xref: /linux/arch/powerpc/kvm/e500_mmu_host.c (revision 4949009eb8d40a441dcddcd96e101e77d31cf1b2)
1 /*
2  * Copyright (C) 2008-2013 Freescale Semiconductor, Inc. All rights reserved.
3  *
4  * Author: Yu Liu, yu.liu@freescale.com
5  *         Scott Wood, scottwood@freescale.com
6  *         Ashish Kalra, ashish.kalra@freescale.com
7  *         Varun Sethi, varun.sethi@freescale.com
8  *         Alexander Graf, agraf@suse.de
9  *
10  * Description:
11  * This file is based on arch/powerpc/kvm/44x_tlb.c,
12  * by Hollis Blanchard <hollisb@us.ibm.com>.
13  *
14  * This program is free software; you can redistribute it and/or modify
15  * it under the terms of the GNU General Public License, version 2, as
16  * published by the Free Software Foundation.
17  */
18 
19 #include <linux/kernel.h>
20 #include <linux/types.h>
21 #include <linux/slab.h>
22 #include <linux/string.h>
23 #include <linux/kvm.h>
24 #include <linux/kvm_host.h>
25 #include <linux/highmem.h>
26 #include <linux/log2.h>
27 #include <linux/uaccess.h>
28 #include <linux/sched.h>
29 #include <linux/rwsem.h>
30 #include <linux/vmalloc.h>
31 #include <linux/hugetlb.h>
32 #include <asm/kvm_ppc.h>
33 
34 #include "e500.h"
35 #include "timing.h"
36 #include "e500_mmu_host.h"
37 
38 #include "trace_booke.h"
39 
40 #define to_htlb1_esel(esel) (host_tlb_params[1].entries - (esel) - 1)
41 
42 static struct kvmppc_e500_tlb_params host_tlb_params[E500_TLB_NUM];
43 
44 static inline unsigned int tlb1_max_shadow_size(void)
45 {
46 	/* reserve one entry for magic page */
47 	return host_tlb_params[1].entries - tlbcam_index - 1;
48 }
49 
50 static inline u32 e500_shadow_mas3_attrib(u32 mas3, int usermode)
51 {
52 	/* Mask off reserved bits. */
53 	mas3 &= MAS3_ATTRIB_MASK;
54 
55 #ifndef CONFIG_KVM_BOOKE_HV
56 	if (!usermode) {
57 		/* Guest is in supervisor mode,
58 		 * so we need to translate guest
59 		 * supervisor permissions into user permissions. */
60 		mas3 &= ~E500_TLB_USER_PERM_MASK;
61 		mas3 |= (mas3 & E500_TLB_SUPER_PERM_MASK) << 1;
62 	}
63 	mas3 |= E500_TLB_SUPER_PERM_MASK;
64 #endif
65 	return mas3;
66 }
67 
68 /*
69  * writing shadow tlb entry to host TLB
70  */
71 static inline void __write_host_tlbe(struct kvm_book3e_206_tlb_entry *stlbe,
72 				     uint32_t mas0,
73 				     uint32_t lpid)
74 {
75 	unsigned long flags;
76 
77 	local_irq_save(flags);
78 	mtspr(SPRN_MAS0, mas0);
79 	mtspr(SPRN_MAS1, stlbe->mas1);
80 	mtspr(SPRN_MAS2, (unsigned long)stlbe->mas2);
81 	mtspr(SPRN_MAS3, (u32)stlbe->mas7_3);
82 	mtspr(SPRN_MAS7, (u32)(stlbe->mas7_3 >> 32));
83 #ifdef CONFIG_KVM_BOOKE_HV
84 	mtspr(SPRN_MAS8, MAS8_TGS | get_thread_specific_lpid(lpid));
85 #endif
86 	asm volatile("isync; tlbwe" : : : "memory");
87 
88 #ifdef CONFIG_KVM_BOOKE_HV
89 	/* Must clear mas8 for other host tlbwe's */
90 	mtspr(SPRN_MAS8, 0);
91 	isync();
92 #endif
93 	local_irq_restore(flags);
94 
95 	trace_kvm_booke206_stlb_write(mas0, stlbe->mas8, stlbe->mas1,
96 	                              stlbe->mas2, stlbe->mas7_3);
97 }
98 
99 /*
100  * Acquire a mas0 with victim hint, as if we just took a TLB miss.
101  *
102  * We don't care about the address we're searching for, other than that it's
103  * in the right set and is not present in the TLB.  Using a zero PID and a
104  * userspace address means we don't have to set and then restore MAS5, or
105  * calculate a proper MAS6 value.
106  */
107 static u32 get_host_mas0(unsigned long eaddr)
108 {
109 	unsigned long flags;
110 	u32 mas0;
111 	u32 mas4;
112 
113 	local_irq_save(flags);
114 	mtspr(SPRN_MAS6, 0);
115 	mas4 = mfspr(SPRN_MAS4);
116 	mtspr(SPRN_MAS4, mas4 & ~MAS4_TLBSEL_MASK);
117 	asm volatile("tlbsx 0, %0" : : "b" (eaddr & ~CONFIG_PAGE_OFFSET));
118 	mas0 = mfspr(SPRN_MAS0);
119 	mtspr(SPRN_MAS4, mas4);
120 	local_irq_restore(flags);
121 
122 	return mas0;
123 }
124 
125 /* sesel is for tlb1 only */
126 static inline void write_host_tlbe(struct kvmppc_vcpu_e500 *vcpu_e500,
127 		int tlbsel, int sesel, struct kvm_book3e_206_tlb_entry *stlbe)
128 {
129 	u32 mas0;
130 
131 	if (tlbsel == 0) {
132 		mas0 = get_host_mas0(stlbe->mas2);
133 		__write_host_tlbe(stlbe, mas0, vcpu_e500->vcpu.kvm->arch.lpid);
134 	} else {
135 		__write_host_tlbe(stlbe,
136 				  MAS0_TLBSEL(1) |
137 				  MAS0_ESEL(to_htlb1_esel(sesel)),
138 				  vcpu_e500->vcpu.kvm->arch.lpid);
139 	}
140 }
141 
142 /* sesel is for tlb1 only */
143 static void write_stlbe(struct kvmppc_vcpu_e500 *vcpu_e500,
144 			struct kvm_book3e_206_tlb_entry *gtlbe,
145 			struct kvm_book3e_206_tlb_entry *stlbe,
146 			int stlbsel, int sesel)
147 {
148 	int stid;
149 
150 	preempt_disable();
151 	stid = kvmppc_e500_get_tlb_stid(&vcpu_e500->vcpu, gtlbe);
152 
153 	stlbe->mas1 |= MAS1_TID(stid);
154 	write_host_tlbe(vcpu_e500, stlbsel, sesel, stlbe);
155 	preempt_enable();
156 }
157 
158 #ifdef CONFIG_KVM_E500V2
159 /* XXX should be a hook in the gva2hpa translation */
160 void kvmppc_map_magic(struct kvm_vcpu *vcpu)
161 {
162 	struct kvmppc_vcpu_e500 *vcpu_e500 = to_e500(vcpu);
163 	struct kvm_book3e_206_tlb_entry magic;
164 	ulong shared_page = ((ulong)vcpu->arch.shared) & PAGE_MASK;
165 	unsigned int stid;
166 	pfn_t pfn;
167 
168 	pfn = (pfn_t)virt_to_phys((void *)shared_page) >> PAGE_SHIFT;
169 	get_page(pfn_to_page(pfn));
170 
171 	preempt_disable();
172 	stid = kvmppc_e500_get_sid(vcpu_e500, 0, 0, 0, 0);
173 
174 	magic.mas1 = MAS1_VALID | MAS1_TS | MAS1_TID(stid) |
175 		     MAS1_TSIZE(BOOK3E_PAGESZ_4K);
176 	magic.mas2 = vcpu->arch.magic_page_ea | MAS2_M;
177 	magic.mas7_3 = ((u64)pfn << PAGE_SHIFT) |
178 		       MAS3_SW | MAS3_SR | MAS3_UW | MAS3_UR;
179 	magic.mas8 = 0;
180 
181 	__write_host_tlbe(&magic, MAS0_TLBSEL(1) | MAS0_ESEL(tlbcam_index), 0);
182 	preempt_enable();
183 }
184 #endif
185 
186 void inval_gtlbe_on_host(struct kvmppc_vcpu_e500 *vcpu_e500, int tlbsel,
187 			 int esel)
188 {
189 	struct kvm_book3e_206_tlb_entry *gtlbe =
190 		get_entry(vcpu_e500, tlbsel, esel);
191 	struct tlbe_ref *ref = &vcpu_e500->gtlb_priv[tlbsel][esel].ref;
192 
193 	/* Don't bother with unmapped entries */
194 	if (!(ref->flags & E500_TLB_VALID)) {
195 		WARN(ref->flags & (E500_TLB_BITMAP | E500_TLB_TLB0),
196 		     "%s: flags %x\n", __func__, ref->flags);
197 		WARN_ON(tlbsel == 1 && vcpu_e500->g2h_tlb1_map[esel]);
198 	}
199 
200 	if (tlbsel == 1 && ref->flags & E500_TLB_BITMAP) {
201 		u64 tmp = vcpu_e500->g2h_tlb1_map[esel];
202 		int hw_tlb_indx;
203 		unsigned long flags;
204 
205 		local_irq_save(flags);
206 		while (tmp) {
207 			hw_tlb_indx = __ilog2_u64(tmp & -tmp);
208 			mtspr(SPRN_MAS0,
209 			      MAS0_TLBSEL(1) |
210 			      MAS0_ESEL(to_htlb1_esel(hw_tlb_indx)));
211 			mtspr(SPRN_MAS1, 0);
212 			asm volatile("tlbwe");
213 			vcpu_e500->h2g_tlb1_rmap[hw_tlb_indx] = 0;
214 			tmp &= tmp - 1;
215 		}
216 		mb();
217 		vcpu_e500->g2h_tlb1_map[esel] = 0;
218 		ref->flags &= ~(E500_TLB_BITMAP | E500_TLB_VALID);
219 		local_irq_restore(flags);
220 	}
221 
222 	if (tlbsel == 1 && ref->flags & E500_TLB_TLB0) {
223 		/*
224 		 * TLB1 entry is backed by 4k pages. This should happen
225 		 * rarely and is not worth optimizing. Invalidate everything.
226 		 */
227 		kvmppc_e500_tlbil_all(vcpu_e500);
228 		ref->flags &= ~(E500_TLB_TLB0 | E500_TLB_VALID);
229 	}
230 
231 	/*
232 	 * If TLB entry is still valid then it's a TLB0 entry, and thus
233 	 * backed by at most one host tlbe per shadow pid
234 	 */
235 	if (ref->flags & E500_TLB_VALID)
236 		kvmppc_e500_tlbil_one(vcpu_e500, gtlbe);
237 
238 	/* Mark the TLB as not backed by the host anymore */
239 	ref->flags = 0;
240 }
241 
242 static inline int tlbe_is_writable(struct kvm_book3e_206_tlb_entry *tlbe)
243 {
244 	return tlbe->mas7_3 & (MAS3_SW|MAS3_UW);
245 }
246 
247 static inline void kvmppc_e500_ref_setup(struct tlbe_ref *ref,
248 					 struct kvm_book3e_206_tlb_entry *gtlbe,
249 					 pfn_t pfn, unsigned int wimg)
250 {
251 	ref->pfn = pfn;
252 	ref->flags = E500_TLB_VALID;
253 
254 	/* Use guest supplied MAS2_G and MAS2_E */
255 	ref->flags |= (gtlbe->mas2 & MAS2_ATTRIB_MASK) | wimg;
256 
257 	/* Mark the page accessed */
258 	kvm_set_pfn_accessed(pfn);
259 
260 	if (tlbe_is_writable(gtlbe))
261 		kvm_set_pfn_dirty(pfn);
262 }
263 
264 static inline void kvmppc_e500_ref_release(struct tlbe_ref *ref)
265 {
266 	if (ref->flags & E500_TLB_VALID) {
267 		/* FIXME: don't log bogus pfn for TLB1 */
268 		trace_kvm_booke206_ref_release(ref->pfn, ref->flags);
269 		ref->flags = 0;
270 	}
271 }
272 
273 static void clear_tlb1_bitmap(struct kvmppc_vcpu_e500 *vcpu_e500)
274 {
275 	if (vcpu_e500->g2h_tlb1_map)
276 		memset(vcpu_e500->g2h_tlb1_map, 0,
277 		       sizeof(u64) * vcpu_e500->gtlb_params[1].entries);
278 	if (vcpu_e500->h2g_tlb1_rmap)
279 		memset(vcpu_e500->h2g_tlb1_rmap, 0,
280 		       sizeof(unsigned int) * host_tlb_params[1].entries);
281 }
282 
283 static void clear_tlb_privs(struct kvmppc_vcpu_e500 *vcpu_e500)
284 {
285 	int tlbsel;
286 	int i;
287 
288 	for (tlbsel = 0; tlbsel <= 1; tlbsel++) {
289 		for (i = 0; i < vcpu_e500->gtlb_params[tlbsel].entries; i++) {
290 			struct tlbe_ref *ref =
291 				&vcpu_e500->gtlb_priv[tlbsel][i].ref;
292 			kvmppc_e500_ref_release(ref);
293 		}
294 	}
295 }
296 
297 void kvmppc_core_flush_tlb(struct kvm_vcpu *vcpu)
298 {
299 	struct kvmppc_vcpu_e500 *vcpu_e500 = to_e500(vcpu);
300 	kvmppc_e500_tlbil_all(vcpu_e500);
301 	clear_tlb_privs(vcpu_e500);
302 	clear_tlb1_bitmap(vcpu_e500);
303 }
304 
305 /* TID must be supplied by the caller */
306 static void kvmppc_e500_setup_stlbe(
307 	struct kvm_vcpu *vcpu,
308 	struct kvm_book3e_206_tlb_entry *gtlbe,
309 	int tsize, struct tlbe_ref *ref, u64 gvaddr,
310 	struct kvm_book3e_206_tlb_entry *stlbe)
311 {
312 	pfn_t pfn = ref->pfn;
313 	u32 pr = vcpu->arch.shared->msr & MSR_PR;
314 
315 	BUG_ON(!(ref->flags & E500_TLB_VALID));
316 
317 	/* Force IPROT=0 for all guest mappings. */
318 	stlbe->mas1 = MAS1_TSIZE(tsize) | get_tlb_sts(gtlbe) | MAS1_VALID;
319 	stlbe->mas2 = (gvaddr & MAS2_EPN) | (ref->flags & E500_TLB_MAS2_ATTR);
320 	stlbe->mas7_3 = ((u64)pfn << PAGE_SHIFT) |
321 			e500_shadow_mas3_attrib(gtlbe->mas7_3, pr);
322 }
323 
324 static inline int kvmppc_e500_shadow_map(struct kvmppc_vcpu_e500 *vcpu_e500,
325 	u64 gvaddr, gfn_t gfn, struct kvm_book3e_206_tlb_entry *gtlbe,
326 	int tlbsel, struct kvm_book3e_206_tlb_entry *stlbe,
327 	struct tlbe_ref *ref)
328 {
329 	struct kvm_memory_slot *slot;
330 	unsigned long pfn = 0; /* silence GCC warning */
331 	unsigned long hva;
332 	int pfnmap = 0;
333 	int tsize = BOOK3E_PAGESZ_4K;
334 	int ret = 0;
335 	unsigned long mmu_seq;
336 	struct kvm *kvm = vcpu_e500->vcpu.kvm;
337 	unsigned long tsize_pages = 0;
338 	pte_t *ptep;
339 	unsigned int wimg = 0;
340 	pgd_t *pgdir;
341 
342 	/* used to check for invalidations in progress */
343 	mmu_seq = kvm->mmu_notifier_seq;
344 	smp_rmb();
345 
346 	/*
347 	 * Translate guest physical to true physical, acquiring
348 	 * a page reference if it is normal, non-reserved memory.
349 	 *
350 	 * gfn_to_memslot() must succeed because otherwise we wouldn't
351 	 * have gotten this far.  Eventually we should just pass the slot
352 	 * pointer through from the first lookup.
353 	 */
354 	slot = gfn_to_memslot(vcpu_e500->vcpu.kvm, gfn);
355 	hva = gfn_to_hva_memslot(slot, gfn);
356 
357 	if (tlbsel == 1) {
358 		struct vm_area_struct *vma;
359 		down_read(&current->mm->mmap_sem);
360 
361 		vma = find_vma(current->mm, hva);
362 		if (vma && hva >= vma->vm_start &&
363 		    (vma->vm_flags & VM_PFNMAP)) {
364 			/*
365 			 * This VMA is a physically contiguous region (e.g.
366 			 * /dev/mem) that bypasses normal Linux page
367 			 * management.  Find the overlap between the
368 			 * vma and the memslot.
369 			 */
370 
371 			unsigned long start, end;
372 			unsigned long slot_start, slot_end;
373 
374 			pfnmap = 1;
375 
376 			start = vma->vm_pgoff;
377 			end = start +
378 			      ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT);
379 
380 			pfn = start + ((hva - vma->vm_start) >> PAGE_SHIFT);
381 
382 			slot_start = pfn - (gfn - slot->base_gfn);
383 			slot_end = slot_start + slot->npages;
384 
385 			if (start < slot_start)
386 				start = slot_start;
387 			if (end > slot_end)
388 				end = slot_end;
389 
390 			tsize = (gtlbe->mas1 & MAS1_TSIZE_MASK) >>
391 				MAS1_TSIZE_SHIFT;
392 
393 			/*
394 			 * e500 doesn't implement the lowest tsize bit,
395 			 * or 1K pages.
396 			 */
397 			tsize = max(BOOK3E_PAGESZ_4K, tsize & ~1);
398 
399 			/*
400 			 * Now find the largest tsize (up to what the guest
401 			 * requested) that will cover gfn, stay within the
402 			 * range, and for which gfn and pfn are mutually
403 			 * aligned.
404 			 */
405 
406 			for (; tsize > BOOK3E_PAGESZ_4K; tsize -= 2) {
407 				unsigned long gfn_start, gfn_end;
408 				tsize_pages = 1 << (tsize - 2);
409 
410 				gfn_start = gfn & ~(tsize_pages - 1);
411 				gfn_end = gfn_start + tsize_pages;
412 
413 				if (gfn_start + pfn - gfn < start)
414 					continue;
415 				if (gfn_end + pfn - gfn > end)
416 					continue;
417 				if ((gfn & (tsize_pages - 1)) !=
418 				    (pfn & (tsize_pages - 1)))
419 					continue;
420 
421 				gvaddr &= ~((tsize_pages << PAGE_SHIFT) - 1);
422 				pfn &= ~(tsize_pages - 1);
423 				break;
424 			}
425 		} else if (vma && hva >= vma->vm_start &&
426 			   (vma->vm_flags & VM_HUGETLB)) {
427 			unsigned long psize = vma_kernel_pagesize(vma);
428 
429 			tsize = (gtlbe->mas1 & MAS1_TSIZE_MASK) >>
430 				MAS1_TSIZE_SHIFT;
431 
432 			/*
433 			 * Take the largest page size that satisfies both host
434 			 * and guest mapping
435 			 */
436 			tsize = min(__ilog2(psize) - 10, tsize);
437 
438 			/*
439 			 * e500 doesn't implement the lowest tsize bit,
440 			 * or 1K pages.
441 			 */
442 			tsize = max(BOOK3E_PAGESZ_4K, tsize & ~1);
443 		}
444 
445 		up_read(&current->mm->mmap_sem);
446 	}
447 
448 	if (likely(!pfnmap)) {
449 		tsize_pages = 1 << (tsize + 10 - PAGE_SHIFT);
450 		pfn = gfn_to_pfn_memslot(slot, gfn);
451 		if (is_error_noslot_pfn(pfn)) {
452 			if (printk_ratelimit())
453 				pr_err("%s: real page not found for gfn %lx\n",
454 				       __func__, (long)gfn);
455 			return -EINVAL;
456 		}
457 
458 		/* Align guest and physical address to page map boundaries */
459 		pfn &= ~(tsize_pages - 1);
460 		gvaddr &= ~((tsize_pages << PAGE_SHIFT) - 1);
461 	}
462 
463 	spin_lock(&kvm->mmu_lock);
464 	if (mmu_notifier_retry(kvm, mmu_seq)) {
465 		ret = -EAGAIN;
466 		goto out;
467 	}
468 
469 
470 	pgdir = vcpu_e500->vcpu.arch.pgdir;
471 	ptep = lookup_linux_ptep(pgdir, hva, &tsize_pages);
472 	if (pte_present(*ptep))
473 		wimg = (*ptep >> PTE_WIMGE_SHIFT) & MAS2_WIMGE_MASK;
474 	else {
475 		if (printk_ratelimit())
476 			pr_err("%s: pte not present: gfn %lx, pfn %lx\n",
477 				__func__, (long)gfn, pfn);
478 		ret = -EINVAL;
479 		goto out;
480 	}
481 	kvmppc_e500_ref_setup(ref, gtlbe, pfn, wimg);
482 
483 	kvmppc_e500_setup_stlbe(&vcpu_e500->vcpu, gtlbe, tsize,
484 				ref, gvaddr, stlbe);
485 
486 	/* Clear i-cache for new pages */
487 	kvmppc_mmu_flush_icache(pfn);
488 
489 out:
490 	spin_unlock(&kvm->mmu_lock);
491 
492 	/* Drop refcount on page, so that mmu notifiers can clear it */
493 	kvm_release_pfn_clean(pfn);
494 
495 	return ret;
496 }
497 
498 /* XXX only map the one-one case, for now use TLB0 */
499 static int kvmppc_e500_tlb0_map(struct kvmppc_vcpu_e500 *vcpu_e500, int esel,
500 				struct kvm_book3e_206_tlb_entry *stlbe)
501 {
502 	struct kvm_book3e_206_tlb_entry *gtlbe;
503 	struct tlbe_ref *ref;
504 	int stlbsel = 0;
505 	int sesel = 0;
506 	int r;
507 
508 	gtlbe = get_entry(vcpu_e500, 0, esel);
509 	ref = &vcpu_e500->gtlb_priv[0][esel].ref;
510 
511 	r = kvmppc_e500_shadow_map(vcpu_e500, get_tlb_eaddr(gtlbe),
512 			get_tlb_raddr(gtlbe) >> PAGE_SHIFT,
513 			gtlbe, 0, stlbe, ref);
514 	if (r)
515 		return r;
516 
517 	write_stlbe(vcpu_e500, gtlbe, stlbe, stlbsel, sesel);
518 
519 	return 0;
520 }
521 
522 static int kvmppc_e500_tlb1_map_tlb1(struct kvmppc_vcpu_e500 *vcpu_e500,
523 				     struct tlbe_ref *ref,
524 				     int esel)
525 {
526 	unsigned int sesel = vcpu_e500->host_tlb1_nv++;
527 
528 	if (unlikely(vcpu_e500->host_tlb1_nv >= tlb1_max_shadow_size()))
529 		vcpu_e500->host_tlb1_nv = 0;
530 
531 	if (vcpu_e500->h2g_tlb1_rmap[sesel]) {
532 		unsigned int idx = vcpu_e500->h2g_tlb1_rmap[sesel] - 1;
533 		vcpu_e500->g2h_tlb1_map[idx] &= ~(1ULL << sesel);
534 	}
535 
536 	vcpu_e500->gtlb_priv[1][esel].ref.flags |= E500_TLB_BITMAP;
537 	vcpu_e500->g2h_tlb1_map[esel] |= (u64)1 << sesel;
538 	vcpu_e500->h2g_tlb1_rmap[sesel] = esel + 1;
539 	WARN_ON(!(ref->flags & E500_TLB_VALID));
540 
541 	return sesel;
542 }
543 
544 /* Caller must ensure that the specified guest TLB entry is safe to insert into
545  * the shadow TLB. */
546 /* For both one-one and one-to-many */
547 static int kvmppc_e500_tlb1_map(struct kvmppc_vcpu_e500 *vcpu_e500,
548 		u64 gvaddr, gfn_t gfn, struct kvm_book3e_206_tlb_entry *gtlbe,
549 		struct kvm_book3e_206_tlb_entry *stlbe, int esel)
550 {
551 	struct tlbe_ref *ref = &vcpu_e500->gtlb_priv[1][esel].ref;
552 	int sesel;
553 	int r;
554 
555 	r = kvmppc_e500_shadow_map(vcpu_e500, gvaddr, gfn, gtlbe, 1, stlbe,
556 				   ref);
557 	if (r)
558 		return r;
559 
560 	/* Use TLB0 when we can only map a page with 4k */
561 	if (get_tlb_tsize(stlbe) == BOOK3E_PAGESZ_4K) {
562 		vcpu_e500->gtlb_priv[1][esel].ref.flags |= E500_TLB_TLB0;
563 		write_stlbe(vcpu_e500, gtlbe, stlbe, 0, 0);
564 		return 0;
565 	}
566 
567 	/* Otherwise map into TLB1 */
568 	sesel = kvmppc_e500_tlb1_map_tlb1(vcpu_e500, ref, esel);
569 	write_stlbe(vcpu_e500, gtlbe, stlbe, 1, sesel);
570 
571 	return 0;
572 }
573 
574 void kvmppc_mmu_map(struct kvm_vcpu *vcpu, u64 eaddr, gpa_t gpaddr,
575 		    unsigned int index)
576 {
577 	struct kvmppc_vcpu_e500 *vcpu_e500 = to_e500(vcpu);
578 	struct tlbe_priv *priv;
579 	struct kvm_book3e_206_tlb_entry *gtlbe, stlbe;
580 	int tlbsel = tlbsel_of(index);
581 	int esel = esel_of(index);
582 
583 	gtlbe = get_entry(vcpu_e500, tlbsel, esel);
584 
585 	switch (tlbsel) {
586 	case 0:
587 		priv = &vcpu_e500->gtlb_priv[tlbsel][esel];
588 
589 		/* Triggers after clear_tlb_privs or on initial mapping */
590 		if (!(priv->ref.flags & E500_TLB_VALID)) {
591 			kvmppc_e500_tlb0_map(vcpu_e500, esel, &stlbe);
592 		} else {
593 			kvmppc_e500_setup_stlbe(vcpu, gtlbe, BOOK3E_PAGESZ_4K,
594 						&priv->ref, eaddr, &stlbe);
595 			write_stlbe(vcpu_e500, gtlbe, &stlbe, 0, 0);
596 		}
597 		break;
598 
599 	case 1: {
600 		gfn_t gfn = gpaddr >> PAGE_SHIFT;
601 		kvmppc_e500_tlb1_map(vcpu_e500, eaddr, gfn, gtlbe, &stlbe,
602 				     esel);
603 		break;
604 	}
605 
606 	default:
607 		BUG();
608 		break;
609 	}
610 }
611 
612 #ifdef CONFIG_KVM_BOOKE_HV
613 int kvmppc_load_last_inst(struct kvm_vcpu *vcpu, enum instruction_type type,
614 			  u32 *instr)
615 {
616 	gva_t geaddr;
617 	hpa_t addr;
618 	hfn_t pfn;
619 	hva_t eaddr;
620 	u32 mas1, mas2, mas3;
621 	u64 mas7_mas3;
622 	struct page *page;
623 	unsigned int addr_space, psize_shift;
624 	bool pr;
625 	unsigned long flags;
626 
627 	/* Search TLB for guest pc to get the real address */
628 	geaddr = kvmppc_get_pc(vcpu);
629 
630 	addr_space = (vcpu->arch.shared->msr & MSR_IS) >> MSR_IR_LG;
631 
632 	local_irq_save(flags);
633 	mtspr(SPRN_MAS6, (vcpu->arch.pid << MAS6_SPID_SHIFT) | addr_space);
634 	mtspr(SPRN_MAS5, MAS5_SGS | get_lpid(vcpu));
635 	asm volatile("tlbsx 0, %[geaddr]\n" : :
636 		     [geaddr] "r" (geaddr));
637 	mtspr(SPRN_MAS5, 0);
638 	mtspr(SPRN_MAS8, 0);
639 	mas1 = mfspr(SPRN_MAS1);
640 	mas2 = mfspr(SPRN_MAS2);
641 	mas3 = mfspr(SPRN_MAS3);
642 #ifdef CONFIG_64BIT
643 	mas7_mas3 = mfspr(SPRN_MAS7_MAS3);
644 #else
645 	mas7_mas3 = ((u64)mfspr(SPRN_MAS7) << 32) | mas3;
646 #endif
647 	local_irq_restore(flags);
648 
649 	/*
650 	 * If the TLB entry for guest pc was evicted, return to the guest.
651 	 * There are high chances to find a valid TLB entry next time.
652 	 */
653 	if (!(mas1 & MAS1_VALID))
654 		return EMULATE_AGAIN;
655 
656 	/*
657 	 * Another thread may rewrite the TLB entry in parallel, don't
658 	 * execute from the address if the execute permission is not set
659 	 */
660 	pr = vcpu->arch.shared->msr & MSR_PR;
661 	if (unlikely((pr && !(mas3 & MAS3_UX)) ||
662 		     (!pr && !(mas3 & MAS3_SX)))) {
663 		pr_err_ratelimited(
664 			"%s: Instruction emulation from guest address %08lx without execute permission\n",
665 			__func__, geaddr);
666 		return EMULATE_AGAIN;
667 	}
668 
669 	/*
670 	 * The real address will be mapped by a cacheable, memory coherent,
671 	 * write-back page. Check for mismatches when LRAT is used.
672 	 */
673 	if (has_feature(vcpu, VCPU_FTR_MMU_V2) &&
674 	    unlikely((mas2 & MAS2_I) || (mas2 & MAS2_W) || !(mas2 & MAS2_M))) {
675 		pr_err_ratelimited(
676 			"%s: Instruction emulation from guest address %08lx mismatches storage attributes\n",
677 			__func__, geaddr);
678 		return EMULATE_AGAIN;
679 	}
680 
681 	/* Get pfn */
682 	psize_shift = MAS1_GET_TSIZE(mas1) + 10;
683 	addr = (mas7_mas3 & (~0ULL << psize_shift)) |
684 	       (geaddr & ((1ULL << psize_shift) - 1ULL));
685 	pfn = addr >> PAGE_SHIFT;
686 
687 	/* Guard against emulation from devices area */
688 	if (unlikely(!page_is_ram(pfn))) {
689 		pr_err_ratelimited("%s: Instruction emulation from non-RAM host address %08llx is not supported\n",
690 			 __func__, addr);
691 		return EMULATE_AGAIN;
692 	}
693 
694 	/* Map a page and get guest's instruction */
695 	page = pfn_to_page(pfn);
696 	eaddr = (unsigned long)kmap_atomic(page);
697 	*instr = *(u32 *)(eaddr | (unsigned long)(addr & ~PAGE_MASK));
698 	kunmap_atomic((u32 *)eaddr);
699 
700 	return EMULATE_DONE;
701 }
702 #else
703 int kvmppc_load_last_inst(struct kvm_vcpu *vcpu, enum instruction_type type,
704 			  u32 *instr)
705 {
706 	return EMULATE_AGAIN;
707 }
708 #endif
709 
710 /************* MMU Notifiers *************/
711 
712 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
713 {
714 	trace_kvm_unmap_hva(hva);
715 
716 	/*
717 	 * Flush all shadow tlb entries everywhere. This is slow, but
718 	 * we are 100% sure that we catch the to be unmapped page
719 	 */
720 	kvm_flush_remote_tlbs(kvm);
721 
722 	return 0;
723 }
724 
725 int kvm_unmap_hva_range(struct kvm *kvm, unsigned long start, unsigned long end)
726 {
727 	/* kvm_unmap_hva flushes everything anyways */
728 	kvm_unmap_hva(kvm, start);
729 
730 	return 0;
731 }
732 
733 int kvm_age_hva(struct kvm *kvm, unsigned long start, unsigned long end)
734 {
735 	/* XXX could be more clever ;) */
736 	return 0;
737 }
738 
739 int kvm_test_age_hva(struct kvm *kvm, unsigned long hva)
740 {
741 	/* XXX could be more clever ;) */
742 	return 0;
743 }
744 
745 void kvm_set_spte_hva(struct kvm *kvm, unsigned long hva, pte_t pte)
746 {
747 	/* The page will get remapped properly on its next fault */
748 	kvm_unmap_hva(kvm, hva);
749 }
750 
751 /*****************************************/
752 
753 int e500_mmu_host_init(struct kvmppc_vcpu_e500 *vcpu_e500)
754 {
755 	host_tlb_params[0].entries = mfspr(SPRN_TLB0CFG) & TLBnCFG_N_ENTRY;
756 	host_tlb_params[1].entries = mfspr(SPRN_TLB1CFG) & TLBnCFG_N_ENTRY;
757 
758 	/*
759 	 * This should never happen on real e500 hardware, but is
760 	 * architecturally possible -- e.g. in some weird nested
761 	 * virtualization case.
762 	 */
763 	if (host_tlb_params[0].entries == 0 ||
764 	    host_tlb_params[1].entries == 0) {
765 		pr_err("%s: need to know host tlb size\n", __func__);
766 		return -ENODEV;
767 	}
768 
769 	host_tlb_params[0].ways = (mfspr(SPRN_TLB0CFG) & TLBnCFG_ASSOC) >>
770 				  TLBnCFG_ASSOC_SHIFT;
771 	host_tlb_params[1].ways = host_tlb_params[1].entries;
772 
773 	if (!is_power_of_2(host_tlb_params[0].entries) ||
774 	    !is_power_of_2(host_tlb_params[0].ways) ||
775 	    host_tlb_params[0].entries < host_tlb_params[0].ways ||
776 	    host_tlb_params[0].ways == 0) {
777 		pr_err("%s: bad tlb0 host config: %u entries %u ways\n",
778 		       __func__, host_tlb_params[0].entries,
779 		       host_tlb_params[0].ways);
780 		return -ENODEV;
781 	}
782 
783 	host_tlb_params[0].sets =
784 		host_tlb_params[0].entries / host_tlb_params[0].ways;
785 	host_tlb_params[1].sets = 1;
786 
787 	vcpu_e500->h2g_tlb1_rmap = kzalloc(sizeof(unsigned int) *
788 					   host_tlb_params[1].entries,
789 					   GFP_KERNEL);
790 	if (!vcpu_e500->h2g_tlb1_rmap)
791 		return -EINVAL;
792 
793 	return 0;
794 }
795 
796 void e500_mmu_host_uninit(struct kvmppc_vcpu_e500 *vcpu_e500)
797 {
798 	kfree(vcpu_e500->h2g_tlb1_rmap);
799 }
800