xref: /linux/arch/powerpc/kvm/e500_mmu_host.c (revision a634dda26186cf9a51567020fcce52bcba5e1e59)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2008-2013 Freescale Semiconductor, Inc. All rights reserved.
4  *
5  * Author: Yu Liu, yu.liu@freescale.com
6  *         Scott Wood, scottwood@freescale.com
7  *         Ashish Kalra, ashish.kalra@freescale.com
8  *         Varun Sethi, varun.sethi@freescale.com
9  *         Alexander Graf, agraf@suse.de
10  *
11  * Description:
12  * This file is based on arch/powerpc/kvm/44x_tlb.c,
13  * by Hollis Blanchard <hollisb@us.ibm.com>.
14  */
15 
16 #include <linux/kernel.h>
17 #include <linux/types.h>
18 #include <linux/slab.h>
19 #include <linux/string.h>
20 #include <linux/kvm.h>
21 #include <linux/kvm_host.h>
22 #include <linux/highmem.h>
23 #include <linux/log2.h>
24 #include <linux/uaccess.h>
25 #include <linux/sched/mm.h>
26 #include <linux/rwsem.h>
27 #include <linux/vmalloc.h>
28 #include <linux/hugetlb.h>
29 #include <asm/kvm_ppc.h>
30 #include <asm/pte-walk.h>
31 
32 #include "e500.h"
33 #include "timing.h"
34 #include "e500_mmu_host.h"
35 
36 #include "trace_booke.h"
37 
38 #define to_htlb1_esel(esel) (host_tlb_params[1].entries - (esel) - 1)
39 
40 static struct kvmppc_e500_tlb_params host_tlb_params[E500_TLB_NUM];
41 
42 static inline unsigned int tlb1_max_shadow_size(void)
43 {
44 	/* reserve one entry for magic page */
45 	return host_tlb_params[1].entries - tlbcam_index - 1;
46 }
47 
48 static inline u32 e500_shadow_mas3_attrib(u32 mas3, bool writable, int usermode)
49 {
50 	/* Mask off reserved bits. */
51 	mas3 &= MAS3_ATTRIB_MASK;
52 
53 	if (!writable)
54 		mas3 &= ~(MAS3_UW|MAS3_SW);
55 
56 #ifndef CONFIG_KVM_BOOKE_HV
57 	if (!usermode) {
58 		/* Guest is in supervisor mode,
59 		 * so we need to translate guest
60 		 * supervisor permissions into user permissions. */
61 		mas3 &= ~E500_TLB_USER_PERM_MASK;
62 		mas3 |= (mas3 & E500_TLB_SUPER_PERM_MASK) << 1;
63 	}
64 	mas3 |= E500_TLB_SUPER_PERM_MASK;
65 #endif
66 	return mas3;
67 }
68 
69 /*
70  * writing shadow tlb entry to host TLB
71  */
72 static inline void __write_host_tlbe(struct kvm_book3e_206_tlb_entry *stlbe,
73 				     uint32_t mas0,
74 				     uint32_t lpid)
75 {
76 	unsigned long flags;
77 
78 	local_irq_save(flags);
79 	mtspr(SPRN_MAS0, mas0);
80 	mtspr(SPRN_MAS1, stlbe->mas1);
81 	mtspr(SPRN_MAS2, (unsigned long)stlbe->mas2);
82 	mtspr(SPRN_MAS3, (u32)stlbe->mas7_3);
83 	mtspr(SPRN_MAS7, (u32)(stlbe->mas7_3 >> 32));
84 #ifdef CONFIG_KVM_BOOKE_HV
85 	mtspr(SPRN_MAS8, MAS8_TGS | get_thread_specific_lpid(lpid));
86 #endif
87 	asm volatile("isync; tlbwe" : : : "memory");
88 
89 #ifdef CONFIG_KVM_BOOKE_HV
90 	/* Must clear mas8 for other host tlbwe's */
91 	mtspr(SPRN_MAS8, 0);
92 	isync();
93 #endif
94 	local_irq_restore(flags);
95 
96 	trace_kvm_booke206_stlb_write(mas0, stlbe->mas8, stlbe->mas1,
97 	                              stlbe->mas2, stlbe->mas7_3);
98 }
99 
100 /*
101  * Acquire a mas0 with victim hint, as if we just took a TLB miss.
102  *
103  * We don't care about the address we're searching for, other than that it's
104  * in the right set and is not present in the TLB.  Using a zero PID and a
105  * userspace address means we don't have to set and then restore MAS5, or
106  * calculate a proper MAS6 value.
107  */
108 static u32 get_host_mas0(unsigned long eaddr)
109 {
110 	unsigned long flags;
111 	u32 mas0;
112 	u32 mas4;
113 
114 	local_irq_save(flags);
115 	mtspr(SPRN_MAS6, 0);
116 	mas4 = mfspr(SPRN_MAS4);
117 	mtspr(SPRN_MAS4, mas4 & ~MAS4_TLBSEL_MASK);
118 	asm volatile("tlbsx 0, %0" : : "b" (eaddr & ~CONFIG_PAGE_OFFSET));
119 	mas0 = mfspr(SPRN_MAS0);
120 	mtspr(SPRN_MAS4, mas4);
121 	local_irq_restore(flags);
122 
123 	return mas0;
124 }
125 
126 /* sesel is for tlb1 only */
127 static inline void write_host_tlbe(struct kvmppc_vcpu_e500 *vcpu_e500,
128 		int tlbsel, int sesel, struct kvm_book3e_206_tlb_entry *stlbe)
129 {
130 	u32 mas0;
131 
132 	if (tlbsel == 0) {
133 		mas0 = get_host_mas0(stlbe->mas2);
134 		__write_host_tlbe(stlbe, mas0, vcpu_e500->vcpu.kvm->arch.lpid);
135 	} else {
136 		__write_host_tlbe(stlbe,
137 				  MAS0_TLBSEL(1) |
138 				  MAS0_ESEL(to_htlb1_esel(sesel)),
139 				  vcpu_e500->vcpu.kvm->arch.lpid);
140 	}
141 }
142 
143 /* sesel is for tlb1 only */
144 static void write_stlbe(struct kvmppc_vcpu_e500 *vcpu_e500,
145 			struct kvm_book3e_206_tlb_entry *gtlbe,
146 			struct kvm_book3e_206_tlb_entry *stlbe,
147 			int stlbsel, int sesel)
148 {
149 	int stid;
150 
151 	preempt_disable();
152 	stid = kvmppc_e500_get_tlb_stid(&vcpu_e500->vcpu, gtlbe);
153 
154 	stlbe->mas1 |= MAS1_TID(stid);
155 	write_host_tlbe(vcpu_e500, stlbsel, sesel, stlbe);
156 	preempt_enable();
157 }
158 
159 #ifdef CONFIG_KVM_E500V2
160 /* XXX should be a hook in the gva2hpa translation */
161 void kvmppc_map_magic(struct kvm_vcpu *vcpu)
162 {
163 	struct kvmppc_vcpu_e500 *vcpu_e500 = to_e500(vcpu);
164 	struct kvm_book3e_206_tlb_entry magic;
165 	ulong shared_page = ((ulong)vcpu->arch.shared) & PAGE_MASK;
166 	unsigned int stid;
167 	kvm_pfn_t pfn;
168 
169 	pfn = (kvm_pfn_t)virt_to_phys((void *)shared_page) >> PAGE_SHIFT;
170 	get_page(pfn_to_page(pfn));
171 
172 	preempt_disable();
173 	stid = kvmppc_e500_get_sid(vcpu_e500, 0, 0, 0, 0);
174 
175 	magic.mas1 = MAS1_VALID | MAS1_TS | MAS1_TID(stid) |
176 		     MAS1_TSIZE(BOOK3E_PAGESZ_4K);
177 	magic.mas2 = vcpu->arch.magic_page_ea | MAS2_M;
178 	magic.mas7_3 = ((u64)pfn << PAGE_SHIFT) |
179 		       MAS3_SW | MAS3_SR | MAS3_UW | MAS3_UR;
180 	magic.mas8 = 0;
181 
182 	__write_host_tlbe(&magic, MAS0_TLBSEL(1) | MAS0_ESEL(tlbcam_index), 0);
183 	preempt_enable();
184 }
185 #endif
186 
187 void inval_gtlbe_on_host(struct kvmppc_vcpu_e500 *vcpu_e500, int tlbsel,
188 			 int esel)
189 {
190 	struct kvm_book3e_206_tlb_entry *gtlbe =
191 		get_entry(vcpu_e500, tlbsel, esel);
192 	struct tlbe_ref *ref = &vcpu_e500->gtlb_priv[tlbsel][esel].ref;
193 
194 	/* Don't bother with unmapped entries */
195 	if (!(ref->flags & E500_TLB_VALID)) {
196 		WARN(ref->flags & (E500_TLB_BITMAP | E500_TLB_TLB0),
197 		     "%s: flags %x\n", __func__, ref->flags);
198 		WARN_ON(tlbsel == 1 && vcpu_e500->g2h_tlb1_map[esel]);
199 	}
200 
201 	if (tlbsel == 1 && ref->flags & E500_TLB_BITMAP) {
202 		u64 tmp = vcpu_e500->g2h_tlb1_map[esel];
203 		int hw_tlb_indx;
204 		unsigned long flags;
205 
206 		local_irq_save(flags);
207 		while (tmp) {
208 			hw_tlb_indx = __ilog2_u64(tmp & -tmp);
209 			mtspr(SPRN_MAS0,
210 			      MAS0_TLBSEL(1) |
211 			      MAS0_ESEL(to_htlb1_esel(hw_tlb_indx)));
212 			mtspr(SPRN_MAS1, 0);
213 			asm volatile("tlbwe");
214 			vcpu_e500->h2g_tlb1_rmap[hw_tlb_indx] = 0;
215 			tmp &= tmp - 1;
216 		}
217 		mb();
218 		vcpu_e500->g2h_tlb1_map[esel] = 0;
219 		ref->flags &= ~(E500_TLB_BITMAP | E500_TLB_VALID);
220 		local_irq_restore(flags);
221 	}
222 
223 	if (tlbsel == 1 && ref->flags & E500_TLB_TLB0) {
224 		/*
225 		 * TLB1 entry is backed by 4k pages. This should happen
226 		 * rarely and is not worth optimizing. Invalidate everything.
227 		 */
228 		kvmppc_e500_tlbil_all(vcpu_e500);
229 		ref->flags &= ~(E500_TLB_TLB0 | E500_TLB_VALID);
230 	}
231 
232 	/*
233 	 * If TLB entry is still valid then it's a TLB0 entry, and thus
234 	 * backed by at most one host tlbe per shadow pid
235 	 */
236 	if (ref->flags & E500_TLB_VALID)
237 		kvmppc_e500_tlbil_one(vcpu_e500, gtlbe);
238 
239 	/* Mark the TLB as not backed by the host anymore */
240 	ref->flags = 0;
241 }
242 
243 static inline int tlbe_is_writable(struct kvm_book3e_206_tlb_entry *tlbe)
244 {
245 	return tlbe->mas7_3 & (MAS3_SW|MAS3_UW);
246 }
247 
248 static inline void kvmppc_e500_ref_setup(struct tlbe_ref *ref,
249 					 struct kvm_book3e_206_tlb_entry *gtlbe,
250 					 kvm_pfn_t pfn, unsigned int wimg,
251 					 bool writable)
252 {
253 	ref->pfn = pfn;
254 	ref->flags = E500_TLB_VALID;
255 	if (writable)
256 		ref->flags |= E500_TLB_WRITABLE;
257 
258 	/* Use guest supplied MAS2_G and MAS2_E */
259 	ref->flags |= (gtlbe->mas2 & MAS2_ATTRIB_MASK) | wimg;
260 }
261 
262 static inline void kvmppc_e500_ref_release(struct tlbe_ref *ref)
263 {
264 	if (ref->flags & E500_TLB_VALID) {
265 		/* FIXME: don't log bogus pfn for TLB1 */
266 		trace_kvm_booke206_ref_release(ref->pfn, ref->flags);
267 		ref->flags = 0;
268 	}
269 }
270 
271 static void clear_tlb1_bitmap(struct kvmppc_vcpu_e500 *vcpu_e500)
272 {
273 	if (vcpu_e500->g2h_tlb1_map)
274 		memset(vcpu_e500->g2h_tlb1_map, 0,
275 		       sizeof(u64) * vcpu_e500->gtlb_params[1].entries);
276 	if (vcpu_e500->h2g_tlb1_rmap)
277 		memset(vcpu_e500->h2g_tlb1_rmap, 0,
278 		       sizeof(unsigned int) * host_tlb_params[1].entries);
279 }
280 
281 static void clear_tlb_privs(struct kvmppc_vcpu_e500 *vcpu_e500)
282 {
283 	int tlbsel;
284 	int i;
285 
286 	for (tlbsel = 0; tlbsel <= 1; tlbsel++) {
287 		for (i = 0; i < vcpu_e500->gtlb_params[tlbsel].entries; i++) {
288 			struct tlbe_ref *ref =
289 				&vcpu_e500->gtlb_priv[tlbsel][i].ref;
290 			kvmppc_e500_ref_release(ref);
291 		}
292 	}
293 }
294 
295 void kvmppc_core_flush_tlb(struct kvm_vcpu *vcpu)
296 {
297 	struct kvmppc_vcpu_e500 *vcpu_e500 = to_e500(vcpu);
298 	kvmppc_e500_tlbil_all(vcpu_e500);
299 	clear_tlb_privs(vcpu_e500);
300 	clear_tlb1_bitmap(vcpu_e500);
301 }
302 
303 /* TID must be supplied by the caller */
304 static void kvmppc_e500_setup_stlbe(
305 	struct kvm_vcpu *vcpu,
306 	struct kvm_book3e_206_tlb_entry *gtlbe,
307 	int tsize, struct tlbe_ref *ref, u64 gvaddr,
308 	struct kvm_book3e_206_tlb_entry *stlbe)
309 {
310 	kvm_pfn_t pfn = ref->pfn;
311 	u32 pr = vcpu->arch.shared->msr & MSR_PR;
312 	bool writable = !!(ref->flags & E500_TLB_WRITABLE);
313 
314 	BUG_ON(!(ref->flags & E500_TLB_VALID));
315 
316 	/* Force IPROT=0 for all guest mappings. */
317 	stlbe->mas1 = MAS1_TSIZE(tsize) | get_tlb_sts(gtlbe) | MAS1_VALID;
318 	stlbe->mas2 = (gvaddr & MAS2_EPN) | (ref->flags & E500_TLB_MAS2_ATTR);
319 	stlbe->mas7_3 = ((u64)pfn << PAGE_SHIFT) |
320 			e500_shadow_mas3_attrib(gtlbe->mas7_3, writable, pr);
321 }
322 
323 static inline int kvmppc_e500_shadow_map(struct kvmppc_vcpu_e500 *vcpu_e500,
324 	u64 gvaddr, gfn_t gfn, struct kvm_book3e_206_tlb_entry *gtlbe,
325 	int tlbsel, struct kvm_book3e_206_tlb_entry *stlbe,
326 	struct tlbe_ref *ref)
327 {
328 	struct kvm_memory_slot *slot;
329 	unsigned int psize;
330 	unsigned long pfn;
331 	struct page *page = NULL;
332 	unsigned long hva;
333 	int tsize = BOOK3E_PAGESZ_4K;
334 	int ret = 0;
335 	unsigned long mmu_seq;
336 	struct kvm *kvm = vcpu_e500->vcpu.kvm;
337 	pte_t *ptep;
338 	unsigned int wimg = 0;
339 	pgd_t *pgdir;
340 	unsigned long flags;
341 	bool writable = false;
342 
343 	/* used to check for invalidations in progress */
344 	mmu_seq = kvm->mmu_invalidate_seq;
345 	smp_rmb();
346 
347 	/*
348 	 * Translate guest physical to true physical, acquiring
349 	 * a page reference if it is normal, non-reserved memory.
350 	 *
351 	 * gfn_to_memslot() must succeed because otherwise we wouldn't
352 	 * have gotten this far.  Eventually we should just pass the slot
353 	 * pointer through from the first lookup.
354 	 */
355 	slot = gfn_to_memslot(vcpu_e500->vcpu.kvm, gfn);
356 	hva = gfn_to_hva_memslot(slot, gfn);
357 
358 	pfn = __kvm_faultin_pfn(slot, gfn, FOLL_WRITE, &writable, &page);
359 	if (is_error_noslot_pfn(pfn)) {
360 		if (printk_ratelimit())
361 			pr_err("%s: real page not found for gfn %lx\n",
362 			       __func__, (long)gfn);
363 		return -EINVAL;
364 	}
365 
366 	spin_lock(&kvm->mmu_lock);
367 	if (mmu_invalidate_retry(kvm, mmu_seq)) {
368 		ret = -EAGAIN;
369 		goto out;
370 	}
371 
372 
373 	pgdir = vcpu_e500->vcpu.arch.pgdir;
374 	/*
375 	 * We are just looking at the wimg bits, so we don't
376 	 * care much about the trans splitting bit.
377 	 * We are holding kvm->mmu_lock so a notifier invalidate
378 	 * can't run hence pfn won't change.
379 	 */
380 	local_irq_save(flags);
381 	ptep = find_linux_pte(pgdir, hva, NULL, &psize);
382 	if (ptep) {
383 		pte_t pte = READ_ONCE(*ptep);
384 
385 		if (pte_present(pte)) {
386 			wimg = (pte_val(pte) >> PTE_WIMGE_SHIFT) &
387 				MAS2_WIMGE_MASK;
388 		} else {
389 			local_irq_restore(flags);
390 			pr_err_ratelimited("%s: pte not present: gfn %lx,pfn %lx\n",
391 					   __func__, (long)gfn, pfn);
392 			ret = -EINVAL;
393 			goto out;
394 		}
395 	}
396 	local_irq_restore(flags);
397 
398 	if (psize && tlbsel == 1) {
399 		unsigned long psize_pages, tsize_pages;
400 		unsigned long start, end;
401 		unsigned long slot_start, slot_end;
402 
403 		psize_pages = 1UL << (psize - PAGE_SHIFT);
404 		start = pfn & ~(psize_pages - 1);
405 		end = start + psize_pages;
406 
407 		slot_start = pfn - (gfn - slot->base_gfn);
408 		slot_end = slot_start + slot->npages;
409 
410 		if (start < slot_start)
411 			start = slot_start;
412 		if (end > slot_end)
413 			end = slot_end;
414 
415 		tsize = (gtlbe->mas1 & MAS1_TSIZE_MASK) >>
416 			MAS1_TSIZE_SHIFT;
417 
418 		/*
419 		 * Any page size that doesn't satisfy the host mapping
420 		 * will fail the start and end tests.
421 		 */
422 		tsize = min(psize - PAGE_SHIFT + BOOK3E_PAGESZ_4K, tsize);
423 
424 		/*
425 		 * e500 doesn't implement the lowest tsize bit,
426 		 * or 1K pages.
427 		 */
428 		tsize = max(BOOK3E_PAGESZ_4K, tsize & ~1);
429 
430 		/*
431 		 * Now find the largest tsize (up to what the guest
432 		 * requested) that will cover gfn, stay within the
433 		 * range, and for which gfn and pfn are mutually
434 		 * aligned.
435 		 */
436 
437 		for (; tsize > BOOK3E_PAGESZ_4K; tsize -= 2) {
438 			unsigned long gfn_start, gfn_end;
439 			tsize_pages = 1UL << (tsize - 2);
440 
441 			gfn_start = gfn & ~(tsize_pages - 1);
442 			gfn_end = gfn_start + tsize_pages;
443 
444 			if (gfn_start + pfn - gfn < start)
445 				continue;
446 			if (gfn_end + pfn - gfn > end)
447 				continue;
448 			if ((gfn & (tsize_pages - 1)) !=
449 			    (pfn & (tsize_pages - 1)))
450 				continue;
451 
452 			gvaddr &= ~((tsize_pages << PAGE_SHIFT) - 1);
453 			pfn &= ~(tsize_pages - 1);
454 			break;
455 		}
456 	}
457 
458 	kvmppc_e500_ref_setup(ref, gtlbe, pfn, wimg, writable);
459 	kvmppc_e500_setup_stlbe(&vcpu_e500->vcpu, gtlbe, tsize,
460 				ref, gvaddr, stlbe);
461 	writable = tlbe_is_writable(stlbe);
462 
463 	/* Clear i-cache for new pages */
464 	kvmppc_mmu_flush_icache(pfn);
465 
466 out:
467 	kvm_release_faultin_page(kvm, page, !!ret, writable);
468 	spin_unlock(&kvm->mmu_lock);
469 	return ret;
470 }
471 
472 /* XXX only map the one-one case, for now use TLB0 */
473 static int kvmppc_e500_tlb0_map(struct kvmppc_vcpu_e500 *vcpu_e500, int esel,
474 				struct kvm_book3e_206_tlb_entry *stlbe)
475 {
476 	struct kvm_book3e_206_tlb_entry *gtlbe;
477 	struct tlbe_ref *ref;
478 	int stlbsel = 0;
479 	int sesel = 0;
480 	int r;
481 
482 	gtlbe = get_entry(vcpu_e500, 0, esel);
483 	ref = &vcpu_e500->gtlb_priv[0][esel].ref;
484 
485 	r = kvmppc_e500_shadow_map(vcpu_e500, get_tlb_eaddr(gtlbe),
486 			get_tlb_raddr(gtlbe) >> PAGE_SHIFT,
487 			gtlbe, 0, stlbe, ref);
488 	if (r)
489 		return r;
490 
491 	write_stlbe(vcpu_e500, gtlbe, stlbe, stlbsel, sesel);
492 
493 	return 0;
494 }
495 
496 static int kvmppc_e500_tlb1_map_tlb1(struct kvmppc_vcpu_e500 *vcpu_e500,
497 				     struct tlbe_ref *ref,
498 				     int esel)
499 {
500 	unsigned int sesel = vcpu_e500->host_tlb1_nv++;
501 
502 	if (unlikely(vcpu_e500->host_tlb1_nv >= tlb1_max_shadow_size()))
503 		vcpu_e500->host_tlb1_nv = 0;
504 
505 	if (vcpu_e500->h2g_tlb1_rmap[sesel]) {
506 		unsigned int idx = vcpu_e500->h2g_tlb1_rmap[sesel] - 1;
507 		vcpu_e500->g2h_tlb1_map[idx] &= ~(1ULL << sesel);
508 	}
509 
510 	vcpu_e500->gtlb_priv[1][esel].ref.flags |= E500_TLB_BITMAP;
511 	vcpu_e500->g2h_tlb1_map[esel] |= (u64)1 << sesel;
512 	vcpu_e500->h2g_tlb1_rmap[sesel] = esel + 1;
513 	WARN_ON(!(ref->flags & E500_TLB_VALID));
514 
515 	return sesel;
516 }
517 
518 /* Caller must ensure that the specified guest TLB entry is safe to insert into
519  * the shadow TLB. */
520 /* For both one-one and one-to-many */
521 static int kvmppc_e500_tlb1_map(struct kvmppc_vcpu_e500 *vcpu_e500,
522 		u64 gvaddr, gfn_t gfn, struct kvm_book3e_206_tlb_entry *gtlbe,
523 		struct kvm_book3e_206_tlb_entry *stlbe, int esel)
524 {
525 	struct tlbe_ref *ref = &vcpu_e500->gtlb_priv[1][esel].ref;
526 	int sesel;
527 	int r;
528 
529 	r = kvmppc_e500_shadow_map(vcpu_e500, gvaddr, gfn, gtlbe, 1, stlbe,
530 				   ref);
531 	if (r)
532 		return r;
533 
534 	/* Use TLB0 when we can only map a page with 4k */
535 	if (get_tlb_tsize(stlbe) == BOOK3E_PAGESZ_4K) {
536 		vcpu_e500->gtlb_priv[1][esel].ref.flags |= E500_TLB_TLB0;
537 		write_stlbe(vcpu_e500, gtlbe, stlbe, 0, 0);
538 		return 0;
539 	}
540 
541 	/* Otherwise map into TLB1 */
542 	sesel = kvmppc_e500_tlb1_map_tlb1(vcpu_e500, ref, esel);
543 	write_stlbe(vcpu_e500, gtlbe, stlbe, 1, sesel);
544 
545 	return 0;
546 }
547 
548 void kvmppc_mmu_map(struct kvm_vcpu *vcpu, u64 eaddr, gpa_t gpaddr,
549 		    unsigned int index)
550 {
551 	struct kvmppc_vcpu_e500 *vcpu_e500 = to_e500(vcpu);
552 	struct tlbe_priv *priv;
553 	struct kvm_book3e_206_tlb_entry *gtlbe, stlbe;
554 	int tlbsel = tlbsel_of(index);
555 	int esel = esel_of(index);
556 
557 	gtlbe = get_entry(vcpu_e500, tlbsel, esel);
558 
559 	switch (tlbsel) {
560 	case 0:
561 		priv = &vcpu_e500->gtlb_priv[tlbsel][esel];
562 
563 		/* Triggers after clear_tlb_privs or on initial mapping */
564 		if (!(priv->ref.flags & E500_TLB_VALID)) {
565 			kvmppc_e500_tlb0_map(vcpu_e500, esel, &stlbe);
566 		} else {
567 			kvmppc_e500_setup_stlbe(vcpu, gtlbe, BOOK3E_PAGESZ_4K,
568 						&priv->ref, eaddr, &stlbe);
569 			write_stlbe(vcpu_e500, gtlbe, &stlbe, 0, 0);
570 		}
571 		break;
572 
573 	case 1: {
574 		gfn_t gfn = gpaddr >> PAGE_SHIFT;
575 		kvmppc_e500_tlb1_map(vcpu_e500, eaddr, gfn, gtlbe, &stlbe,
576 				     esel);
577 		break;
578 	}
579 
580 	default:
581 		BUG();
582 		break;
583 	}
584 }
585 
586 #ifdef CONFIG_KVM_BOOKE_HV
587 int kvmppc_load_last_inst(struct kvm_vcpu *vcpu,
588 		enum instruction_fetch_type type, unsigned long *instr)
589 {
590 	gva_t geaddr;
591 	hpa_t addr;
592 	hfn_t pfn;
593 	hva_t eaddr;
594 	u32 mas1, mas2, mas3;
595 	u64 mas7_mas3;
596 	struct page *page;
597 	unsigned int addr_space, psize_shift;
598 	bool pr;
599 	unsigned long flags;
600 
601 	/* Search TLB for guest pc to get the real address */
602 	geaddr = kvmppc_get_pc(vcpu);
603 
604 	addr_space = (vcpu->arch.shared->msr & MSR_IS) >> MSR_IR_LG;
605 
606 	local_irq_save(flags);
607 	mtspr(SPRN_MAS6, (vcpu->arch.pid << MAS6_SPID_SHIFT) | addr_space);
608 	mtspr(SPRN_MAS5, MAS5_SGS | get_lpid(vcpu));
609 	asm volatile("tlbsx 0, %[geaddr]\n" : :
610 		     [geaddr] "r" (geaddr));
611 	mtspr(SPRN_MAS5, 0);
612 	mtspr(SPRN_MAS8, 0);
613 	mas1 = mfspr(SPRN_MAS1);
614 	mas2 = mfspr(SPRN_MAS2);
615 	mas3 = mfspr(SPRN_MAS3);
616 #ifdef CONFIG_64BIT
617 	mas7_mas3 = mfspr(SPRN_MAS7_MAS3);
618 #else
619 	mas7_mas3 = ((u64)mfspr(SPRN_MAS7) << 32) | mas3;
620 #endif
621 	local_irq_restore(flags);
622 
623 	/*
624 	 * If the TLB entry for guest pc was evicted, return to the guest.
625 	 * There are high chances to find a valid TLB entry next time.
626 	 */
627 	if (!(mas1 & MAS1_VALID))
628 		return EMULATE_AGAIN;
629 
630 	/*
631 	 * Another thread may rewrite the TLB entry in parallel, don't
632 	 * execute from the address if the execute permission is not set
633 	 */
634 	pr = vcpu->arch.shared->msr & MSR_PR;
635 	if (unlikely((pr && !(mas3 & MAS3_UX)) ||
636 		     (!pr && !(mas3 & MAS3_SX)))) {
637 		pr_err_ratelimited(
638 			"%s: Instruction emulation from guest address %08lx without execute permission\n",
639 			__func__, geaddr);
640 		return EMULATE_AGAIN;
641 	}
642 
643 	/*
644 	 * The real address will be mapped by a cacheable, memory coherent,
645 	 * write-back page. Check for mismatches when LRAT is used.
646 	 */
647 	if (has_feature(vcpu, VCPU_FTR_MMU_V2) &&
648 	    unlikely((mas2 & MAS2_I) || (mas2 & MAS2_W) || !(mas2 & MAS2_M))) {
649 		pr_err_ratelimited(
650 			"%s: Instruction emulation from guest address %08lx mismatches storage attributes\n",
651 			__func__, geaddr);
652 		return EMULATE_AGAIN;
653 	}
654 
655 	/* Get pfn */
656 	psize_shift = MAS1_GET_TSIZE(mas1) + 10;
657 	addr = (mas7_mas3 & (~0ULL << psize_shift)) |
658 	       (geaddr & ((1ULL << psize_shift) - 1ULL));
659 	pfn = addr >> PAGE_SHIFT;
660 
661 	/* Guard against emulation from devices area */
662 	if (unlikely(!page_is_ram(pfn))) {
663 		pr_err_ratelimited("%s: Instruction emulation from non-RAM host address %08llx is not supported\n",
664 			 __func__, addr);
665 		return EMULATE_AGAIN;
666 	}
667 
668 	/* Map a page and get guest's instruction */
669 	page = pfn_to_page(pfn);
670 	eaddr = (unsigned long)kmap_atomic(page);
671 	*instr = *(u32 *)(eaddr | (unsigned long)(addr & ~PAGE_MASK));
672 	kunmap_atomic((u32 *)eaddr);
673 
674 	return EMULATE_DONE;
675 }
676 #else
677 int kvmppc_load_last_inst(struct kvm_vcpu *vcpu,
678 		enum instruction_fetch_type type, unsigned long *instr)
679 {
680 	return EMULATE_AGAIN;
681 }
682 #endif
683 
684 /************* MMU Notifiers *************/
685 
686 static bool kvm_e500_mmu_unmap_gfn(struct kvm *kvm, struct kvm_gfn_range *range)
687 {
688 	/*
689 	 * Flush all shadow tlb entries everywhere. This is slow, but
690 	 * we are 100% sure that we catch the to be unmapped page
691 	 */
692 	return true;
693 }
694 
695 bool kvm_unmap_gfn_range(struct kvm *kvm, struct kvm_gfn_range *range)
696 {
697 	return kvm_e500_mmu_unmap_gfn(kvm, range);
698 }
699 
700 bool kvm_age_gfn(struct kvm *kvm, struct kvm_gfn_range *range)
701 {
702 	/* XXX could be more clever ;) */
703 	return false;
704 }
705 
706 bool kvm_test_age_gfn(struct kvm *kvm, struct kvm_gfn_range *range)
707 {
708 	/* XXX could be more clever ;) */
709 	return false;
710 }
711 
712 /*****************************************/
713 
714 int e500_mmu_host_init(struct kvmppc_vcpu_e500 *vcpu_e500)
715 {
716 	host_tlb_params[0].entries = mfspr(SPRN_TLB0CFG) & TLBnCFG_N_ENTRY;
717 	host_tlb_params[1].entries = mfspr(SPRN_TLB1CFG) & TLBnCFG_N_ENTRY;
718 
719 	/*
720 	 * This should never happen on real e500 hardware, but is
721 	 * architecturally possible -- e.g. in some weird nested
722 	 * virtualization case.
723 	 */
724 	if (host_tlb_params[0].entries == 0 ||
725 	    host_tlb_params[1].entries == 0) {
726 		pr_err("%s: need to know host tlb size\n", __func__);
727 		return -ENODEV;
728 	}
729 
730 	host_tlb_params[0].ways = (mfspr(SPRN_TLB0CFG) & TLBnCFG_ASSOC) >>
731 				  TLBnCFG_ASSOC_SHIFT;
732 	host_tlb_params[1].ways = host_tlb_params[1].entries;
733 
734 	if (!is_power_of_2(host_tlb_params[0].entries) ||
735 	    !is_power_of_2(host_tlb_params[0].ways) ||
736 	    host_tlb_params[0].entries < host_tlb_params[0].ways ||
737 	    host_tlb_params[0].ways == 0) {
738 		pr_err("%s: bad tlb0 host config: %u entries %u ways\n",
739 		       __func__, host_tlb_params[0].entries,
740 		       host_tlb_params[0].ways);
741 		return -ENODEV;
742 	}
743 
744 	host_tlb_params[0].sets =
745 		host_tlb_params[0].entries / host_tlb_params[0].ways;
746 	host_tlb_params[1].sets = 1;
747 	vcpu_e500->h2g_tlb1_rmap = kcalloc(host_tlb_params[1].entries,
748 					   sizeof(*vcpu_e500->h2g_tlb1_rmap),
749 					   GFP_KERNEL);
750 	if (!vcpu_e500->h2g_tlb1_rmap)
751 		return -EINVAL;
752 
753 	return 0;
754 }
755 
756 void e500_mmu_host_uninit(struct kvmppc_vcpu_e500 *vcpu_e500)
757 {
758 	kfree(vcpu_e500->h2g_tlb1_rmap);
759 }
760