xref: /linux/arch/powerpc/kvm/e500_mmu_host.c (revision 3932b9ca55b0be314a36d3e84faff3e823c081f5)
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 {
74 	unsigned long flags;
75 
76 	local_irq_save(flags);
77 	mtspr(SPRN_MAS0, mas0);
78 	mtspr(SPRN_MAS1, stlbe->mas1);
79 	mtspr(SPRN_MAS2, (unsigned long)stlbe->mas2);
80 	mtspr(SPRN_MAS3, (u32)stlbe->mas7_3);
81 	mtspr(SPRN_MAS7, (u32)(stlbe->mas7_3 >> 32));
82 #ifdef CONFIG_KVM_BOOKE_HV
83 	mtspr(SPRN_MAS8, stlbe->mas8);
84 #endif
85 	asm volatile("isync; tlbwe" : : : "memory");
86 
87 #ifdef CONFIG_KVM_BOOKE_HV
88 	/* Must clear mas8 for other host tlbwe's */
89 	mtspr(SPRN_MAS8, 0);
90 	isync();
91 #endif
92 	local_irq_restore(flags);
93 
94 	trace_kvm_booke206_stlb_write(mas0, stlbe->mas8, stlbe->mas1,
95 	                              stlbe->mas2, stlbe->mas7_3);
96 }
97 
98 /*
99  * Acquire a mas0 with victim hint, as if we just took a TLB miss.
100  *
101  * We don't care about the address we're searching for, other than that it's
102  * in the right set and is not present in the TLB.  Using a zero PID and a
103  * userspace address means we don't have to set and then restore MAS5, or
104  * calculate a proper MAS6 value.
105  */
106 static u32 get_host_mas0(unsigned long eaddr)
107 {
108 	unsigned long flags;
109 	u32 mas0;
110 	u32 mas4;
111 
112 	local_irq_save(flags);
113 	mtspr(SPRN_MAS6, 0);
114 	mas4 = mfspr(SPRN_MAS4);
115 	mtspr(SPRN_MAS4, mas4 & ~MAS4_TLBSEL_MASK);
116 	asm volatile("tlbsx 0, %0" : : "b" (eaddr & ~CONFIG_PAGE_OFFSET));
117 	mas0 = mfspr(SPRN_MAS0);
118 	mtspr(SPRN_MAS4, mas4);
119 	local_irq_restore(flags);
120 
121 	return mas0;
122 }
123 
124 /* sesel is for tlb1 only */
125 static inline void write_host_tlbe(struct kvmppc_vcpu_e500 *vcpu_e500,
126 		int tlbsel, int sesel, struct kvm_book3e_206_tlb_entry *stlbe)
127 {
128 	u32 mas0;
129 
130 	if (tlbsel == 0) {
131 		mas0 = get_host_mas0(stlbe->mas2);
132 		__write_host_tlbe(stlbe, mas0);
133 	} else {
134 		__write_host_tlbe(stlbe,
135 				  MAS0_TLBSEL(1) |
136 				  MAS0_ESEL(to_htlb1_esel(sesel)));
137 	}
138 }
139 
140 /* sesel is for tlb1 only */
141 static void write_stlbe(struct kvmppc_vcpu_e500 *vcpu_e500,
142 			struct kvm_book3e_206_tlb_entry *gtlbe,
143 			struct kvm_book3e_206_tlb_entry *stlbe,
144 			int stlbsel, int sesel)
145 {
146 	int stid;
147 
148 	preempt_disable();
149 	stid = kvmppc_e500_get_tlb_stid(&vcpu_e500->vcpu, gtlbe);
150 
151 	stlbe->mas1 |= MAS1_TID(stid);
152 	write_host_tlbe(vcpu_e500, stlbsel, sesel, stlbe);
153 	preempt_enable();
154 }
155 
156 #ifdef CONFIG_KVM_E500V2
157 /* XXX should be a hook in the gva2hpa translation */
158 void kvmppc_map_magic(struct kvm_vcpu *vcpu)
159 {
160 	struct kvmppc_vcpu_e500 *vcpu_e500 = to_e500(vcpu);
161 	struct kvm_book3e_206_tlb_entry magic;
162 	ulong shared_page = ((ulong)vcpu->arch.shared) & PAGE_MASK;
163 	unsigned int stid;
164 	pfn_t pfn;
165 
166 	pfn = (pfn_t)virt_to_phys((void *)shared_page) >> PAGE_SHIFT;
167 	get_page(pfn_to_page(pfn));
168 
169 	preempt_disable();
170 	stid = kvmppc_e500_get_sid(vcpu_e500, 0, 0, 0, 0);
171 
172 	magic.mas1 = MAS1_VALID | MAS1_TS | MAS1_TID(stid) |
173 		     MAS1_TSIZE(BOOK3E_PAGESZ_4K);
174 	magic.mas2 = vcpu->arch.magic_page_ea | MAS2_M;
175 	magic.mas7_3 = ((u64)pfn << PAGE_SHIFT) |
176 		       MAS3_SW | MAS3_SR | MAS3_UW | MAS3_UR;
177 	magic.mas8 = 0;
178 
179 	__write_host_tlbe(&magic, MAS0_TLBSEL(1) | MAS0_ESEL(tlbcam_index));
180 	preempt_enable();
181 }
182 #endif
183 
184 void inval_gtlbe_on_host(struct kvmppc_vcpu_e500 *vcpu_e500, int tlbsel,
185 			 int esel)
186 {
187 	struct kvm_book3e_206_tlb_entry *gtlbe =
188 		get_entry(vcpu_e500, tlbsel, esel);
189 	struct tlbe_ref *ref = &vcpu_e500->gtlb_priv[tlbsel][esel].ref;
190 
191 	/* Don't bother with unmapped entries */
192 	if (!(ref->flags & E500_TLB_VALID)) {
193 		WARN(ref->flags & (E500_TLB_BITMAP | E500_TLB_TLB0),
194 		     "%s: flags %x\n", __func__, ref->flags);
195 		WARN_ON(tlbsel == 1 && vcpu_e500->g2h_tlb1_map[esel]);
196 	}
197 
198 	if (tlbsel == 1 && ref->flags & E500_TLB_BITMAP) {
199 		u64 tmp = vcpu_e500->g2h_tlb1_map[esel];
200 		int hw_tlb_indx;
201 		unsigned long flags;
202 
203 		local_irq_save(flags);
204 		while (tmp) {
205 			hw_tlb_indx = __ilog2_u64(tmp & -tmp);
206 			mtspr(SPRN_MAS0,
207 			      MAS0_TLBSEL(1) |
208 			      MAS0_ESEL(to_htlb1_esel(hw_tlb_indx)));
209 			mtspr(SPRN_MAS1, 0);
210 			asm volatile("tlbwe");
211 			vcpu_e500->h2g_tlb1_rmap[hw_tlb_indx] = 0;
212 			tmp &= tmp - 1;
213 		}
214 		mb();
215 		vcpu_e500->g2h_tlb1_map[esel] = 0;
216 		ref->flags &= ~(E500_TLB_BITMAP | E500_TLB_VALID);
217 		local_irq_restore(flags);
218 	}
219 
220 	if (tlbsel == 1 && ref->flags & E500_TLB_TLB0) {
221 		/*
222 		 * TLB1 entry is backed by 4k pages. This should happen
223 		 * rarely and is not worth optimizing. Invalidate everything.
224 		 */
225 		kvmppc_e500_tlbil_all(vcpu_e500);
226 		ref->flags &= ~(E500_TLB_TLB0 | E500_TLB_VALID);
227 	}
228 
229 	/*
230 	 * If TLB entry is still valid then it's a TLB0 entry, and thus
231 	 * backed by at most one host tlbe per shadow pid
232 	 */
233 	if (ref->flags & E500_TLB_VALID)
234 		kvmppc_e500_tlbil_one(vcpu_e500, gtlbe);
235 
236 	/* Mark the TLB as not backed by the host anymore */
237 	ref->flags = 0;
238 }
239 
240 static inline int tlbe_is_writable(struct kvm_book3e_206_tlb_entry *tlbe)
241 {
242 	return tlbe->mas7_3 & (MAS3_SW|MAS3_UW);
243 }
244 
245 static inline void kvmppc_e500_ref_setup(struct tlbe_ref *ref,
246 					 struct kvm_book3e_206_tlb_entry *gtlbe,
247 					 pfn_t pfn, unsigned int wimg)
248 {
249 	ref->pfn = pfn;
250 	ref->flags = E500_TLB_VALID;
251 
252 	/* Use guest supplied MAS2_G and MAS2_E */
253 	ref->flags |= (gtlbe->mas2 & MAS2_ATTRIB_MASK) | wimg;
254 
255 	/* Mark the page accessed */
256 	kvm_set_pfn_accessed(pfn);
257 
258 	if (tlbe_is_writable(gtlbe))
259 		kvm_set_pfn_dirty(pfn);
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 	pfn_t pfn = ref->pfn;
311 	u32 pr = vcpu->arch.shared->msr & MSR_PR;
312 
313 	BUG_ON(!(ref->flags & E500_TLB_VALID));
314 
315 	/* Force IPROT=0 for all guest mappings. */
316 	stlbe->mas1 = MAS1_TSIZE(tsize) | get_tlb_sts(gtlbe) | MAS1_VALID;
317 	stlbe->mas2 = (gvaddr & MAS2_EPN) | (ref->flags & E500_TLB_MAS2_ATTR);
318 	stlbe->mas7_3 = ((u64)pfn << PAGE_SHIFT) |
319 			e500_shadow_mas3_attrib(gtlbe->mas7_3, pr);
320 
321 #ifdef CONFIG_KVM_BOOKE_HV
322 	stlbe->mas8 = MAS8_TGS | vcpu->kvm->arch.lpid;
323 #endif
324 }
325 
326 static inline int kvmppc_e500_shadow_map(struct kvmppc_vcpu_e500 *vcpu_e500,
327 	u64 gvaddr, gfn_t gfn, struct kvm_book3e_206_tlb_entry *gtlbe,
328 	int tlbsel, struct kvm_book3e_206_tlb_entry *stlbe,
329 	struct tlbe_ref *ref)
330 {
331 	struct kvm_memory_slot *slot;
332 	unsigned long pfn = 0; /* silence GCC warning */
333 	unsigned long hva;
334 	int pfnmap = 0;
335 	int tsize = BOOK3E_PAGESZ_4K;
336 	int ret = 0;
337 	unsigned long mmu_seq;
338 	struct kvm *kvm = vcpu_e500->vcpu.kvm;
339 	unsigned long tsize_pages = 0;
340 	pte_t *ptep;
341 	unsigned int wimg = 0;
342 	pgd_t *pgdir;
343 
344 	/* used to check for invalidations in progress */
345 	mmu_seq = kvm->mmu_notifier_seq;
346 	smp_rmb();
347 
348 	/*
349 	 * Translate guest physical to true physical, acquiring
350 	 * a page reference if it is normal, non-reserved memory.
351 	 *
352 	 * gfn_to_memslot() must succeed because otherwise we wouldn't
353 	 * have gotten this far.  Eventually we should just pass the slot
354 	 * pointer through from the first lookup.
355 	 */
356 	slot = gfn_to_memslot(vcpu_e500->vcpu.kvm, gfn);
357 	hva = gfn_to_hva_memslot(slot, gfn);
358 
359 	if (tlbsel == 1) {
360 		struct vm_area_struct *vma;
361 		down_read(&current->mm->mmap_sem);
362 
363 		vma = find_vma(current->mm, hva);
364 		if (vma && hva >= vma->vm_start &&
365 		    (vma->vm_flags & VM_PFNMAP)) {
366 			/*
367 			 * This VMA is a physically contiguous region (e.g.
368 			 * /dev/mem) that bypasses normal Linux page
369 			 * management.  Find the overlap between the
370 			 * vma and the memslot.
371 			 */
372 
373 			unsigned long start, end;
374 			unsigned long slot_start, slot_end;
375 
376 			pfnmap = 1;
377 
378 			start = vma->vm_pgoff;
379 			end = start +
380 			      ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT);
381 
382 			pfn = start + ((hva - vma->vm_start) >> PAGE_SHIFT);
383 
384 			slot_start = pfn - (gfn - slot->base_gfn);
385 			slot_end = slot_start + slot->npages;
386 
387 			if (start < slot_start)
388 				start = slot_start;
389 			if (end > slot_end)
390 				end = slot_end;
391 
392 			tsize = (gtlbe->mas1 & MAS1_TSIZE_MASK) >>
393 				MAS1_TSIZE_SHIFT;
394 
395 			/*
396 			 * e500 doesn't implement the lowest tsize bit,
397 			 * or 1K pages.
398 			 */
399 			tsize = max(BOOK3E_PAGESZ_4K, tsize & ~1);
400 
401 			/*
402 			 * Now find the largest tsize (up to what the guest
403 			 * requested) that will cover gfn, stay within the
404 			 * range, and for which gfn and pfn are mutually
405 			 * aligned.
406 			 */
407 
408 			for (; tsize > BOOK3E_PAGESZ_4K; tsize -= 2) {
409 				unsigned long gfn_start, gfn_end;
410 				tsize_pages = 1 << (tsize - 2);
411 
412 				gfn_start = gfn & ~(tsize_pages - 1);
413 				gfn_end = gfn_start + tsize_pages;
414 
415 				if (gfn_start + pfn - gfn < start)
416 					continue;
417 				if (gfn_end + pfn - gfn > end)
418 					continue;
419 				if ((gfn & (tsize_pages - 1)) !=
420 				    (pfn & (tsize_pages - 1)))
421 					continue;
422 
423 				gvaddr &= ~((tsize_pages << PAGE_SHIFT) - 1);
424 				pfn &= ~(tsize_pages - 1);
425 				break;
426 			}
427 		} else if (vma && hva >= vma->vm_start &&
428 			   (vma->vm_flags & VM_HUGETLB)) {
429 			unsigned long psize = vma_kernel_pagesize(vma);
430 
431 			tsize = (gtlbe->mas1 & MAS1_TSIZE_MASK) >>
432 				MAS1_TSIZE_SHIFT;
433 
434 			/*
435 			 * Take the largest page size that satisfies both host
436 			 * and guest mapping
437 			 */
438 			tsize = min(__ilog2(psize) - 10, tsize);
439 
440 			/*
441 			 * e500 doesn't implement the lowest tsize bit,
442 			 * or 1K pages.
443 			 */
444 			tsize = max(BOOK3E_PAGESZ_4K, tsize & ~1);
445 		}
446 
447 		up_read(&current->mm->mmap_sem);
448 	}
449 
450 	if (likely(!pfnmap)) {
451 		tsize_pages = 1 << (tsize + 10 - PAGE_SHIFT);
452 		pfn = gfn_to_pfn_memslot(slot, gfn);
453 		if (is_error_noslot_pfn(pfn)) {
454 			if (printk_ratelimit())
455 				pr_err("%s: real page not found for gfn %lx\n",
456 				       __func__, (long)gfn);
457 			return -EINVAL;
458 		}
459 
460 		/* Align guest and physical address to page map boundaries */
461 		pfn &= ~(tsize_pages - 1);
462 		gvaddr &= ~((tsize_pages << PAGE_SHIFT) - 1);
463 	}
464 
465 	spin_lock(&kvm->mmu_lock);
466 	if (mmu_notifier_retry(kvm, mmu_seq)) {
467 		ret = -EAGAIN;
468 		goto out;
469 	}
470 
471 
472 	pgdir = vcpu_e500->vcpu.arch.pgdir;
473 	ptep = lookup_linux_ptep(pgdir, hva, &tsize_pages);
474 	if (pte_present(*ptep))
475 		wimg = (*ptep >> PTE_WIMGE_SHIFT) & MAS2_WIMGE_MASK;
476 	else {
477 		if (printk_ratelimit())
478 			pr_err("%s: pte not present: gfn %lx, pfn %lx\n",
479 				__func__, (long)gfn, pfn);
480 		ret = -EINVAL;
481 		goto out;
482 	}
483 	kvmppc_e500_ref_setup(ref, gtlbe, pfn, wimg);
484 
485 	kvmppc_e500_setup_stlbe(&vcpu_e500->vcpu, gtlbe, tsize,
486 				ref, gvaddr, stlbe);
487 
488 	/* Clear i-cache for new pages */
489 	kvmppc_mmu_flush_icache(pfn);
490 
491 out:
492 	spin_unlock(&kvm->mmu_lock);
493 
494 	/* Drop refcount on page, so that mmu notifiers can clear it */
495 	kvm_release_pfn_clean(pfn);
496 
497 	return ret;
498 }
499 
500 /* XXX only map the one-one case, for now use TLB0 */
501 static int kvmppc_e500_tlb0_map(struct kvmppc_vcpu_e500 *vcpu_e500, int esel,
502 				struct kvm_book3e_206_tlb_entry *stlbe)
503 {
504 	struct kvm_book3e_206_tlb_entry *gtlbe;
505 	struct tlbe_ref *ref;
506 	int stlbsel = 0;
507 	int sesel = 0;
508 	int r;
509 
510 	gtlbe = get_entry(vcpu_e500, 0, esel);
511 	ref = &vcpu_e500->gtlb_priv[0][esel].ref;
512 
513 	r = kvmppc_e500_shadow_map(vcpu_e500, get_tlb_eaddr(gtlbe),
514 			get_tlb_raddr(gtlbe) >> PAGE_SHIFT,
515 			gtlbe, 0, stlbe, ref);
516 	if (r)
517 		return r;
518 
519 	write_stlbe(vcpu_e500, gtlbe, stlbe, stlbsel, sesel);
520 
521 	return 0;
522 }
523 
524 static int kvmppc_e500_tlb1_map_tlb1(struct kvmppc_vcpu_e500 *vcpu_e500,
525 				     struct tlbe_ref *ref,
526 				     int esel)
527 {
528 	unsigned int sesel = vcpu_e500->host_tlb1_nv++;
529 
530 	if (unlikely(vcpu_e500->host_tlb1_nv >= tlb1_max_shadow_size()))
531 		vcpu_e500->host_tlb1_nv = 0;
532 
533 	if (vcpu_e500->h2g_tlb1_rmap[sesel]) {
534 		unsigned int idx = vcpu_e500->h2g_tlb1_rmap[sesel] - 1;
535 		vcpu_e500->g2h_tlb1_map[idx] &= ~(1ULL << sesel);
536 	}
537 
538 	vcpu_e500->gtlb_priv[1][esel].ref.flags |= E500_TLB_BITMAP;
539 	vcpu_e500->g2h_tlb1_map[esel] |= (u64)1 << sesel;
540 	vcpu_e500->h2g_tlb1_rmap[sesel] = esel + 1;
541 	WARN_ON(!(ref->flags & E500_TLB_VALID));
542 
543 	return sesel;
544 }
545 
546 /* Caller must ensure that the specified guest TLB entry is safe to insert into
547  * the shadow TLB. */
548 /* For both one-one and one-to-many */
549 static int kvmppc_e500_tlb1_map(struct kvmppc_vcpu_e500 *vcpu_e500,
550 		u64 gvaddr, gfn_t gfn, struct kvm_book3e_206_tlb_entry *gtlbe,
551 		struct kvm_book3e_206_tlb_entry *stlbe, int esel)
552 {
553 	struct tlbe_ref *ref = &vcpu_e500->gtlb_priv[1][esel].ref;
554 	int sesel;
555 	int r;
556 
557 	r = kvmppc_e500_shadow_map(vcpu_e500, gvaddr, gfn, gtlbe, 1, stlbe,
558 				   ref);
559 	if (r)
560 		return r;
561 
562 	/* Use TLB0 when we can only map a page with 4k */
563 	if (get_tlb_tsize(stlbe) == BOOK3E_PAGESZ_4K) {
564 		vcpu_e500->gtlb_priv[1][esel].ref.flags |= E500_TLB_TLB0;
565 		write_stlbe(vcpu_e500, gtlbe, stlbe, 0, 0);
566 		return 0;
567 	}
568 
569 	/* Otherwise map into TLB1 */
570 	sesel = kvmppc_e500_tlb1_map_tlb1(vcpu_e500, ref, esel);
571 	write_stlbe(vcpu_e500, gtlbe, stlbe, 1, sesel);
572 
573 	return 0;
574 }
575 
576 void kvmppc_mmu_map(struct kvm_vcpu *vcpu, u64 eaddr, gpa_t gpaddr,
577 		    unsigned int index)
578 {
579 	struct kvmppc_vcpu_e500 *vcpu_e500 = to_e500(vcpu);
580 	struct tlbe_priv *priv;
581 	struct kvm_book3e_206_tlb_entry *gtlbe, stlbe;
582 	int tlbsel = tlbsel_of(index);
583 	int esel = esel_of(index);
584 
585 	gtlbe = get_entry(vcpu_e500, tlbsel, esel);
586 
587 	switch (tlbsel) {
588 	case 0:
589 		priv = &vcpu_e500->gtlb_priv[tlbsel][esel];
590 
591 		/* Triggers after clear_tlb_privs or on initial mapping */
592 		if (!(priv->ref.flags & E500_TLB_VALID)) {
593 			kvmppc_e500_tlb0_map(vcpu_e500, esel, &stlbe);
594 		} else {
595 			kvmppc_e500_setup_stlbe(vcpu, gtlbe, BOOK3E_PAGESZ_4K,
596 						&priv->ref, eaddr, &stlbe);
597 			write_stlbe(vcpu_e500, gtlbe, &stlbe, 0, 0);
598 		}
599 		break;
600 
601 	case 1: {
602 		gfn_t gfn = gpaddr >> PAGE_SHIFT;
603 		kvmppc_e500_tlb1_map(vcpu_e500, eaddr, gfn, gtlbe, &stlbe,
604 				     esel);
605 		break;
606 	}
607 
608 	default:
609 		BUG();
610 		break;
611 	}
612 }
613 
614 #ifdef CONFIG_KVM_BOOKE_HV
615 int kvmppc_load_last_inst(struct kvm_vcpu *vcpu, enum instruction_type type,
616 			  u32 *instr)
617 {
618 	gva_t geaddr;
619 	hpa_t addr;
620 	hfn_t pfn;
621 	hva_t eaddr;
622 	u32 mas1, mas2, mas3;
623 	u64 mas7_mas3;
624 	struct page *page;
625 	unsigned int addr_space, psize_shift;
626 	bool pr;
627 	unsigned long flags;
628 
629 	/* Search TLB for guest pc to get the real address */
630 	geaddr = kvmppc_get_pc(vcpu);
631 
632 	addr_space = (vcpu->arch.shared->msr & MSR_IS) >> MSR_IR_LG;
633 
634 	local_irq_save(flags);
635 	mtspr(SPRN_MAS6, (vcpu->arch.pid << MAS6_SPID_SHIFT) | addr_space);
636 	mtspr(SPRN_MAS5, MAS5_SGS | vcpu->kvm->arch.lpid);
637 	asm volatile("tlbsx 0, %[geaddr]\n" : :
638 		     [geaddr] "r" (geaddr));
639 	mtspr(SPRN_MAS5, 0);
640 	mtspr(SPRN_MAS8, 0);
641 	mas1 = mfspr(SPRN_MAS1);
642 	mas2 = mfspr(SPRN_MAS2);
643 	mas3 = mfspr(SPRN_MAS3);
644 #ifdef CONFIG_64BIT
645 	mas7_mas3 = mfspr(SPRN_MAS7_MAS3);
646 #else
647 	mas7_mas3 = ((u64)mfspr(SPRN_MAS7) << 32) | mas3;
648 #endif
649 	local_irq_restore(flags);
650 
651 	/*
652 	 * If the TLB entry for guest pc was evicted, return to the guest.
653 	 * There are high chances to find a valid TLB entry next time.
654 	 */
655 	if (!(mas1 & MAS1_VALID))
656 		return EMULATE_AGAIN;
657 
658 	/*
659 	 * Another thread may rewrite the TLB entry in parallel, don't
660 	 * execute from the address if the execute permission is not set
661 	 */
662 	pr = vcpu->arch.shared->msr & MSR_PR;
663 	if (unlikely((pr && !(mas3 & MAS3_UX)) ||
664 		     (!pr && !(mas3 & MAS3_SX)))) {
665 		pr_err_ratelimited(
666 			"%s: Instuction emulation from guest addres %08lx without execute permission\n",
667 			__func__, geaddr);
668 		return EMULATE_AGAIN;
669 	}
670 
671 	/*
672 	 * The real address will be mapped by a cacheable, memory coherent,
673 	 * write-back page. Check for mismatches when LRAT is used.
674 	 */
675 	if (has_feature(vcpu, VCPU_FTR_MMU_V2) &&
676 	    unlikely((mas2 & MAS2_I) || (mas2 & MAS2_W) || !(mas2 & MAS2_M))) {
677 		pr_err_ratelimited(
678 			"%s: Instuction emulation from guest addres %08lx mismatches storage attributes\n",
679 			__func__, geaddr);
680 		return EMULATE_AGAIN;
681 	}
682 
683 	/* Get pfn */
684 	psize_shift = MAS1_GET_TSIZE(mas1) + 10;
685 	addr = (mas7_mas3 & (~0ULL << psize_shift)) |
686 	       (geaddr & ((1ULL << psize_shift) - 1ULL));
687 	pfn = addr >> PAGE_SHIFT;
688 
689 	/* Guard against emulation from devices area */
690 	if (unlikely(!page_is_ram(pfn))) {
691 		pr_err_ratelimited("%s: Instruction emulation from non-RAM host addres %08llx is not supported\n",
692 			 __func__, addr);
693 		return EMULATE_AGAIN;
694 	}
695 
696 	/* Map a page and get guest's instruction */
697 	page = pfn_to_page(pfn);
698 	eaddr = (unsigned long)kmap_atomic(page);
699 	*instr = *(u32 *)(eaddr | (unsigned long)(addr & ~PAGE_MASK));
700 	kunmap_atomic((u32 *)eaddr);
701 
702 	return EMULATE_DONE;
703 }
704 #else
705 int kvmppc_load_last_inst(struct kvm_vcpu *vcpu, enum instruction_type type,
706 			  u32 *instr)
707 {
708 	return EMULATE_AGAIN;
709 }
710 #endif
711 
712 /************* MMU Notifiers *************/
713 
714 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
715 {
716 	trace_kvm_unmap_hva(hva);
717 
718 	/*
719 	 * Flush all shadow tlb entries everywhere. This is slow, but
720 	 * we are 100% sure that we catch the to be unmapped page
721 	 */
722 	kvm_flush_remote_tlbs(kvm);
723 
724 	return 0;
725 }
726 
727 int kvm_unmap_hva_range(struct kvm *kvm, unsigned long start, unsigned long end)
728 {
729 	/* kvm_unmap_hva flushes everything anyways */
730 	kvm_unmap_hva(kvm, start);
731 
732 	return 0;
733 }
734 
735 int kvm_age_hva(struct kvm *kvm, unsigned long hva)
736 {
737 	/* XXX could be more clever ;) */
738 	return 0;
739 }
740 
741 int kvm_test_age_hva(struct kvm *kvm, unsigned long hva)
742 {
743 	/* XXX could be more clever ;) */
744 	return 0;
745 }
746 
747 void kvm_set_spte_hva(struct kvm *kvm, unsigned long hva, pte_t pte)
748 {
749 	/* The page will get remapped properly on its next fault */
750 	kvm_unmap_hva(kvm, hva);
751 }
752 
753 /*****************************************/
754 
755 int e500_mmu_host_init(struct kvmppc_vcpu_e500 *vcpu_e500)
756 {
757 	host_tlb_params[0].entries = mfspr(SPRN_TLB0CFG) & TLBnCFG_N_ENTRY;
758 	host_tlb_params[1].entries = mfspr(SPRN_TLB1CFG) & TLBnCFG_N_ENTRY;
759 
760 	/*
761 	 * This should never happen on real e500 hardware, but is
762 	 * architecturally possible -- e.g. in some weird nested
763 	 * virtualization case.
764 	 */
765 	if (host_tlb_params[0].entries == 0 ||
766 	    host_tlb_params[1].entries == 0) {
767 		pr_err("%s: need to know host tlb size\n", __func__);
768 		return -ENODEV;
769 	}
770 
771 	host_tlb_params[0].ways = (mfspr(SPRN_TLB0CFG) & TLBnCFG_ASSOC) >>
772 				  TLBnCFG_ASSOC_SHIFT;
773 	host_tlb_params[1].ways = host_tlb_params[1].entries;
774 
775 	if (!is_power_of_2(host_tlb_params[0].entries) ||
776 	    !is_power_of_2(host_tlb_params[0].ways) ||
777 	    host_tlb_params[0].entries < host_tlb_params[0].ways ||
778 	    host_tlb_params[0].ways == 0) {
779 		pr_err("%s: bad tlb0 host config: %u entries %u ways\n",
780 		       __func__, host_tlb_params[0].entries,
781 		       host_tlb_params[0].ways);
782 		return -ENODEV;
783 	}
784 
785 	host_tlb_params[0].sets =
786 		host_tlb_params[0].entries / host_tlb_params[0].ways;
787 	host_tlb_params[1].sets = 1;
788 
789 	vcpu_e500->h2g_tlb1_rmap = kzalloc(sizeof(unsigned int) *
790 					   host_tlb_params[1].entries,
791 					   GFP_KERNEL);
792 	if (!vcpu_e500->h2g_tlb1_rmap)
793 		return -EINVAL;
794 
795 	return 0;
796 }
797 
798 void e500_mmu_host_uninit(struct kvmppc_vcpu_e500 *vcpu_e500)
799 {
800 	kfree(vcpu_e500->h2g_tlb1_rmap);
801 }
802