xref: /netbsd-src/sys/arch/powerpc/booke/e500_tlb.c (revision c38e7cc395b1472a774ff828e46123de44c628e9)
1 /*	$NetBSD: e500_tlb.c,v 1.18 2016/07/11 16:06:52 matt Exp $	*/
2 /*-
3  * Copyright (c) 2010, 2011 The NetBSD Foundation, Inc.
4  * All rights reserved.
5  *
6  * This code is derived from software contributed to The NetBSD Foundation
7  * by Raytheon BBN Technologies Corp and Defense Advanced Research Projects
8  * Agency and which was developed by Matt Thomas of 3am Software Foundry.
9  *
10  * This material is based upon work supported by the Defense Advanced Research
11  * Projects Agency and Space and Naval Warfare Systems Center, Pacific, under
12  * Contract No. N66001-09-C-2073.
13  * Approved for Public Release, Distribution Unlimited
14  *
15  * Redistribution and use in source and binary forms, with or without
16  * modification, are permitted provided that the following conditions
17  * are met:
18  * 1. Redistributions of source code must retain the above copyright
19  *    notice, this list of conditions and the following disclaimer.
20  * 2. Redistributions in binary form must reproduce the above copyright
21  *    notice, this list of conditions and the following disclaimer in the
22  *    documentation and/or other materials provided with the distribution.
23  *
24  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
25  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
26  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
27  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
28  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
29  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
30  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
31  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
32  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
33  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
34  * POSSIBILITY OF SUCH DAMAGE.
35  */
36 
37 #include "opt_ppcparam.h"
38 
39 #define	__PMAP_PRIVATE
40 
41 #include <sys/cdefs.h>
42 
43 __KERNEL_RCSID(0, "$NetBSD: e500_tlb.c,v 1.18 2016/07/11 16:06:52 matt Exp $");
44 
45 #include <sys/param.h>
46 
47 #include <uvm/uvm_extern.h>
48 
49 #include <powerpc/spr.h>
50 #include <powerpc/booke/spr.h>
51 #include <powerpc/booke/cpuvar.h>
52 #include <powerpc/booke/e500reg.h>
53 #include <powerpc/booke/e500var.h>
54 #include <powerpc/booke/pmap.h>
55 
56 struct e500_tlb {
57 	vaddr_t tlb_va;
58 	uint32_t tlb_pte;
59 	uint32_t tlb_asid;
60 	vsize_t tlb_size;
61 };
62 
63 struct e500_hwtlb {
64 	uint32_t hwtlb_mas0;
65 	uint32_t hwtlb_mas1;
66 	uint32_t hwtlb_mas2;
67 	uint32_t hwtlb_mas3;
68 };
69 
70 struct e500_xtlb {
71 	struct e500_tlb e_tlb;
72 	struct e500_hwtlb e_hwtlb;
73 	u_long e_refcnt;
74 };
75 
76 static struct e500_tlb1 {
77 	uint32_t tlb1_maxsize;
78 	uint32_t tlb1_minsize;
79 	u_int tlb1_numentries;
80 	u_int tlb1_numfree;
81 	u_int tlb1_freelist[32];
82 	struct e500_xtlb tlb1_entries[32];
83 } e500_tlb1;
84 
85 static inline register_t mftlb0cfg(void) __pure;
86 static inline register_t mftlb1cfg(void) __pure;
87 
88 static inline register_t
89 mftlb0cfg(void)
90 {
91 	register_t tlb0cfg;
92 	__asm("mfspr %0, %1" : "=r"(tlb0cfg) : "n"(SPR_TLB0CFG));
93 	return tlb0cfg;
94 }
95 
96 static inline register_t
97 mftlb1cfg(void)
98 {
99 	register_t tlb1cfg;
100 	__asm("mfspr %0, %1" : "=r"(tlb1cfg) : "n"(SPR_TLB1CFG));
101 	return tlb1cfg;
102 }
103 
104 static struct e500_tlb
105 hwtlb_to_tlb(const struct e500_hwtlb hwtlb)
106 {
107 	struct e500_tlb tlb;
108 	register_t prot_mask;
109 	u_int prot_shift;
110 
111 	tlb.tlb_va = MAS2_EPN & hwtlb.hwtlb_mas2;
112 	tlb.tlb_size = 1024 << (2 * MASX_TSIZE_GET(hwtlb.hwtlb_mas1));
113 	tlb.tlb_asid = MASX_TID_GET(hwtlb.hwtlb_mas1);
114 	tlb.tlb_pte = (hwtlb.hwtlb_mas2 & MAS2_WIMGE)
115 	    | (hwtlb.hwtlb_mas3 & MAS3_RPN);
116 	if (hwtlb.hwtlb_mas1 & MAS1_TS) {
117 		prot_mask = MAS3_UX|MAS3_UW|MAS3_UR;
118 		prot_shift = PTE_RWX_SHIFT - 1;
119 	} else {
120 		prot_mask = MAS3_SX|MAS3_SW|MAS3_SR;
121 		prot_shift = PTE_RWX_SHIFT;
122 	}
123 	tlb.tlb_pte |= (prot_mask & hwtlb.hwtlb_mas3) << prot_shift;
124 	return tlb;
125 }
126 
127 static inline struct e500_hwtlb
128 hwtlb_read(uint32_t mas0, u_int slot)
129 {
130 	struct e500_hwtlb hwtlb;
131 	register_t tlbcfg;
132 
133 	if (__predict_true(mas0 == MAS0_TLBSEL_TLB0)) {
134 		tlbcfg = mftlb0cfg();
135 	} else if (mas0 == MAS0_TLBSEL_TLB1) {
136 		tlbcfg = mftlb1cfg();
137 	} else {
138 		panic("%s:%d: unexpected MAS0 %#" PRIx32,
139 		    __func__, __LINE__, mas0);
140 	}
141 
142 	/*
143 	 * ESEL is the way we want to look up.
144 	 * If tlbassoc is the same as tlbentries (like in TLB1) then the TLB is
145 	 * fully associative, the entire slot is placed into ESEL.  If tlbassoc
146 	 * is less than the number of tlb entries, the slot is split in two
147 	 * fields.  Since the TLB is M rows by N ways, the lowers bits are for
148 	 * row (MAS2[EPN]) and the upper for the way (MAS1[ESEL]).
149 	 */
150 	const u_int tlbassoc = TLBCFG_ASSOC(tlbcfg);
151 	const u_int tlbentries = TLBCFG_NENTRY(tlbcfg);
152 	const u_int esel_shift =
153 	    __builtin_clz(tlbassoc) - __builtin_clz(tlbentries);
154 
155 	/*
156 	 * Disable interrupts since we don't want anyone else mucking with
157 	 * the MMU Assist registers
158 	 */
159 	const register_t msr = wrtee(0);
160 	const register_t saved_mas0 = mfspr(SPR_MAS0);
161 	mtspr(SPR_MAS0, mas0 | MAS0_ESEL_MAKE(slot >> esel_shift));
162 
163 	if (__predict_true(tlbassoc > tlbentries))
164 		mtspr(SPR_MAS2, slot << PAGE_SHIFT);
165 
166 	/*
167 	 * Now select the entry and grab its contents.
168 	 */
169 	__asm volatile("tlbre");
170 
171 	hwtlb.hwtlb_mas0 = mfspr(SPR_MAS0);
172 	hwtlb.hwtlb_mas1 = mfspr(SPR_MAS1);
173 	hwtlb.hwtlb_mas2 = mfspr(SPR_MAS2);
174 	hwtlb.hwtlb_mas3 = mfspr(SPR_MAS3);
175 
176 	mtspr(SPR_MAS0, saved_mas0);
177 	wrtee(msr);	/* restore interrupts */
178 
179 	return hwtlb;
180 }
181 
182 static inline void
183 hwtlb_write(const struct e500_hwtlb hwtlb, bool needs_sync)
184 {
185 	const register_t msr = wrtee(0);
186 	const uint32_t saved_mas0 = mfspr(SPR_MAS0);
187 
188 	/*
189 	 * Need to always write MAS0 and MAS1
190 	 */
191 	mtspr(SPR_MAS0, hwtlb.hwtlb_mas0);
192 	mtspr(SPR_MAS1, hwtlb.hwtlb_mas1);
193 
194 	/*
195 	 * Only write the VPN/WIMGE if this is in TLB0 or if a valid mapping.
196 	 */
197 	if ((hwtlb.hwtlb_mas0 & MAS0_TLBSEL) == MAS0_TLBSEL_TLB0
198 	    || (hwtlb.hwtlb_mas1 & MAS1_V)) {
199 		mtspr(SPR_MAS2, hwtlb.hwtlb_mas2);
200 	}
201 	/*
202 	 * Only need to write the RPN/prot if we are dealing with a valid
203 	 * mapping.
204 	 */
205 	if (hwtlb.hwtlb_mas1 & MAS1_V) {
206 		mtspr(SPR_MAS3, hwtlb.hwtlb_mas3);
207 		//mtspr(SPR_MAS7, 0);
208 	}
209 
210 #if 0
211 	printf("%s->[%x,%x,%x,%x]\n",
212 	    __func__,
213 	    hwtlb.hwtlb_mas0, hwtlb.hwtlb_mas1,
214 	    hwtlb.hwtlb_mas2, hwtlb.hwtlb_mas3);
215 #endif
216 	__asm volatile("tlbwe");
217 	if (needs_sync) {
218 		__asm volatile("tlbsync\n\tisync\n\tsync");
219 	}
220 
221 	mtspr(SPR_MAS0, saved_mas0);
222 	wrtee(msr);
223 }
224 
225 static struct e500_hwtlb
226 tlb_to_hwtlb(const struct e500_tlb tlb)
227 {
228 	struct e500_hwtlb hwtlb;
229 
230 	KASSERT(trunc_page(tlb.tlb_va) == tlb.tlb_va);
231 	KASSERT(tlb.tlb_size != 0);
232 	KASSERT((tlb.tlb_size & (tlb.tlb_size - 1)) == 0);
233 	const uint32_t prot_mask = tlb.tlb_pte & PTE_RWX_MASK;
234 	if (__predict_true(tlb.tlb_size == PAGE_SIZE)) {
235 		hwtlb.hwtlb_mas0 = 0;
236 		hwtlb.hwtlb_mas1 = MAS1_V | MASX_TSIZE_MAKE(1);
237 		/*
238 		 * A non-zero ASID means this is a user page so mark it as
239 		 * being in the user's address space.
240 		 */
241 		if (tlb.tlb_asid) {
242 			hwtlb.hwtlb_mas1 |= MAS1_TS
243 			    | MASX_TID_MAKE(tlb.tlb_asid);
244 			hwtlb.hwtlb_mas3 = (prot_mask >> (PTE_RWX_SHIFT - 1))
245 			    | ((prot_mask & ~PTE_xX) >> PTE_RWX_SHIFT);
246 			KASSERT(prot_mask & PTE_xR);
247 			KASSERT(hwtlb.hwtlb_mas3 & MAS3_UR);
248 			CTASSERT(MAS3_UR == (PTE_xR >> (PTE_RWX_SHIFT - 1)));
249 			CTASSERT(MAS3_SR == (PTE_xR >> PTE_RWX_SHIFT));
250 		} else {
251 			hwtlb.hwtlb_mas3 = prot_mask >> PTE_RWX_SHIFT;
252 		}
253 		if (tlb.tlb_pte & PTE_UNMODIFIED)
254 			hwtlb.hwtlb_mas3 &= ~(MAS3_UW|MAS3_SW);
255 		if (tlb.tlb_pte & PTE_UNSYNCED)
256 			hwtlb.hwtlb_mas3 &= ~(MAS3_UX|MAS3_SX);
257 	} else {
258 		KASSERT(tlb.tlb_asid == 0);
259 		KASSERT((tlb.tlb_size & 0xaaaaa7ff) == 0);
260 		u_int cntlz = __builtin_clz(tlb.tlb_size);
261 		KASSERT(cntlz & 1);
262 		KASSERT(cntlz <= 19);
263 		hwtlb.hwtlb_mas0 = MAS0_TLBSEL_TLB1;
264 		/*
265 		 * TSIZE is defined (4^TSIZE) Kbytes except a TSIZE of 0 is not
266 		 * allowed.  So 1K would be 0x00000400 giving 21 leading zero
267 		 * bits.  Subtracting the leading number of zero bits from 21
268 		 * and dividing by 2 gives us the number that the MMU wants.
269 		 */
270 		hwtlb.hwtlb_mas1 = MASX_TSIZE_MAKE(((31 - 10) - cntlz) / 2)
271 		    | MAS1_IPROT | MAS1_V;
272 		hwtlb.hwtlb_mas3 = prot_mask >> PTE_RWX_SHIFT;
273 	}
274 	/* We are done with MAS1, on to MAS2 ... */
275 	hwtlb.hwtlb_mas2 = tlb.tlb_va | (tlb.tlb_pte & PTE_WIMGE_MASK);
276 	hwtlb.hwtlb_mas3 |= tlb.tlb_pte & PTE_RPN_MASK;
277 
278 	return hwtlb;
279 }
280 
281 void *
282 e500_tlb1_fetch(size_t slot)
283 {
284 	struct e500_tlb1 * const tlb1 = &e500_tlb1;
285 
286 	return &tlb1->tlb1_entries[slot].e_hwtlb;
287 }
288 
289 void
290 e500_tlb1_sync(void)
291 {
292 	struct e500_tlb1 * const tlb1 = &e500_tlb1;
293 	for (u_int slot = 1; slot < tlb1->tlb1_numentries; slot++) {
294 		const struct e500_hwtlb * const new_hwtlb =
295 		    &tlb1->tlb1_entries[slot].e_hwtlb;
296 		const struct e500_hwtlb old_hwtlb =
297 		    hwtlb_read(MAS0_TLBSEL_TLB1, slot);
298 #define CHANGED(n,o,f)	((n)->f != (o).f)
299 		bool mas1_changed_p = CHANGED(new_hwtlb, old_hwtlb, hwtlb_mas1);
300 		bool mas2_changed_p = CHANGED(new_hwtlb, old_hwtlb, hwtlb_mas2);
301 		bool mas3_changed_p = CHANGED(new_hwtlb, old_hwtlb, hwtlb_mas3);
302 #undef CHANGED
303 		bool new_valid_p = (new_hwtlb->hwtlb_mas1 & MAS1_V) != 0;
304 		bool old_valid_p = (old_hwtlb.hwtlb_mas1 & MAS1_V) != 0;
305 		if ((new_valid_p || old_valid_p)
306 		    && (mas1_changed_p
307 			|| (new_valid_p
308 			    && (mas2_changed_p || mas3_changed_p))))
309 			hwtlb_write(*new_hwtlb, true);
310 	}
311 }
312 
313 static int
314 e500_alloc_tlb1_entry(void)
315 {
316 	struct e500_tlb1 * const tlb1 = &e500_tlb1;
317 
318 	if (tlb1->tlb1_numfree == 0)
319 		return -1;
320 	const u_int slot = tlb1->tlb1_freelist[--tlb1->tlb1_numfree];
321 	KASSERT((tlb1->tlb1_entries[slot].e_hwtlb.hwtlb_mas1 & MAS1_V) == 0);
322 	tlb1->tlb1_entries[slot].e_hwtlb.hwtlb_mas0 =
323 	    MAS0_TLBSEL_TLB1 | __SHIFTIN(slot, MAS0_ESEL);
324 	return (int)slot;
325 }
326 
327 static void
328 e500_free_tlb1_entry(struct e500_xtlb *xtlb, u_int slot, bool needs_sync)
329 {
330 	struct e500_tlb1 * const tlb1 = &e500_tlb1;
331 	KASSERT(slot < tlb1->tlb1_numentries);
332 	KASSERT(&tlb1->tlb1_entries[slot] == xtlb);
333 
334 	KASSERT(xtlb->e_hwtlb.hwtlb_mas0 == (MAS0_TLBSEL_TLB1|__SHIFTIN(slot, MAS0_ESEL)));
335 	xtlb->e_hwtlb.hwtlb_mas1 &= ~(MAS1_V|MAS1_IPROT);
336 	hwtlb_write(xtlb->e_hwtlb, needs_sync);
337 
338 	const register_t msr = wrtee(0);
339 	tlb1->tlb1_freelist[tlb1->tlb1_numfree++] = slot;
340 	wrtee(msr);
341 }
342 
343 static tlb_asid_t
344 e500_tlb_get_asid(void)
345 {
346 	return mfspr(SPR_PID0);
347 }
348 
349 static void
350 e500_tlb_set_asid(tlb_asid_t asid)
351 {
352 	mtspr(SPR_PID0, asid);
353 }
354 
355 static void
356 e500_tlb_invalidate_all(void)
357 {
358 	/*
359 	 * This does a flash invalidate of all entries in TLB0.
360 	 * We don't touch TLB1 since we don't expect those to be volatile.
361 	 */
362 #if 1
363 	__asm volatile("tlbivax\t0, %0" :: "b"(4));	/* INV_ALL */
364 	__asm volatile("tlbsync\n\tisync\n\tsync");
365 #else
366 	mtspr(SPR_MMUCSR0, MMUCSR0_TLB0_FI);
367 	while (mfspr(SPR_MMUCSR0) != 0)
368 		;
369 #endif
370 }
371 
372 static void
373 e500_tlb_invalidate_globals(void)
374 {
375 #if defined(MULTIPROCESSOR)
376 	e500_tlb_invalidate_all();
377 #else	/* !MULTIPROCESSOR */
378 	const size_t tlbassoc = TLBCFG_ASSOC(mftlb0cfg());
379 	const size_t tlbentries = TLBCFG_NENTRY(mftlb0cfg());
380 	const size_t max_epn = (tlbentries / tlbassoc) << PAGE_SHIFT;
381 	const vaddr_t kstack_lo = (uintptr_t)curlwp->l_addr;
382 	const vaddr_t kstack_hi = kstack_lo + USPACE - 1;
383 	const vaddr_t epn_kstack_lo = kstack_lo & (max_epn - 1);
384 	const vaddr_t epn_kstack_hi = kstack_hi & (max_epn - 1);
385 
386 	const register_t msr = wrtee(0);
387 	for (size_t assoc = 0; assoc < tlbassoc; assoc++) {
388 		mtspr(SPR_MAS0, MAS0_ESEL_MAKE(assoc) | MAS0_TLBSEL_TLB0);
389 		for (size_t epn = 0; epn < max_epn; epn += PAGE_SIZE) {
390 			mtspr(SPR_MAS2, epn);
391 			__asm volatile("tlbre");
392 			uint32_t mas1 = mfspr(SPR_MAS1);
393 
394 			/*
395 			 * Make sure this is a valid kernel entry first.
396 			 */
397 			if ((mas1 & (MAS1_V|MAS1_TID|MAS1_TS)) != MAS1_V)
398 				continue;
399 
400 			/*
401 			 * We have a valid kernel TLB entry.  But if it matches
402 			 * the stack we are currently running on, it would
403 			 * unwise to invalidate it.  First see if the epn
404 			 * overlaps the stack.  If it does then get the
405 			 * VA and see if it really is part of the stack.
406 			 */
407 			if (epn_kstack_lo < epn_kstack_hi
408 			    ? (epn_kstack_lo <= epn && epn <= epn_kstack_hi)
409 			    : (epn <= epn_kstack_hi || epn_kstack_lo <= epn)) {
410 				const uint32_t mas2_epn =
411 				    mfspr(SPR_MAS2) & MAS2_EPN;
412 				if (kstack_lo <= mas2_epn
413 				    && mas2_epn <= kstack_hi)
414 					continue;
415 			}
416 			mtspr(SPR_MAS1, mas1 ^ MAS1_V);
417 			__asm volatile("tlbwe");
418 		}
419 	}
420 	__asm volatile("isync\n\tsync");
421 	wrtee(msr);
422 #endif	/* MULTIPROCESSOR */
423 }
424 
425 static void
426 e500_tlb_invalidate_asids(tlb_asid_t asid_lo, tlb_asid_t asid_hi)
427 {
428 #if defined(MULTIPROCESSOR)
429 	e500_tlb_invalidate_all();
430 #else	/* !MULTIPROCESSOR */
431 	const size_t tlbassoc = TLBCFG_ASSOC(mftlb0cfg());
432 	const size_t tlbentries = TLBCFG_NENTRY(mftlb0cfg());
433 	const size_t max_epn = (tlbentries / tlbassoc) << PAGE_SHIFT;
434 
435 	asid_lo = __SHIFTIN(asid_lo, MAS1_TID);
436 	asid_hi = __SHIFTIN(asid_hi, MAS1_TID);
437 
438 	const register_t msr = wrtee(0);
439 	for (size_t assoc = 0; assoc < tlbassoc; assoc++) {
440 		mtspr(SPR_MAS0, MAS0_ESEL_MAKE(assoc) | MAS0_TLBSEL_TLB0);
441 		for (size_t epn = 0; epn < max_epn; epn += PAGE_SIZE) {
442 			mtspr(SPR_MAS2, epn);
443 			__asm volatile("tlbre");
444 			const uint32_t mas1 = mfspr(SPR_MAS1);
445 			/*
446 			 * If this is a valid entry for AS space 1 and
447 			 * its asid matches the constraints of the caller,
448 			 * clear its valid bit.
449 			 */
450 			if ((mas1 & (MAS1_V|MAS1_TS)) == (MAS1_V|MAS1_TS)
451 			    && asid_lo <= (mas1 & MAS1_TID)
452 			    && (mas1 & MAS1_TID) <= asid_hi) {
453 				mtspr(SPR_MAS1, mas1 ^ MAS1_V);
454 #if 0
455 				printf("%s[%zu,%zu]->[%x]\n",
456 				    __func__, assoc, epn, mas1);
457 #endif
458 				__asm volatile("tlbwe");
459 			}
460 		}
461 	}
462 	__asm volatile("isync\n\tsync");
463 	wrtee(msr);
464 #endif	/* MULTIPROCESSOR */
465 }
466 
467 static u_int
468 e500_tlb_record_asids(u_long *bitmap, tlb_asid_t asid_max)
469 {
470 	const size_t tlbassoc = TLBCFG_ASSOC(mftlb0cfg());
471 	const size_t tlbentries = TLBCFG_NENTRY(mftlb0cfg());
472 	const size_t max_epn = (tlbentries / tlbassoc) << PAGE_SHIFT;
473 	const size_t nbits = 8 * sizeof(bitmap[0]);
474 	u_int found = 0;
475 
476 	const register_t msr = wrtee(0);
477 	for (size_t assoc = 0; assoc < tlbassoc; assoc++) {
478 		mtspr(SPR_MAS0, MAS0_ESEL_MAKE(assoc) | MAS0_TLBSEL_TLB0);
479 		for (size_t epn = 0; epn < max_epn; epn += PAGE_SIZE) {
480 			mtspr(SPR_MAS2, epn);
481 			__asm volatile("tlbre");
482 			const uint32_t mas1 = mfspr(SPR_MAS1);
483 			/*
484 			 * If this is a valid entry for AS space 1 and
485 			 * its asid matches the constraints of the caller,
486 			 * clear its valid bit.
487 			 */
488 			if ((mas1 & (MAS1_V|MAS1_TS)) == (MAS1_V|MAS1_TS)) {
489 				const uint32_t asid = MASX_TID_GET(mas1);
490 				const u_int i = asid / nbits;
491 				const u_long mask = 1UL << (asid & (nbits - 1));
492 				if ((bitmap[i] & mask) == 0) {
493 					bitmap[i] |= mask;
494 					found++;
495 				}
496 			}
497 		}
498 	}
499 	wrtee(msr);
500 
501 	return found;
502 }
503 
504 static void
505 e500_tlb_invalidate_addr(vaddr_t va, tlb_asid_t asid)
506 {
507 	KASSERT((va & PAGE_MASK) == 0);
508 	/*
509 	 * Bits 60 & 61 have meaning
510 	 */
511 	if (asid == KERNEL_PID) {
512 		/*
513 		 * For data accesses, the context-synchronizing instruction
514 		 * before tlbwe or tlbivax ensures that all memory accesses
515 		 * due to preceding instructions have completed to a point
516 		 * at which they have reported all exceptions they will cause.
517 		 */
518 		__asm volatile("isync");
519 	}
520 	__asm volatile("tlbivax\t0, %0" :: "b"(va));
521 	__asm volatile("tlbsync");
522 	__asm volatile("tlbsync");	/* Why? */
523 	if (asid == KERNEL_PID) {
524 		/*
525 		 * The context-synchronizing instruction after tlbwe or tlbivax
526 		 * ensures that subsequent accesses (data and instruction) use
527 		 * the updated value in any TLB entries affected.
528 		 */
529 		__asm volatile("isync\n\tsync");
530 	}
531 }
532 
533 static bool
534 e500_tlb_update_addr(vaddr_t va, tlb_asid_t asid, pt_entry_t pte, bool insert)
535 {
536 #if defined(MULTIPROCESSOR)
537 	e500_tlb_invalidate_addr(va, asid);
538 	return true;
539 #else	/* !MULTIPROCESSOR */
540 	struct e500_hwtlb hwtlb = tlb_to_hwtlb(
541 	    (struct e500_tlb){ .tlb_va = va, .tlb_asid = asid,
542 		.tlb_size = PAGE_SIZE, .tlb_pte = pte,});
543 
544 	register_t msr = wrtee(0);
545 	mtspr(SPR_MAS6, asid ? __SHIFTIN(asid, MAS6_SPID0) | MAS6_SAS : 0);
546 	__asm volatile("tlbsx 0, %0" :: "b"(va));
547 	register_t mas1 = mfspr(SPR_MAS1);
548 	if ((mas1 & MAS1_V) == 0) {
549 		if (!insert) {
550 			wrtee(msr);
551 #if 0
552 			printf("%s(%#lx,%#x,%#x,%x)<no update>\n",
553 			    __func__, va, asid, pte, insert);
554 #endif
555 			return false;
556 		}
557 		mas1 = hwtlb.hwtlb_mas1 | MAS1_V;
558 		mtspr(SPR_MAS1, mas1);
559 	}
560 	mtspr(SPR_MAS2, hwtlb.hwtlb_mas2);
561 	mtspr(SPR_MAS3, hwtlb.hwtlb_mas3);
562 	//mtspr(SPR_MAS7, 0);
563 	__asm volatile("tlbwe");
564 	if (asid == KERNEL_PID)
565 		__asm volatile("isync\n\tsync");
566 	wrtee(msr);
567 #if 0
568 	if (asid)
569 	printf("%s(%#lx,%#x,%#x,%x)->[%x,%x,%x]\n",
570 	    __func__, va, asid, pte, insert,
571 	    hwtlb.hwtlb_mas1, hwtlb.hwtlb_mas2, hwtlb.hwtlb_mas3);
572 #endif
573 	return (mas1 & MAS1_V) != 0;
574 #endif	/* MULTIPROCESSOR */
575 }
576 
577 static void
578 e500_tlb_write_entry(size_t index, const struct tlbmask *tlb)
579 {
580 }
581 
582 static void
583 e500_tlb_read_entry(size_t index, struct tlbmask *tlb)
584 {
585 }
586 
587 static void
588 e500_tlb_dump(void (*pr)(const char *, ...))
589 {
590 	const size_t tlbassoc = TLBCFG_ASSOC(mftlb0cfg());
591 	const size_t tlbentries = TLBCFG_NENTRY(mftlb0cfg());
592 	const size_t max_epn = (tlbentries / tlbassoc) << PAGE_SHIFT;
593 	const uint32_t saved_mas0 = mfspr(SPR_MAS0);
594 	size_t valid = 0;
595 
596 	if (pr == NULL)
597 		pr = printf;
598 
599 	const register_t msr = wrtee(0);
600 	for (size_t assoc = 0; assoc < tlbassoc; assoc++) {
601 		struct e500_hwtlb hwtlb;
602 		hwtlb.hwtlb_mas0 = MAS0_ESEL_MAKE(assoc) | MAS0_TLBSEL_TLB0;
603 		mtspr(SPR_MAS0, hwtlb.hwtlb_mas0);
604 		for (size_t epn = 0; epn < max_epn; epn += PAGE_SIZE) {
605 			mtspr(SPR_MAS2, epn);
606 			__asm volatile("tlbre");
607 			hwtlb.hwtlb_mas1 = mfspr(SPR_MAS1);
608 			/*
609 			 * If this is a valid entry for AS space 1 and
610 			 * its asid matches the constraints of the caller,
611 			 * clear its valid bit.
612 			 */
613 			if (hwtlb.hwtlb_mas1 & MAS1_V) {
614 				hwtlb.hwtlb_mas2 = mfspr(SPR_MAS2);
615 				hwtlb.hwtlb_mas3 = mfspr(SPR_MAS3);
616 				struct e500_tlb tlb = hwtlb_to_tlb(hwtlb);
617 				(*pr)("[%zu,%zu]->[%x,%x,%x]",
618 				    assoc, atop(epn),
619 				    hwtlb.hwtlb_mas1,
620 				    hwtlb.hwtlb_mas2,
621 				    hwtlb.hwtlb_mas3);
622 				(*pr)(": VA=%#lx size=4KB asid=%u pte=%x",
623 				    tlb.tlb_va, tlb.tlb_asid, tlb.tlb_pte);
624 				(*pr)(" (RPN=%#x,%s%s%s%s%s,%s%s%s%s%s)\n",
625 				    tlb.tlb_pte & PTE_RPN_MASK,
626 				    tlb.tlb_pte & PTE_xR ? "R" : "",
627 				    tlb.tlb_pte & PTE_xW ? "W" : "",
628 				    tlb.tlb_pte & PTE_UNMODIFIED ? "*" : "",
629 				    tlb.tlb_pte & PTE_xX ? "X" : "",
630 				    tlb.tlb_pte & PTE_UNSYNCED ? "*" : "",
631 				    tlb.tlb_pte & PTE_W ? "W" : "",
632 				    tlb.tlb_pte & PTE_I ? "I" : "",
633 				    tlb.tlb_pte & PTE_M ? "M" : "",
634 				    tlb.tlb_pte & PTE_G ? "G" : "",
635 				    tlb.tlb_pte & PTE_E ? "E" : "");
636 				valid++;
637 			}
638 		}
639 	}
640 	mtspr(SPR_MAS0, saved_mas0);
641 	wrtee(msr);
642 	(*pr)("%s: %zu valid entries\n", __func__, valid);
643 }
644 
645 static void
646 e500_tlb_walk(void *ctx, bool (*func)(void *, vaddr_t, uint32_t, uint32_t))
647 {
648 	const size_t tlbassoc = TLBCFG_ASSOC(mftlb0cfg());
649 	const size_t tlbentries = TLBCFG_NENTRY(mftlb0cfg());
650 	const size_t max_epn = (tlbentries / tlbassoc) << PAGE_SHIFT;
651 	const uint32_t saved_mas0 = mfspr(SPR_MAS0);
652 
653 	const register_t msr = wrtee(0);
654 	for (size_t assoc = 0; assoc < tlbassoc; assoc++) {
655 		struct e500_hwtlb hwtlb;
656 		hwtlb.hwtlb_mas0 = MAS0_ESEL_MAKE(assoc) | MAS0_TLBSEL_TLB0;
657 		mtspr(SPR_MAS0, hwtlb.hwtlb_mas0);
658 		for (size_t epn = 0; epn < max_epn; epn += PAGE_SIZE) {
659 			mtspr(SPR_MAS2, epn);
660 			__asm volatile("tlbre");
661 			hwtlb.hwtlb_mas1 = mfspr(SPR_MAS1);
662 			if (hwtlb.hwtlb_mas1 & MAS1_V) {
663 				hwtlb.hwtlb_mas2 = mfspr(SPR_MAS2);
664 				hwtlb.hwtlb_mas3 = mfspr(SPR_MAS3);
665 				struct e500_tlb tlb = hwtlb_to_tlb(hwtlb);
666 				if (!(*func)(ctx, tlb.tlb_va, tlb.tlb_asid,
667 				    tlb.tlb_pte))
668 					break;
669 			}
670 		}
671 	}
672 	mtspr(SPR_MAS0, saved_mas0);
673 	wrtee(msr);
674 }
675 
676 static struct e500_xtlb *
677 e500_tlb_lookup_xtlb_pa(vaddr_t pa, u_int *slotp)
678 {
679 	struct e500_tlb1 * const tlb1 = &e500_tlb1;
680 	struct e500_xtlb *xtlb = tlb1->tlb1_entries;
681 
682 	/*
683 	 * See if we have a TLB entry for the pa.
684 	 */
685 	for (u_int i = 0; i < tlb1->tlb1_numentries; i++, xtlb++) {
686 		psize_t mask = ~(xtlb->e_tlb.tlb_size - 1);
687 		if ((xtlb->e_hwtlb.hwtlb_mas1 & MAS1_V)
688 		    && ((pa ^ xtlb->e_tlb.tlb_pte) & mask) == 0) {
689 			if (slotp != NULL)
690 				*slotp = i;
691 			return xtlb;
692 		}
693 	}
694 
695 	return NULL;
696 }
697 
698 struct e500_xtlb *
699 e500_tlb_lookup_xtlb(vaddr_t va, u_int *slotp)
700 {
701 	struct e500_tlb1 * const tlb1 = &e500_tlb1;
702 	struct e500_xtlb *xtlb = tlb1->tlb1_entries;
703 
704 	/*
705 	 * See if we have a TLB entry for the va.
706 	 */
707 	for (u_int i = 0; i < tlb1->tlb1_numentries; i++, xtlb++) {
708 		vsize_t mask = ~(xtlb->e_tlb.tlb_size - 1);
709 		if ((xtlb->e_hwtlb.hwtlb_mas1 & MAS1_V)
710 		    && ((va ^ xtlb->e_tlb.tlb_va) & mask) == 0) {
711 			if (slotp != NULL)
712 				*slotp = i;
713 			return xtlb;
714 		}
715 	}
716 
717 	return NULL;
718 }
719 
720 static struct e500_xtlb *
721 e500_tlb_lookup_xtlb2(vaddr_t va, vsize_t len)
722 {
723 	struct e500_tlb1 * const tlb1 = &e500_tlb1;
724 	struct e500_xtlb *xtlb = tlb1->tlb1_entries;
725 
726 	/*
727 	 * See if we have a TLB entry for the pa.
728 	 */
729 	for (u_int i = 0; i < tlb1->tlb1_numentries; i++, xtlb++) {
730 		vsize_t mask = ~(xtlb->e_tlb.tlb_size - 1);
731 		if ((xtlb->e_hwtlb.hwtlb_mas1 & MAS1_V)
732 		    && ((va ^ xtlb->e_tlb.tlb_va) & mask) == 0
733 		    && (((va + len - 1) ^ va) & mask) == 0) {
734 			return xtlb;
735 		}
736 	}
737 
738 	return NULL;
739 }
740 
741 static void *
742 e500_tlb_mapiodev(paddr_t pa, psize_t len, bool prefetchable)
743 {
744 	struct e500_xtlb * const xtlb = e500_tlb_lookup_xtlb_pa(pa, NULL);
745 
746 	/*
747 	 * See if we have a TLB entry for the pa.  If completely falls within
748 	 * mark the reference and return the pa.  But only if the tlb entry
749 	 * is not cacheable.
750 	 */
751 	if (xtlb
752 	    && (prefetchable
753 		|| (xtlb->e_tlb.tlb_pte & PTE_WIG) == (PTE_I|PTE_G))) {
754 		xtlb->e_refcnt++;
755 		return (void *) (xtlb->e_tlb.tlb_va
756 		    + pa - (xtlb->e_tlb.tlb_pte & PTE_RPN_MASK));
757 	}
758 	return NULL;
759 }
760 
761 static void
762 e500_tlb_unmapiodev(vaddr_t va, vsize_t len)
763 {
764 	if (va < VM_MIN_KERNEL_ADDRESS || VM_MAX_KERNEL_ADDRESS <= va) {
765 		struct e500_xtlb * const xtlb = e500_tlb_lookup_xtlb(va, NULL);
766 		if (xtlb)
767 			xtlb->e_refcnt--;
768 	}
769 }
770 
771 static int
772 e500_tlb_ioreserve(vaddr_t va, vsize_t len, pt_entry_t pte)
773 {
774 	struct e500_tlb1 * const tlb1 = &e500_tlb1;
775 	struct e500_xtlb *xtlb;
776 
777 	KASSERT(len & 0x55555000);
778 	KASSERT((len & ~0x55555000) == 0);
779 	KASSERT(len >= PAGE_SIZE);
780 	KASSERT((len & (len - 1)) == 0);
781 	KASSERT((va & (len - 1)) == 0);
782 	KASSERT(((pte & PTE_RPN_MASK) & (len - 1)) == 0);
783 
784 	if ((xtlb = e500_tlb_lookup_xtlb2(va, len)) != NULL) {
785 		psize_t mask __diagused = ~(xtlb->e_tlb.tlb_size - 1);
786 		KASSERT(len <= xtlb->e_tlb.tlb_size);
787 		KASSERT((pte & mask) == (xtlb->e_tlb.tlb_pte & mask));
788 		xtlb->e_refcnt++;
789 		return 0;
790 	}
791 
792 	const int slot = e500_alloc_tlb1_entry();
793 	if (slot < 0)
794 		return ENOMEM;
795 
796 	xtlb = &tlb1->tlb1_entries[slot];
797 	xtlb->e_tlb.tlb_va = va;
798 	xtlb->e_tlb.tlb_size = len;
799 	xtlb->e_tlb.tlb_pte = pte;
800 	xtlb->e_tlb.tlb_asid = KERNEL_PID;
801 
802 	xtlb->e_hwtlb = tlb_to_hwtlb(xtlb->e_tlb);
803 	xtlb->e_hwtlb.hwtlb_mas0 |= __SHIFTIN(slot, MAS0_ESEL);
804 	hwtlb_write(xtlb->e_hwtlb, true);
805 
806 #if defined(MULTIPROCESSOR)
807 	cpu_send_ipi(IPI_DST_NOTME, IPI_TLB1SYNC);
808 #endif
809 
810 	return 0;
811 }
812 
813 static int
814 e500_tlb_iorelease(vaddr_t va)
815 {
816 	u_int slot;
817 	struct e500_xtlb * const xtlb = e500_tlb_lookup_xtlb(va, &slot);
818 
819 	if (xtlb == NULL)
820 		return ENOENT;
821 
822 	if (xtlb->e_refcnt)
823 		return EBUSY;
824 
825 	e500_free_tlb1_entry(xtlb, slot, true);
826 
827 #if defined(MULTIPROCESSOR)
828 	cpu_send_ipi(IPI_DST_NOTME, IPI_TLB1SYNC);
829 #endif
830 
831 	return 0;
832 }
833 
834 static u_int
835 e500_tlbmemmap(paddr_t memstart, psize_t memsize, struct e500_tlb1 *tlb1)
836 {
837 	u_int slotmask = 0;
838 	u_int slots = 0, nextslot = 0;
839 	KASSERT(tlb1->tlb1_numfree > 1);
840 	KASSERT(((memstart + memsize - 1) & -memsize) == memstart);
841 	for (paddr_t lastaddr = memstart; 0 < memsize; ) {
842 		u_int cnt = __builtin_clz(memsize);
843 		psize_t size = min(1UL << (31 - (cnt | 1)), tlb1->tlb1_maxsize);
844 		slots += memsize / size;
845 		if (slots > 4)
846 			panic("%s: %d: can't map memory (%#lx) into TLB1: %s",
847 			    __func__, __LINE__, memsize, "too fragmented");
848 		if (slots > tlb1->tlb1_numfree - 1)
849 			panic("%s: %d: can't map memory (%#lx) into TLB1: %s",
850 			    __func__, __LINE__, memsize,
851 			    "insufficent TLB entries");
852 		for (; nextslot < slots; nextslot++) {
853 			const u_int freeslot = e500_alloc_tlb1_entry();
854 			struct e500_xtlb * const xtlb =
855 			    &tlb1->tlb1_entries[freeslot];
856 			xtlb->e_tlb.tlb_asid = KERNEL_PID;
857 			xtlb->e_tlb.tlb_size = size;
858 			xtlb->e_tlb.tlb_va = lastaddr;
859 			xtlb->e_tlb.tlb_pte = lastaddr
860 			    | PTE_M | PTE_xX | PTE_xW | PTE_xR;
861 			lastaddr += size;
862 			memsize -= size;
863 			slotmask |= 1 << (31 - freeslot); /* clz friendly */
864 		}
865 	}
866 
867 #if defined(MULTIPROCESSOR)
868 	cpu_send_ipi(IPI_DST_NOTME, IPI_TLB1SYNC);
869 #endif
870 
871 	return nextslot;
872 }
873 
874 static const struct tlb_md_ops e500_tlb_ops = {
875 	.md_tlb_get_asid = e500_tlb_get_asid,
876 	.md_tlb_set_asid = e500_tlb_set_asid,
877 	.md_tlb_invalidate_all = e500_tlb_invalidate_all,
878 	.md_tlb_invalidate_globals = e500_tlb_invalidate_globals,
879 	.md_tlb_invalidate_asids = e500_tlb_invalidate_asids,
880 	.md_tlb_invalidate_addr = e500_tlb_invalidate_addr,
881 	.md_tlb_update_addr = e500_tlb_update_addr,
882 	.md_tlb_record_asids = e500_tlb_record_asids,
883 	.md_tlb_write_entry = e500_tlb_write_entry,
884 	.md_tlb_read_entry = e500_tlb_read_entry,
885 	.md_tlb_dump = e500_tlb_dump,
886 	.md_tlb_walk = e500_tlb_walk,
887 };
888 
889 static const struct tlb_md_io_ops e500_tlb_io_ops = {
890 	.md_tlb_mapiodev = e500_tlb_mapiodev,
891 	.md_tlb_unmapiodev = e500_tlb_unmapiodev,
892 	.md_tlb_ioreserve = e500_tlb_ioreserve,
893 	.md_tlb_iorelease = e500_tlb_iorelease,
894 };
895 
896 void
897 e500_tlb_init(vaddr_t endkernel, psize_t memsize)
898 {
899 	struct e500_tlb1 * const tlb1 = &e500_tlb1;
900 
901 #if 0
902 	register_t mmucfg = mfspr(SPR_MMUCFG);
903 	register_t mas4 = mfspr(SPR_MAS4);
904 #endif
905 
906 	const uint32_t tlb1cfg = mftlb1cfg();
907 	tlb1->tlb1_numentries = TLBCFG_NENTRY(tlb1cfg);
908 	KASSERT(tlb1->tlb1_numentries <= __arraycount(tlb1->tlb1_entries));
909 	/*
910 	 * Limit maxsize to 1G since 4G isn't really useful to us.
911 	 */
912 	tlb1->tlb1_minsize = 1024 << (2 * TLBCFG_MINSIZE(tlb1cfg));
913 	tlb1->tlb1_maxsize = 1024 << (2 * min(10, TLBCFG_MAXSIZE(tlb1cfg)));
914 
915 #ifdef VERBOSE_INITPPC
916 	printf(" tlb1cfg=%#x numentries=%u minsize=%#xKB maxsize=%#xKB",
917 	    tlb1cfg, tlb1->tlb1_numentries, tlb1->tlb1_minsize >> 10,
918 	    tlb1->tlb1_maxsize >> 10);
919 #endif
920 
921 	/*
922 	 * Let's see what's in TLB1 and we need to invalidate any entry that
923 	 * would fit within the kernel's mapped address space.
924 	 */
925 	psize_t memmapped = 0;
926 	for (u_int i = 0; i < tlb1->tlb1_numentries; i++) {
927 		struct e500_xtlb * const xtlb = &tlb1->tlb1_entries[i];
928 
929 		xtlb->e_hwtlb = hwtlb_read(MAS0_TLBSEL_TLB1, i);
930 
931 		if ((xtlb->e_hwtlb.hwtlb_mas1 & MAS1_V) == 0) {
932 			tlb1->tlb1_freelist[tlb1->tlb1_numfree++] = i;
933 #ifdef VERBOSE_INITPPC
934 			printf(" TLB1[%u]=<unused>", i);
935 #endif
936 			continue;
937 		}
938 
939 		xtlb->e_tlb = hwtlb_to_tlb(xtlb->e_hwtlb);
940 #ifdef VERBOSE_INITPPC
941 		printf(" TLB1[%u]=<%#lx,%#lx,%#x,%#x>",
942 		    i, xtlb->e_tlb.tlb_va, xtlb->e_tlb.tlb_size,
943 		    xtlb->e_tlb.tlb_asid, xtlb->e_tlb.tlb_pte);
944 #endif
945 		if ((VM_MIN_KERNEL_ADDRESS <= xtlb->e_tlb.tlb_va
946 		    && xtlb->e_tlb.tlb_va < VM_MAX_KERNEL_ADDRESS)
947 		    || (xtlb->e_tlb.tlb_va < VM_MIN_KERNEL_ADDRESS
948 		        && VM_MIN_KERNEL_ADDRESS <
949 			   xtlb->e_tlb.tlb_va + xtlb->e_tlb.tlb_size)) {
950 #ifdef VERBOSE_INITPPC
951 			printf("free");
952 #endif
953 			e500_free_tlb1_entry(xtlb, i, false);
954 #ifdef VERBOSE_INITPPC
955 			printf("d");
956 #endif
957 			continue;
958 		}
959 		if ((xtlb->e_hwtlb.hwtlb_mas1 & MAS1_IPROT) == 0) {
960 			xtlb->e_hwtlb.hwtlb_mas1 |= MAS1_IPROT;
961 			hwtlb_write(xtlb->e_hwtlb, false);
962 #ifdef VERBOSE_INITPPC
963 			printf("+iprot");
964 #endif
965 		}
966 		if (xtlb->e_tlb.tlb_pte & PTE_I)
967 			continue;
968 
969 		if (xtlb->e_tlb.tlb_va == 0
970 		    || xtlb->e_tlb.tlb_va + xtlb->e_tlb.tlb_size <= memsize) {
971 			memmapped += xtlb->e_tlb.tlb_size;
972 			/*
973 			 * Let make sure main memory is setup so it's memory
974 			 * coherent.  For some reason u-boot doesn't set it up
975 			 * that way.
976 			 */
977 			if ((xtlb->e_hwtlb.hwtlb_mas2 & MAS2_M) == 0) {
978 				xtlb->e_hwtlb.hwtlb_mas2 |= MAS2_M;
979 				hwtlb_write(xtlb->e_hwtlb, true);
980 			}
981 		}
982 	}
983 
984 	cpu_md_ops.md_tlb_ops = &e500_tlb_ops;
985 	cpu_md_ops.md_tlb_io_ops = &e500_tlb_io_ops;
986 
987 	if (__predict_false(memmapped < memsize)) {
988 		/*
989 		 * Let's see how many TLB entries are needed to map memory.
990 		 */
991 		u_int slotmask = e500_tlbmemmap(0, memsize, tlb1);
992 
993 		/*
994 		 * To map main memory into the TLB, we need to flush any
995 		 * existing entries from the TLB that overlap the virtual
996 		 * address space needed to map physical memory.  That may
997 		 * include the entries for the pages currently used by the
998 		 * stack or that we are executing.  So to avoid problems, we
999 		 * are going to temporarily map the kernel and stack into AS 1,
1000 		 * switch to it, and clear out the TLB entries from AS 0,
1001 		 * install the new TLB entries to map memory, and then switch
1002 		 * back to AS 0 and free the temp entry used for AS1.
1003 		 */
1004 		u_int b = __builtin_clz(endkernel);
1005 
1006 		/*
1007 		 * If the kernel doesn't end on a clean power of 2, we need
1008 		 * to round the size up (by decrementing the number of leading
1009 		 * zero bits).  If the size isn't a power of 4KB, decrement
1010 		 * again to make it one.
1011 		 */
1012 		if (endkernel & (endkernel - 1))
1013 			b--;
1014 		if ((b & 1) == 0)
1015 			b--;
1016 
1017 		/*
1018 		 * Create a TLB1 mapping for the kernel in AS1.
1019 		 */
1020 		const u_int kslot = e500_alloc_tlb1_entry();
1021 		struct e500_xtlb * const kxtlb = &tlb1->tlb1_entries[kslot];
1022 		kxtlb->e_tlb.tlb_va = 0;
1023 		kxtlb->e_tlb.tlb_size = 1UL << (31 - b);
1024 		kxtlb->e_tlb.tlb_pte = PTE_M|PTE_xR|PTE_xW|PTE_xX;
1025 		kxtlb->e_tlb.tlb_asid = KERNEL_PID;
1026 
1027 		kxtlb->e_hwtlb = tlb_to_hwtlb(kxtlb->e_tlb);
1028 		kxtlb->e_hwtlb.hwtlb_mas0 |= __SHIFTIN(kslot, MAS0_ESEL);
1029 		kxtlb->e_hwtlb.hwtlb_mas1 |= MAS1_TS;
1030 		hwtlb_write(kxtlb->e_hwtlb, true);
1031 
1032 		/*
1033 		 * Now that we have a TLB mapping in AS1 for the kernel and its
1034 		 * stack, we switch to AS1 to cleanup the TLB mappings for TLB0.
1035 		 */
1036 		const register_t saved_msr = mfmsr();
1037 		mtmsr(saved_msr | PSL_DS | PSL_IS);
1038 		__asm volatile("isync");
1039 
1040 		/*
1041 		 *** Invalidate all the TLB0 entries.
1042 		 */
1043 		e500_tlb_invalidate_all();
1044 
1045 		/*
1046 		 *** Now let's see if we have any entries in TLB1 that would
1047 		 *** overlap the ones we are about to install.  If so, nuke 'em.
1048 		 */
1049 		for (u_int i = 0; i < tlb1->tlb1_numentries; i++) {
1050 			struct e500_xtlb * const xtlb = &tlb1->tlb1_entries[i];
1051 			struct e500_hwtlb * const hwtlb = &xtlb->e_hwtlb;
1052 			if ((hwtlb->hwtlb_mas1 & (MAS1_V|MAS1_TS)) == MAS1_V
1053 			    && (hwtlb->hwtlb_mas2 & MAS2_EPN) < memsize) {
1054 				e500_free_tlb1_entry(xtlb, i, false);
1055 			}
1056 		}
1057 
1058 		/*
1059 		 *** Now we can add the TLB entries that will map physical
1060 		 *** memory.  If bit 0 [MSB] in slotmask is set, then tlb
1061 		 *** entry 0 contains a mapping for physical memory...
1062 		 */
1063 		struct e500_xtlb *entries = tlb1->tlb1_entries;
1064 		while (slotmask != 0) {
1065 			const u_int slot = __builtin_clz(slotmask);
1066 			hwtlb_write(entries[slot].e_hwtlb, false);
1067 			entries += slot + 1;
1068 			slotmask <<= slot + 1;
1069 		}
1070 
1071 		/*
1072 		 *** Synchronize the TLB and the instruction stream.
1073 		 */
1074 		__asm volatile("tlbsync");
1075 		__asm volatile("isync");
1076 
1077 		/*
1078 		 *** Switch back to AS 0.
1079 		 */
1080 		mtmsr(saved_msr);
1081 		__asm volatile("isync");
1082 
1083 		/*
1084 		 * Free the temporary TLB1 entry.
1085 		 */
1086 		e500_free_tlb1_entry(kxtlb, kslot, true);
1087 	}
1088 
1089 	/*
1090 	 * Finally set the MAS4 defaults.
1091 	 */
1092 	mtspr(SPR_MAS4, MAS4_TSIZED_4KB | MAS4_MD);
1093 
1094 	/*
1095 	 * Invalidate all the TLB0 entries.
1096 	 */
1097 	e500_tlb_invalidate_all();
1098 }
1099 
1100 void
1101 e500_tlb_minimize(vaddr_t endkernel)
1102 {
1103 #ifdef PMAP_MINIMALTLB
1104 	struct e500_tlb1 * const tlb1 = &e500_tlb1;
1105 	extern uint32_t _fdata[];
1106 
1107 	u_int slot;
1108 
1109 	paddr_t boot_page = cpu_read_4(GUR_BPTR);
1110 	if (boot_page & BPTR_EN) {
1111 		/*
1112 		 * shift it to an address
1113 		 */
1114 		boot_page = (boot_page & BPTR_BOOT_PAGE) << PAGE_SHIFT;
1115 		pmap_kvptefill(boot_page, boot_page + NBPG,
1116 		    PTE_M | PTE_xR | PTE_xW | PTE_xX);
1117 	}
1118 
1119 
1120 	KASSERT(endkernel - (uintptr_t)_fdata < 0x400000);
1121 	KASSERT((uintptr_t)_fdata == 0x400000);
1122 
1123 	struct e500_xtlb *xtlb = e500_tlb_lookup_xtlb(endkernel, &slot);
1124 
1125 	KASSERT(xtlb == e500_tlb_lookup_xtlb2(0, endkernel));
1126 	const u_int tmp_slot = e500_alloc_tlb1_entry();
1127 	KASSERT(tmp_slot != (u_int) -1);
1128 
1129 	struct e500_xtlb * const tmp_xtlb = &tlb1->tlb1_entries[tmp_slot];
1130 	tmp_xtlb->e_tlb = xtlb->e_tlb;
1131 	tmp_xtlb->e_hwtlb = tlb_to_hwtlb(tmp_xtlb->e_tlb);
1132 	tmp_xtlb->e_hwtlb.hwtlb_mas1 |= MAS1_TS;
1133 	KASSERT((tmp_xtlb->e_hwtlb.hwtlb_mas0 & MAS0_TLBSEL) == MAS0_TLBSEL_TLB1);
1134 	tmp_xtlb->e_hwtlb.hwtlb_mas0 |= __SHIFTIN(tmp_slot, MAS0_ESEL);
1135 	hwtlb_write(tmp_xtlb->e_hwtlb, true);
1136 
1137 	const u_int text_slot = e500_alloc_tlb1_entry();
1138 	KASSERT(text_slot != (u_int)-1);
1139 	struct e500_xtlb * const text_xtlb = &tlb1->tlb1_entries[text_slot];
1140 	text_xtlb->e_tlb.tlb_va = 0;
1141 	text_xtlb->e_tlb.tlb_size = 0x400000;
1142 	text_xtlb->e_tlb.tlb_pte = PTE_M | PTE_xR | PTE_xX | text_xtlb->e_tlb.tlb_va;
1143 	text_xtlb->e_tlb.tlb_asid = 0;
1144 	text_xtlb->e_hwtlb = tlb_to_hwtlb(text_xtlb->e_tlb);
1145 	KASSERT((text_xtlb->e_hwtlb.hwtlb_mas0 & MAS0_TLBSEL) == MAS0_TLBSEL_TLB1);
1146 	text_xtlb->e_hwtlb.hwtlb_mas0 |= __SHIFTIN(text_slot, MAS0_ESEL);
1147 
1148 	const u_int data_slot = e500_alloc_tlb1_entry();
1149 	KASSERT(data_slot != (u_int)-1);
1150 	struct e500_xtlb * const data_xtlb = &tlb1->tlb1_entries[data_slot];
1151 	data_xtlb->e_tlb.tlb_va = 0x400000;
1152 	data_xtlb->e_tlb.tlb_size = 0x400000;
1153 	data_xtlb->e_tlb.tlb_pte = PTE_M | PTE_xR | PTE_xW | data_xtlb->e_tlb.tlb_va;
1154 	data_xtlb->e_tlb.tlb_asid = 0;
1155 	data_xtlb->e_hwtlb = tlb_to_hwtlb(data_xtlb->e_tlb);
1156 	KASSERT((data_xtlb->e_hwtlb.hwtlb_mas0 & MAS0_TLBSEL) == MAS0_TLBSEL_TLB1);
1157 	data_xtlb->e_hwtlb.hwtlb_mas0 |= __SHIFTIN(data_slot, MAS0_ESEL);
1158 
1159 	const register_t msr = mfmsr();
1160 	const register_t ts_msr = (msr | PSL_DS | PSL_IS) & ~PSL_EE;
1161 
1162 	__asm __volatile(
1163 		"mtmsr	%[ts_msr]"	"\n\t"
1164 		"sync"			"\n\t"
1165 		"isync"
1166 	    ::	[ts_msr] "r" (ts_msr));
1167 
1168 #if 0
1169 	hwtlb_write(text_xtlb->e_hwtlb, false);
1170 	hwtlb_write(data_xtlb->e_hwtlb, false);
1171 	e500_free_tlb1_entry(xtlb, slot, true);
1172 #endif
1173 
1174 	__asm __volatile(
1175 		"mtmsr	%[msr]"		"\n\t"
1176 		"sync"			"\n\t"
1177 		"isync"
1178 	    ::	[msr] "r" (msr));
1179 
1180 	e500_free_tlb1_entry(tmp_xtlb, tmp_slot, true);
1181 #endif	/* PMAP_MINIMALTLB */
1182 }
1183