native(void *ctx, void *buf,
260 		    u_long *nbytes);
261 
262 /*
263  * PTE calls.
264  */
265 static int64_t	moea64_pte_insert_native(struct pvo_entry *);
266 static int64_t	moea64_pte_synch_native(struct pvo_entry *);
267 static int64_t	moea64_pte_clear_native(struct pvo_entry *, uint64_t);
268 static int64_t	moea64_pte_replace_native(struct pvo_entry *, int);
269 static int64_t	moea64_pte_unset_native(struct pvo_entry *);
270 static int64_t	moea64_pte_insert_sp_native(struct pvo_entry *);
271 static int64_t	moea64_pte_unset_sp_native(struct pvo_entry *);
272 static int64_t	moea64_pte_replace_sp_native(struct pvo_entry *);
273 
274 /*
275  * Utility routines.
276  */
277 static void	moea64_bootstrap_native(
278 		    vm_offset_t kernelstart, vm_offset_t kernelend);
279 static void	moea64_cpu_bootstrap_native(int ap);
280 static void	tlbia(void);
281 static void	moea64_install_native(void);
282 
283 static struct pmap_funcs moea64_native_methods = {
284 	.install = moea64_install_native,
285 
286 	/* Internal interfaces */
287 	.bootstrap = moea64_bootstrap_native,
288 	.cpu_bootstrap = moea64_cpu_bootstrap_native,
289         .dumpsys_dump_pmap =         moea64_dump_pmap_native,
290 };
291 
292 static struct moea64_funcs moea64_native_funcs = {
293 	.pte_synch = moea64_pte_synch_native,
294 	.pte_clear = moea64_pte_clear_native,
295 	.pte_unset = moea64_pte_unset_native,
296 	.pte_replace = moea64_pte_replace_native,
297 	.pte_insert = moea64_pte_insert_native,
298 	.pte_insert_sp = moea64_pte_insert_sp_native,
299 	.pte_unset_sp = moea64_pte_unset_sp_native,
300 	.pte_replace_sp = moea64_pte_replace_sp_native,
301 };
302 
303 MMU_DEF_INHERIT(oea64_mmu_native, MMU_TYPE_G5, moea64_native_methods, oea64_mmu);
304 
305 static void
306 moea64_install_native(void)
307 {
308 
309 	/* Install the MOEA64 ops. */
310 	moea64_ops = &moea64_native_funcs;
311 
312 	moea64_install();
313 }
314 
315 static int64_t
316 moea64_pte_synch_native(struct pvo_entry *pvo)
317 {
318 	volatile struct lpte *pt = moea64_pteg_table + pvo->pvo_pte.slot;
319 	uint64_t ptelo, pvo_ptevpn;
320 
321 	PMAP_LOCK_ASSERT(pvo->pvo_pmap, MA_OWNED);
322 
323 	pvo_ptevpn = moea64_pte_vpn_from_pvo_vpn(pvo);
324 
325 	rw_rlock(&moea64_eviction_lock);
326 	if ((be64toh(pt->pte_hi) & LPTE_AVPN_MASK) != pvo_ptevpn) {
327 		/* Evicted */
328 		rw_runlock(&moea64_eviction_lock);
329 		return (-1);
330 	}
331 
332 	PTESYNC();
333 	ptelo = be64toh(pt->pte_lo);
334 
335 	rw_runlock(&moea64_eviction_lock);
336 
337 	return (ptelo & (LPTE_REF | LPTE_CHG));
338 }
339 
340 static int64_t
341 moea64_pte_clear_native(struct pvo_entry *pvo, uint64_t ptebit)
342 {
343 	volatile struct lpte *pt = moea64_pteg_table + pvo->pvo_pte.slot;
344 	struct lpte properpt;
345 	uint64_t ptelo;
346 
347 	PMAP_LOCK_ASSERT(pvo->pvo_pmap, MA_OWNED);
348 
349 	moea64_pte_from_pvo(pvo, &properpt);
350 
351 	rw_rlock(&moea64_eviction_lock);
352 	if ((be64toh(pt->pte_hi) & LPTE_AVPN_MASK) !=
353 	    (properpt.pte_hi & LPTE_AVPN_MASK)) {
354 		/* Evicted */
355 		rw_runlock(&moea64_eviction_lock);
356 		return (-1);
357 	}
358 
359 	if (ptebit == LPTE_REF) {
360 		/* See "Resetting the Reference Bit" in arch manual */
361 		PTESYNC();
362 		/* 2-step here safe: precision is not guaranteed */
363 		ptelo = be64toh(pt->pte_lo);
364 
365 		/* One-byte store to avoid touching the C bit */
366 		((volatile uint8_t *)(&pt->pte_lo))[6] =
367 #if BYTE_ORDER == BIG_ENDIAN
368 		    ((uint8_t *)(&properpt.pte_lo))[6];
369 #else
370 		    ((uint8_t *)(&properpt.pte_lo))[1];
371 #endif
372 		rw_runlock(&moea64_eviction_lock);
373 
374 		critical_enter();
375 		TLBIE(pvo->pvo_vpn, properpt.pte_hi);
376 		critical_exit();
377 	} else {
378 		rw_runlock(&moea64_eviction_lock);
379 		ptelo = moea64_pte_unset_native(pvo);
380 		moea64_pte_insert_native(pvo);
381 	}
382 
383 	return (ptelo & (LPTE_REF | LPTE_CHG));
384 }
385 
386 static __always_inline int64_t
387 moea64_pte_unset_locked(volatile struct lpte *pt, uint64_t vpn)
388 {
389 	uint64_t ptelo, ptehi;
390 
391 	/*
392 	 * Invalidate the pte, briefly locking it to collect RC bits. No
393 	 * atomics needed since this is protected against eviction by the lock.
394 	 */
395 	isync();
396 	critical_enter();
397 	ptehi = (be64toh(pt->pte_hi) & ~LPTE_VALID) | LPTE_LOCKED;
398 	pt->pte_hi = htobe64(ptehi);
399 	PTESYNC();
400 	TLBIE(vpn, ptehi);
401 	ptelo = be64toh(pt->pte_lo);
402 	*((volatile int32_t *)(&pt->pte_hi) + 1) = 0; /* Release lock */
403 	critical_exit();
404 
405 	/* Keep statistics */
406 	STAT_MOEA64(moea64_pte_valid--);
407 
408 	return (ptelo & (LPTE_CHG | LPTE_REF));
409 }
410 
411 static int64_t
412 moea64_pte_unset_native(struct pvo_entry *pvo)
413 {
414 	volatile struct lpte *pt = moea64_pteg_table + pvo->pvo_pte.slot;
415 	int64_t ret;
416 	uint64_t pvo_ptevpn;
417 
418 	pvo_ptevpn = moea64_pte_vpn_from_pvo_vpn(pvo);
419 
420 	rw_rlock(&moea64_eviction_lock);
421 
422 	if ((be64toh(pt->pte_hi) & LPTE_AVPN_MASK) != pvo_ptevpn) {
423 		/* Evicted */
424 		STAT_MOEA64(moea64_pte_overflow--);
425 		ret = -1;
426 	} else
427 		ret = moea64_pte_unset_locked(pt, pvo->pvo_vpn);
428 
429 	rw_runlock(&moea64_eviction_lock);
430 
431 	return (ret);
432 }
433 
434 static int64_t
435 moea64_pte_replace_inval_native(struct pvo_entry *pvo,
436     volatile struct lpte *pt)
437 {
438 	struct lpte properpt;
439 	uint64_t ptelo, ptehi;
440 
441 	moea64_pte_from_pvo(pvo, &properpt);
442 
443 	rw_rlock(&moea64_eviction_lock);
444 	if ((be64toh(pt->pte_hi) & LPTE_AVPN_MASK) !=
445 	    (properpt.pte_hi & LPTE_AVPN_MASK)) {
446 		/* Evicted */
447 		STAT_MOEA64(moea64_pte_overflow--);
448 		rw_runlock(&moea64_eviction_lock);
449 		return (-1);
450 	}
451 
452 	/*
453 	 * Replace the pte, briefly locking it to collect RC bits. No
454 	 * atomics needed since this is protected against eviction by the lock.
455 	 */
456 	isync();
457 	critical_enter();
458 	ptehi = (be64toh(pt->pte_hi) & ~LPTE_VALID) | LPTE_LOCKED;
459 	pt->pte_hi = htobe64(ptehi);
460 	PTESYNC();
461 	TLBIE(pvo->pvo_vpn, ptehi);
462 	ptelo = be64toh(pt->pte_lo);
463 	EIEIO();
464 	pt->pte_lo = htobe64(properpt.pte_lo);
465 	EIEIO();
466 	pt->pte_hi = htobe64(properpt.pte_hi); /* Release lock */
467 	PTESYNC();
468 	critical_exit();
469 	rw_runlock(&moea64_eviction_lock);
470 
471 	return (ptelo & (LPTE_CHG | LPTE_REF));
472 }
473 
474 static int64_t
475 moea64_pte_replace_native(struct pvo_entry *pvo, int flags)
476 {
477 	volatile struct lpte *pt = moea64_pteg_table + pvo->pvo_pte.slot;
478 	struct lpte properpt;
479 	int64_t ptelo;
480 
481 	if (flags == 0) {
482 		/* Just some software bits changing. */
483 		moea64_pte_from_pvo(pvo, &properpt);
484 
485 		rw_rlock(&moea64_eviction_lock);
486 		if ((be64toh(pt->pte_hi) & LPTE_AVPN_MASK) !=
487 		    (properpt.pte_hi & LPTE_AVPN_MASK)) {
488 			rw_runlock(&moea64_eviction_lock);
489 			return (-1);
490 		}
491 		pt->pte_hi = htobe64(properpt.pte_hi);
492 		ptelo = be64toh(pt->pte_lo);
493 		rw_runlock(&moea64_eviction_lock);
494 	} else {
495 		/* Otherwise, need reinsertion and deletion */
496 		ptelo = moea64_pte_replace_inval_native(pvo, pt);
497 	}
498 
499 	return (ptelo);
500 }
501 
502 static void
503 moea64_cpu_bootstrap_native(int ap)
504 {
505 	int i = 0;
506 	#ifdef __powerpc64__
507 	struct slb *slb = PCPU_GET(aim.slb);
508 	register_t seg0;
509 	#endif
510 
511 	/*
512 	 * Initialize segment registers and MMU
513 	 */
514 
515 	mtmsr(mfmsr() & ~PSL_DR & ~PSL_IR);
516 
517 	switch(mfpvr() >> 16) {
518 	case IBMPOWER9:
519 		mtspr(SPR_HID0, mfspr(SPR_HID0) & ~HID0_RADIX);
520 		break;
521 	}
522 
523 	/*
524 	 * Install kernel SLB entries
525 	 */
526 
527 	#ifdef __powerpc64__
528 		__asm __volatile ("slbia");
529 		__asm __volatile ("slbmfee %0,%1; slbie %0;" : "=r"(seg0) :
530 		    "r"(0));
531 
532 		for (i = 0; i < n_slbs; i++) {
533 			if (!(slb[i].slbe & SLBE_VALID))
534 				continue;
535 
536 			__asm __volatile ("slbmte %0, %1" ::
537 			    "r"(slb[i].slbv), "r"(slb[i].slbe));
538 		}
539 	#else
540 		for (i = 0; i < 16; i++)
541 			mtsrin(i << ADDR_SR_SHFT, kernel_pmap->pm_sr[i]);
542 	#endif
543 
544 	/*
545 	 * Install page table
546 	 */
547 
548 	if (cpu_features2 & PPC_FEATURE2_ARCH_3_00)
549 		mtspr(SPR_PTCR,
550 		    ((uintptr_t)moea64_part_table & ~DMAP_BASE_ADDRESS) |
551 		     flsl((PART_SIZE >> 12) - 1));
552 	else
553 		__asm __volatile ("ptesync; mtsdr1 %0; isync"
554 		    :: "r"(((uintptr_t)moea64_pteg_table & ~DMAP_BASE_ADDRESS)
555 			     | (uintptr_t)(flsl(moea64_pteg_mask >> 11))));
556 	tlbia();
557 }
558 
559 static void
560 moea64_bootstrap_native(vm_offset_t kernelstart, vm_offset_t kernelend)
561 {
562 	vm_size_t	size;
563 	vm_offset_t	off;
564 	vm_paddr_t	pa;
565 	register_t	msr;
566 
567 	moea64_early_bootstrap(kernelstart, kernelend);
568 
569 	switch (mfpvr() >> 16) {
570 	case IBMPOWER9:
571 		moea64_need_lock = false;
572 		break;
573 	case IBMPOWER4:
574 	case IBMPOWER4PLUS:
575 	case IBM970:
576 	case IBM970FX:
577 	case IBM970GX:
578 	case IBM970MP:
579 		moea64_crop_tlbie = true;
580 	default:
581 		moea64_need_lock = true;
582 	}
583 	/*
584 	 * Allocate PTEG table.
585 	 */
586 
587 	size = moea64_pteg_count * sizeof(struct lpteg);
588 	CTR2(KTR_PMAP, "moea64_bootstrap: %lu PTEGs, %lu bytes",
589 	    moea64_pteg_count, size);
590 	rw_init(&moea64_eviction_lock, "pte eviction");
591 
592 	/*
593 	 * We now need to allocate memory. This memory, to be allocated,
594 	 * has to reside in a page table. The page table we are about to
595 	 * allocate. We don't have BAT. So drop to data real mode for a minute
596 	 * as a measure of last resort. We do this a couple times.
597 	 */
598 	/*
599 	 * PTEG table must be aligned on a 256k boundary, but can be placed
600 	 * anywhere with that alignment on POWER ISA 3+ systems. On earlier
601 	 * systems, offset addition is done by the CPU with bitwise OR rather
602 	 * than addition, so the table must also be aligned on a boundary of
603 	 * its own size. Pick the larger of the two, which works on all
604 	 * systems.
605 	 */
606 	moea64_pteg_table = (struct lpte *)moea64_bootstrap_alloc(size,
607 	    MAX(256*1024, size));
608 	if (hw_direct_map)
609 		moea64_pteg_table =
610 		    (struct lpte *)PHYS_TO_DMAP((vm_offset_t)moea64_pteg_table);
611 	/* Allocate partition table (ISA 3.0). */
612 	if (cpu_features2 & PPC_FEATURE2_ARCH_3_00) {
613 		moea64_part_table =
614 		    (struct pate *)moea64_bootstrap_alloc(PART_SIZE, PART_SIZE);
615 		moea64_part_table =
616 		    (struct pate *)PHYS_TO_DMAP((vm_offset_t)moea64_part_table);
617 	}
618 	DISABLE_TRANS(msr);
619 	bzero(__DEVOLATILE(void *, moea64_pteg_table), moea64_pteg_count *
620 	    sizeof(struct lpteg));
621 	if (cpu_features2 & PPC_FEATURE2_ARCH_3_00) {
622 		bzero(__DEVOLATILE(void *, moea64_part_table), PART_SIZE);
623 		moea64_part_table[0].pagetab = htobe64(
624 			(DMAP_TO_PHYS((vm_offset_t)moea64_pteg_table)) |
625 			(uintptr_t)(flsl((moea64_pteg_count - 1) >> 11)));
626 	}
627 	ENABLE_TRANS(msr);
628 
629 	CTR1(KTR_PMAP, "moea64_bootstrap: PTEG table at %p", moea64_pteg_table);
630 
631 	moea64_mid_bootstrap(kernelstart, kernelend);
632 
633 	/*
634 	 * Add a mapping for the page table itself if there is no direct map.
635 	 */
636 	if (!hw_direct_map) {
637 		size = moea64_pteg_count * sizeof(struct lpteg);
638 		off = (vm_offset_t)(moea64_pteg_table);
639 		DISABLE_TRANS(msr);
640 		for (pa = off; pa < off + size; pa += PAGE_SIZE)
641 			pmap_kenter(pa, pa);
642 		ENABLE_TRANS(msr);
643 	}
644 
645 	/* Bring up virtual memory */
646 	moea64_late_bootstrap(kernelstart, kernelend);
647 }
648 
649 static void
650 tlbia(void)
651 {
652 	vm_offset_t i;
653 	#ifndef __powerpc64__
654 	register_t msr, scratch;
655 	#endif
656 
657 	i = 0xc00; /* IS = 11 */
658 	switch (mfpvr() >> 16) {
659 	case IBM970:
660 	case IBM970FX:
661 	case IBM970MP:
662 	case IBM970GX:
663 	case IBMPOWER4:
664 	case IBMPOWER4PLUS:
665 	case IBMPOWER5:
666 	case IBMPOWER5PLUS:
667 		i = 0; /* IS not supported */
668 		break;
669 	}
670 
671 	TLBSYNC();
672 
673 	for (; i < 0x400000; i += 0x00001000) {
674 		#ifdef __powerpc64__
675 		__asm __volatile("tlbiel %0" :: "r"(i));
676 		#else
677 		__asm __volatile("\
678 		    mfmsr %0; \
679 		    mr %1, %0; \
680 		    insrdi %1,%3,1,0; \
681 		    mtmsrd %1; \
682 		    isync; \
683 		    \
684 		    tlbiel %2; \
685 		    \
686 		    mtmsrd %0; \
687 		    isync;"
688 		: "=r"(msr), "=r"(scratch) : "r"(i), "r"(1));
689 		#endif
690 	}
691 
692 	EIEIO();
693 	TLBSYNC();
694 }
695 
696 static int
697 atomic_pte_lock(volatile struct lpte *pte, uint64_t bitmask, uint64_t *oldhi)
698 {
699 	int	ret;
700 #ifdef __powerpc64__
701 	uint64_t temp;
702 #else
703 	uint32_t oldhihalf;
704 #endif
705 
706 	/*
707 	 * Note: in principle, if just the locked bit were set here, we
708 	 * could avoid needing the eviction lock. However, eviction occurs
709 	 * so rarely that it isn't worth bothering about in practice.
710 	 */
711 #ifdef __powerpc64__
712 	/*
713 	 * Note: Success of this sequence has the side effect of invalidating
714 	 * the PTE, as we are setting it to LPTE_LOCKED and discarding the
715 	 * other bits, including LPTE_V.
716 	 */
717 	__asm __volatile (
718 		"1:\tldarx %1, 0, %3\n\t"	/* load old value */
719 		"and. %0,%1,%4\n\t"		/* check if any bits set */
720 		"bne 2f\n\t"			/* exit if any set */
721 		"stdcx. %5, 0, %3\n\t"		/* attempt to store */
722 		"bne- 1b\n\t"			/* spin if failed */
723 		"li %0, 1\n\t"			/* success - retval = 1 */
724 		"b 3f\n\t"			/* we've succeeded */
725 		"2:\n\t"
726 		"stdcx. %1, 0, %3\n\t"       	/* clear reservation (74xx) */
727 		"li %0, 0\n\t"			/* failure - retval = 0 */
728 		"3:\n\t"
729 		: "=&r" (ret), "=&r"(temp), "=m" (pte->pte_hi)
730 		: "r" ((volatile char *)&pte->pte_hi),
731 		  "r" (htobe64(bitmask)), "r" (htobe64(LPTE_LOCKED)),
732 		  "m" (pte->pte_hi)
733 		: "cr0", "cr1", "cr2", "memory");
734 	*oldhi = be64toh(temp);
735 #else
736 	/*
737 	 * This code is used on bridge mode only.
738 	 */
739 	__asm __volatile (
740 		"1:\tlwarx %1, 0, %3\n\t"	/* load old value */
741 		"and. %0,%1,%4\n\t"		/* check if any bits set */
742 		"bne 2f\n\t"			/* exit if any set */
743 		"stwcx. %5, 0, %3\n\t"      	/* attempt to store */
744 		"bne- 1b\n\t"			/* spin if failed */
745 		"li %0, 1\n\t"			/* success - retval = 1 */
746 		"b 3f\n\t"			/* we've succeeded */
747 		"2:\n\t"
748 		"stwcx. %1, 0, %3\n\t"       	/* clear reservation (74xx) */
749 		"li %0, 0\n\t"			/* failure - retval = 0 */
750 		"3:\n\t"
751 		: "=&r" (ret), "=&r"(oldhihalf), "=m" (pte->pte_hi)
752 		: "r" ((volatile char *)&pte->pte_hi + 4),
753 		  "r" ((uint32_t)bitmask), "r" ((uint32_t)LPTE_LOCKED),
754 		  "m" (pte->pte_hi)
755 		: "cr0", "cr1", "cr2", "memory");
756 
757 	*oldhi = (pte->pte_hi & 0xffffffff00000000ULL) | oldhihalf;
758 #endif
759 
760 	return (ret);
761 }
762 
763 static uintptr_t
764 moea64_insert_to_pteg_native(struct lpte *pvo_pt, uintptr_t slotbase,
765     uint64_t mask)
766 {
767 	volatile struct lpte *pt;
768 	uint64_t oldptehi, va;
769 	uintptr_t k;
770 	int i, j;
771 
772 	/* Start at a random slot */
773 	i = mftb() % 8;
774 	for (j = 0; j < 8; j++) {
775 		k = slotbase + (i + j) % 8;
776 		pt = &moea64_pteg_table[k];
777 		/* Invalidate and seize lock only if no bits in mask set */
778 		if (atomic_pte_lock(pt, mask, &oldptehi)) /* Lock obtained */
779 			break;
780 	}
781 
782 	if (j == 8)
783 		return (-1);
784 
785 	if (oldptehi & LPTE_VALID) {
786 		KASSERT(!(oldptehi & LPTE_WIRED), ("Unmapped wired entry"));
787 		/*
788 		 * Need to invalidate old entry completely: see
789 		 * "Modifying a Page Table Entry". Need to reconstruct
790 		 * the virtual address for the outgoing entry to do that.
791 		 */
792 		va = oldptehi >> (ADDR_SR_SHFT - ADDR_API_SHFT64);
793 		if (oldptehi & LPTE_HID)
794 			va = (((k >> 3) ^ moea64_pteg_mask) ^ va) &
795 			    (ADDR_PIDX >> ADDR_PIDX_SHFT);
796 		else
797 			va = ((k >> 3) ^ va) & (ADDR_PIDX >> ADDR_PIDX_SHFT);
798 		va |= (oldptehi & LPTE_AVPN_MASK) <<
799 		    (ADDR_API_SHFT64 - ADDR_PIDX_SHFT);
800 		PTESYNC();
801 		TLBIE(va, oldptehi);
802 		STAT_MOEA64(moea64_pte_valid--);
803 		STAT_MOEA64(moea64_pte_overflow++);
804 	}
805 
806 	/*
807 	 * Update the PTE as per "Adding a Page Table Entry". Lock is released
808 	 * by setting the high doubleworld.
809 	 */
810 	pt->pte_lo = htobe64(pvo_pt->pte_lo);
811 	EIEIO();
812 	pt->pte_hi = htobe64(pvo_pt->pte_hi);
813 	PTESYNC();
814 
815 	/* Keep statistics */
816 	STAT_MOEA64(moea64_pte_valid++);
817 
818 	return (k);
819 }
820 
821 static __always_inline int64_t
822 moea64_pte_insert_locked(struct pvo_entry *pvo, struct lpte *insertpt,
823     uint64_t mask)
824 {
825 	uintptr_t slot;
826 
827 	/*
828 	 * First try primary hash.
829 	 */
830 	slot = moea64_insert_to_pteg_native(insertpt, pvo->pvo_pte.slot,
831 	    mask | LPTE_WIRED | LPTE_LOCKED);
832 	if (slot != -1) {
833 		pvo->pvo_pte.slot = slot;
834 		return (0);
835 	}
836 
837 	/*
838 	 * Now try secondary hash.
839 	 */
840 	pvo->pvo_vaddr ^= PVO_HID;
841 	insertpt->pte_hi ^= LPTE_HID;
842 	pvo->pvo_pte.slot ^= (moea64_pteg_mask << 3);
843 	slot = moea64_insert_to_pteg_native(insertpt, pvo->pvo_pte.slot,
844 	    mask | LPTE_WIRED | LPTE_LOCKED);
845 	if (slot != -1) {
846 		pvo->pvo_pte.slot = slot;
847 		return (0);
848 	}
849 
850 	return (-1);
851 }
852 
853 static int64_t
854 moea64_pte_insert_native(struct pvo_entry *pvo)
855 {
856 	struct lpte insertpt;
857 	int64_t ret;
858 
859 	/* Initialize PTE */
860 	moea64_pte_from_pvo(pvo, &insertpt);
861 
862 	/* Make sure further insertion is locked out during evictions */
863 	rw_rlock(&moea64_eviction_lock);
864 
865 	pvo->pvo_pte.slot &= ~7ULL; /* Base slot address */
866 	ret = moea64_pte_insert_locked(pvo, &insertpt, LPTE_VALID);
867 	if (ret == -1) {
868 		/*
869 		 * Out of luck. Find a PTE to sacrifice.
870 		 */
871 
872 		/* Lock out all insertions for a bit */
873 		if (!rw_try_upgrade(&moea64_eviction_lock)) {
874 			rw_runlock(&moea64_eviction_lock);
875 			rw_wlock(&moea64_eviction_lock);
876 		}
877 		/* Don't evict large pages */
878 		ret = moea64_pte_insert_locked(pvo, &insertpt, LPTE_BIG);
879 		rw_wunlock(&moea64_eviction_lock);
880 		/* No freeable slots in either PTEG? We're hosed. */
881 		if (ret == -1)
882 			panic("moea64_pte_insert: overflow");
883 	} else
884 		rw_runlock(&moea64_eviction_lock);
885 
886 	return (0);
887 }
888 
889 static void *
890 moea64_dump_pmap_native(void *ctx, void *buf, u_long *nbytes)
891 {
892 	struct dump_context *dctx;
893 	u_long ptex, ptex_end;
894 
895 	dctx = (struct dump_context *)ctx;
896 	ptex = dctx->ptex;
897 	ptex_end = ptex + dctx->blksz / sizeof(struct lpte);
898 	ptex_end = MIN(ptex_end, dctx->ptex_end);
899 	*nbytes = (ptex_end - ptex) * sizeof(struct lpte);
900 
901 	if (*nbytes == 0)
902 		return (NULL);
903 
904 	dctx->ptex = ptex_end;
905 	return (__DEVOLATILE(struct lpte *, moea64_pteg_table) + ptex);
906 }
907 
908 static __always_inline uint64_t
909 moea64_vpn_from_pte(uint64_t ptehi, uintptr_t slot)
910 {
911 	uint64_t pgn, pgnlo, vsid;
912 
913 	vsid = (ptehi & LPTE_AVA_MASK) >> LPTE_VSID_SHIFT;
914 	if ((ptehi & LPTE_HID) != 0)
915 		slot ^= (moea64_pteg_mask << 3);
916 	pgnlo = ((vsid & VSID_HASH_MASK) ^ (slot >> 3)) & EA_PAGELO_MASK;
917 	pgn = ((ptehi & LPTE_AVA_PGNHI_MASK) << (EA_PAGELO_SHIFT -
918 	    LPTE_AVA_PGNHI_SHIFT)) | pgnlo;
919 	return ((vsid << 16) | pgn);
920 }
921 
922 static __always_inline int64_t
923 moea64_pte_unset_sp_locked(struct pvo_entry *pvo)
924 {
925 	volatile struct lpte *pt;
926 	uint64_t ptehi, refchg, vpn;
927 	vm_offset_t eva;
928 
929 	refchg = 0;
930 	eva = PVO_VADDR(pvo) + HPT_SP_SIZE;
931 
932 	for (; pvo != NULL && PVO_VADDR(pvo) < eva;
933 	    pvo = RB_NEXT(pvo_tree, &pvo->pvo_pmap->pmap_pvo, pvo)) {
934 		pt = moea64_pteg_table + pvo->pvo_pte.slot;
935 		ptehi = be64toh(pt->pte_hi);
936 		if ((ptehi & LPTE_AVPN_MASK) !=
937 		    moea64_pte_vpn_from_pvo_vpn(pvo)) {
938 			/* Evicted: invalidate new entry */
939 			STAT_MOEA64(moea64_pte_overflow--);
940 			vpn = moea64_vpn_from_pte(ptehi, pvo->pvo_pte.slot);
941 			CTR1(KTR_PMAP, "Evicted page in pte_unset_sp: vpn=%jx",
942 			    (uintmax_t)vpn);
943 			/* Assume evicted page was modified */
944 			refchg |= LPTE_CHG;
945 		} else
946 			vpn = pvo->pvo_vpn;
947 
948 		refchg |= moea64_pte_unset_locked(pt, vpn);
949 	}
950 
951 	return (refchg);
952 }
953 
954 static int64_t
955 moea64_pte_unset_sp_native(struct pvo_entry *pvo)
956 {
957 	uint64_t refchg;
958 
959 	PMAP_LOCK_ASSERT(pvo->pvo_pmap, MA_OWNED);
960 	KASSERT((PVO_VADDR(pvo) & HPT_SP_MASK) == 0,
961 	    ("%s: va %#jx unaligned", __func__, (uintmax_t)PVO_VADDR(pvo)));
962 
963 	rw_rlock(&moea64_eviction_lock);
964 	refchg = moea64_pte_unset_sp_locked(pvo);
965 	rw_runlock(&moea64_eviction_lock);
966 
967 	return (refchg);
968 }
969 
970 static __always_inline int64_t
971 moea64_pte_insert_sp_locked(struct pvo_entry *pvo)
972 {
973 	struct lpte insertpt;
974 	int64_t ret;
975 	vm_offset_t eva;
976 
977 	eva = PVO_VADDR(pvo) + HPT_SP_SIZE;
978 
979 	for (; pvo != NULL && PVO_VADDR(pvo) < eva;
980 	    pvo = RB_NEXT(pvo_tree, &pvo->pvo_pmap->pmap_pvo, pvo)) {
981 		moea64_pte_from_pvo(pvo, &insertpt);
982 		pvo->pvo_pte.slot &= ~7ULL; /* Base slot address */
983 
984 		ret = moea64_pte_insert_locked(pvo, &insertpt, LPTE_VALID);
985 		if (ret == -1) {
986 			/* Lock out all insertions for a bit */
987 			if (!rw_try_upgrade(&moea64_eviction_lock)) {
988 				rw_runlock(&moea64_eviction_lock);
989 				rw_wlock(&moea64_eviction_lock);
990 			}
991 			/* Don't evict large pages */
992 			ret = moea64_pte_insert_locked(pvo, &insertpt,
993 			    LPTE_BIG);
994 			rw_downgrade(&moea64_eviction_lock);
995 			/* No freeable slots in either PTEG? We're hosed. */
996 			if (ret == -1)
997 				panic("moea64_pte_insert_sp: overflow");
998 		}
999 	}
1000 
1001 	return (0);
1002 }
1003 
1004 static int64_t
1005 moea64_pte_insert_sp_native(struct pvo_entry *pvo)
1006 {
1007 	PMAP_LOCK_ASSERT(pvo->pvo_pmap, MA_OWNED);
1008 	KASSERT((PVO_VADDR(pvo) & HPT_SP_MASK) == 0,
1009 	    ("%s: va %#jx unaligned", __func__, (uintmax_t)PVO_VADDR(pvo)));
1010 
1011 	rw_rlock(&moea64_eviction_lock);
1012 	moea64_pte_insert_sp_locked(pvo);
1013 	rw_runlock(&moea64_eviction_lock);
1014 
1015 	return (0);
1016 }
1017 
1018 static int64_t
1019 moea64_pte_replace_sp_native(struct pvo_entry *pvo)
1020 {
1021 	uint64_t refchg;
1022 
1023 	PMAP_LOCK_ASSERT(pvo->pvo_pmap, MA_OWNED);
1024 	KASSERT((PVO_VADDR(pvo) & HPT_SP_MASK) == 0,
1025 	    ("%s: va %#jx unaligned", __func__, (uintmax_t)PVO_VADDR(pvo)));
1026 
1027 	rw_rlock(&moea64_eviction_lock);
1028 	refchg = moea64_pte_unset_sp_locked(pvo);
1029 	moea64_pte_insert_sp_locked(pvo);
1030 	rw_runlock(&moea64_eviction_lock);
1031 
1032 	return (refchg);
1033 }
1034