xref: /openbmc/linux/arch/powerpc/kvm/e500_mmu_host.c (revision b23bf21f)
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, int usermode)
49 {
50 	/* Mask off reserved bits. */
51 	mas3 &= MAS3_ATTRIB_MASK;
52 
53 #ifndef CONFIG_KVM_BOOKE_HV
54 	if (!usermode) {
55 		/* Guest is in supervisor mode,
56 		 * so we need to translate guest
57 		 * supervisor permissions into user permissions. */
58 		mas3 &= ~E500_TLB_USER_PERM_MASK;
59 		mas3 |= (mas3 & E500_TLB_SUPER_PERM_MASK) << 1;
60 	}
61 	mas3 |= E500_TLB_SUPER_PERM_MASK;
62 #endif
63 	return mas3;
64 }
65 
66 /*
67  * writing shadow tlb entry to host TLB
68  */
69 static inline void __write_host_tlbe(struct kvm_book3e_206_tlb_entry *stlbe,
70 				     uint32_t mas0,
71 				     uint32_t lpid)
72 {
73 	unsigned long flags;
74 
75 	local_irq_save(flags);
76 	mtspr(SPRN_MAS0, mas0);
77 	mtspr(SPRN_MAS1, stlbe->mas1);
78 	mtspr(SPRN_MAS2, (unsigned long)stlbe->mas2);
79 	mtspr(SPRN_MAS3, (u32)stlbe->mas7_3);
80 	mtspr(SPRN_MAS7, (u32)(stlbe->mas7_3 >> 32));
81 #ifdef CONFIG_KVM_BOOKE_HV
82 	mtspr(SPRN_MAS8, MAS8_TGS | get_thread_specific_lpid(lpid));
83 #endif
84 	asm volatile("isync; tlbwe" : : : "memory");
85 
86 #ifdef CONFIG_KVM_BOOKE_HV
87 	/* Must clear mas8 for other host tlbwe's */
88 	mtspr(SPRN_MAS8, 0);
89 	isync();
90 #endif
91 	local_irq_restore(flags);
92 
93 	trace_kvm_booke206_stlb_write(mas0, stlbe->mas8, stlbe->mas1,
94 	                              stlbe->mas2, stlbe->mas7_3);
95 }
96 
97 /*
98  * Acquire a mas0 with victim hint, as if we just took a TLB miss.
99  *
100  * We don't care about the address we're searching for, other than that it's
101  * in the right set and is not present in the TLB.  Using a zero PID and a
102  * userspace address means we don't have to set and then restore MAS5, or
103  * calculate a proper MAS6 value.
104  */
105 static u32 get_host_mas0(unsigned long eaddr)
106 {
107 	unsigned long flags;
108 	u32 mas0;
109 	u32 mas4;
110 
111 	local_irq_save(flags);
112 	mtspr(SPRN_MAS6, 0);
113 	mas4 = mfspr(SPRN_MAS4);
114 	mtspr(SPRN_MAS4, mas4 & ~MAS4_TLBSEL_MASK);
115 	asm volatile("tlbsx 0, %0" : : "b" (eaddr & ~CONFIG_PAGE_OFFSET));
116 	mas0 = mfspr(SPRN_MAS0);
117 	mtspr(SPRN_MAS4, mas4);
118 	local_irq_restore(flags);
119 
120 	return mas0;
121 }
122 
123 /* sesel is for tlb1 only */
124 static inline void write_host_tlbe(struct kvmppc_vcpu_e500 *vcpu_e500,
125 		int tlbsel, int sesel, struct kvm_book3e_206_tlb_entry *stlbe)
126 {
127 	u32 mas0;
128 
129 	if (tlbsel == 0) {
130 		mas0 = get_host_mas0(stlbe->mas2);
131 		__write_host_tlbe(stlbe, mas0, vcpu_e500->vcpu.kvm->arch.lpid);
132 	} else {
133 		__write_host_tlbe(stlbe,
134 				  MAS0_TLBSEL(1) |
135 				  MAS0_ESEL(to_htlb1_esel(sesel)),
136 				  vcpu_e500->vcpu.kvm->arch.lpid);
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 	kvm_pfn_t pfn;
165 
166 	pfn = (kvm_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), 0);
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 					 kvm_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 	kvm_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 
322 static inline int kvmppc_e500_shadow_map(struct kvmppc_vcpu_e500 *vcpu_e500,
323 	u64 gvaddr, gfn_t gfn, struct kvm_book3e_206_tlb_entry *gtlbe,
324 	int tlbsel, struct kvm_book3e_206_tlb_entry *stlbe,
325 	struct tlbe_ref *ref)
326 {
327 	struct kvm_memory_slot *slot;
328 	unsigned long pfn = 0; /* silence GCC warning */
329 	unsigned long hva;
330 	int pfnmap = 0;
331 	int tsize = BOOK3E_PAGESZ_4K;
332 	int ret = 0;
333 	unsigned long mmu_seq;
334 	struct kvm *kvm = vcpu_e500->vcpu.kvm;
335 	unsigned long tsize_pages = 0;
336 	pte_t *ptep;
337 	unsigned int wimg = 0;
338 	pgd_t *pgdir;
339 	unsigned long flags;
340 
341 	/* used to check for invalidations in progress */
342 	mmu_seq = kvm->mmu_notifier_seq;
343 	smp_rmb();
344 
345 	/*
346 	 * Translate guest physical to true physical, acquiring
347 	 * a page reference if it is normal, non-reserved memory.
348 	 *
349 	 * gfn_to_memslot() must succeed because otherwise we wouldn't
350 	 * have gotten this far.  Eventually we should just pass the slot
351 	 * pointer through from the first lookup.
352 	 */
353 	slot = gfn_to_memslot(vcpu_e500->vcpu.kvm, gfn);
354 	hva = gfn_to_hva_memslot(slot, gfn);
355 
356 	if (tlbsel == 1) {
357 		struct vm_area_struct *vma;
358 		down_read(&current->mm->mmap_sem);
359 
360 		vma = find_vma(current->mm, hva);
361 		if (vma && hva >= vma->vm_start &&
362 		    (vma->vm_flags & VM_PFNMAP)) {
363 			/*
364 			 * This VMA is a physically contiguous region (e.g.
365 			 * /dev/mem) that bypasses normal Linux page
366 			 * management.  Find the overlap between the
367 			 * vma and the memslot.
368 			 */
369 
370 			unsigned long start, end;
371 			unsigned long slot_start, slot_end;
372 
373 			pfnmap = 1;
374 
375 			start = vma->vm_pgoff;
376 			end = start +
377 			      vma_pages(vma);
378 
379 			pfn = start + ((hva - vma->vm_start) >> PAGE_SHIFT);
380 
381 			slot_start = pfn - (gfn - slot->base_gfn);
382 			slot_end = slot_start + slot->npages;
383 
384 			if (start < slot_start)
385 				start = slot_start;
386 			if (end > slot_end)
387 				end = slot_end;
388 
389 			tsize = (gtlbe->mas1 & MAS1_TSIZE_MASK) >>
390 				MAS1_TSIZE_SHIFT;
391 
392 			/*
393 			 * e500 doesn't implement the lowest tsize bit,
394 			 * or 1K pages.
395 			 */
396 			tsize = max(BOOK3E_PAGESZ_4K, tsize & ~1);
397 
398 			/*
399 			 * Now find the largest tsize (up to what the guest
400 			 * requested) that will cover gfn, stay within the
401 			 * range, and for which gfn and pfn are mutually
402 			 * aligned.
403 			 */
404 
405 			for (; tsize > BOOK3E_PAGESZ_4K; tsize -= 2) {
406 				unsigned long gfn_start, gfn_end;
407 				tsize_pages = 1UL << (tsize - 2);
408 
409 				gfn_start = gfn & ~(tsize_pages - 1);
410 				gfn_end = gfn_start + tsize_pages;
411 
412 				if (gfn_start + pfn - gfn < start)
413 					continue;
414 				if (gfn_end + pfn - gfn > end)
415 					continue;
416 				if ((gfn & (tsize_pages - 1)) !=
417 				    (pfn & (tsize_pages - 1)))
418 					continue;
419 
420 				gvaddr &= ~((tsize_pages << PAGE_SHIFT) - 1);
421 				pfn &= ~(tsize_pages - 1);
422 				break;
423 			}
424 		} else if (vma && hva >= vma->vm_start &&
425 			   (vma->vm_flags & VM_HUGETLB)) {
426 			unsigned long psize = vma_kernel_pagesize(vma);
427 
428 			tsize = (gtlbe->mas1 & MAS1_TSIZE_MASK) >>
429 				MAS1_TSIZE_SHIFT;
430 
431 			/*
432 			 * Take the largest page size that satisfies both host
433 			 * and guest mapping
434 			 */
435 			tsize = min(__ilog2(psize) - 10, tsize);
436 
437 			/*
438 			 * e500 doesn't implement the lowest tsize bit,
439 			 * or 1K pages.
440 			 */
441 			tsize = max(BOOK3E_PAGESZ_4K, tsize & ~1);
442 		}
443 
444 		up_read(&current->mm->mmap_sem);
445 	}
446 
447 	if (likely(!pfnmap)) {
448 		tsize_pages = 1UL << (tsize + 10 - PAGE_SHIFT);
449 		pfn = gfn_to_pfn_memslot(slot, gfn);
450 		if (is_error_noslot_pfn(pfn)) {
451 			if (printk_ratelimit())
452 				pr_err("%s: real page not found for gfn %lx\n",
453 				       __func__, (long)gfn);
454 			return -EINVAL;
455 		}
456 
457 		/* Align guest and physical address to page map boundaries */
458 		pfn &= ~(tsize_pages - 1);
459 		gvaddr &= ~((tsize_pages << PAGE_SHIFT) - 1);
460 	}
461 
462 	spin_lock(&kvm->mmu_lock);
463 	if (mmu_notifier_retry(kvm, mmu_seq)) {
464 		ret = -EAGAIN;
465 		goto out;
466 	}
467 
468 
469 	pgdir = vcpu_e500->vcpu.arch.pgdir;
470 	/*
471 	 * We are just looking at the wimg bits, so we don't
472 	 * care much about the trans splitting bit.
473 	 * We are holding kvm->mmu_lock so a notifier invalidate
474 	 * can't run hence pfn won't change.
475 	 */
476 	local_irq_save(flags);
477 	ptep = find_linux_pte(pgdir, hva, NULL, NULL);
478 	if (ptep) {
479 		pte_t pte = READ_ONCE(*ptep);
480 
481 		if (pte_present(pte)) {
482 			wimg = (pte_val(pte) >> PTE_WIMGE_SHIFT) &
483 				MAS2_WIMGE_MASK;
484 			local_irq_restore(flags);
485 		} else {
486 			local_irq_restore(flags);
487 			pr_err_ratelimited("%s: pte not present: gfn %lx,pfn %lx\n",
488 					   __func__, (long)gfn, pfn);
489 			ret = -EINVAL;
490 			goto out;
491 		}
492 	}
493 	kvmppc_e500_ref_setup(ref, gtlbe, pfn, wimg);
494 
495 	kvmppc_e500_setup_stlbe(&vcpu_e500->vcpu, gtlbe, tsize,
496 				ref, gvaddr, stlbe);
497 
498 	/* Clear i-cache for new pages */
499 	kvmppc_mmu_flush_icache(pfn);
500 
501 out:
502 	spin_unlock(&kvm->mmu_lock);
503 
504 	/* Drop refcount on page, so that mmu notifiers can clear it */
505 	kvm_release_pfn_clean(pfn);
506 
507 	return ret;
508 }
509 
510 /* XXX only map the one-one case, for now use TLB0 */
511 static int kvmppc_e500_tlb0_map(struct kvmppc_vcpu_e500 *vcpu_e500, int esel,
512 				struct kvm_book3e_206_tlb_entry *stlbe)
513 {
514 	struct kvm_book3e_206_tlb_entry *gtlbe;
515 	struct tlbe_ref *ref;
516 	int stlbsel = 0;
517 	int sesel = 0;
518 	int r;
519 
520 	gtlbe = get_entry(vcpu_e500, 0, esel);
521 	ref = &vcpu_e500->gtlb_priv[0][esel].ref;
522 
523 	r = kvmppc_e500_shadow_map(vcpu_e500, get_tlb_eaddr(gtlbe),
524 			get_tlb_raddr(gtlbe) >> PAGE_SHIFT,
525 			gtlbe, 0, stlbe, ref);
526 	if (r)
527 		return r;
528 
529 	write_stlbe(vcpu_e500, gtlbe, stlbe, stlbsel, sesel);
530 
531 	return 0;
532 }
533 
534 static int kvmppc_e500_tlb1_map_tlb1(struct kvmppc_vcpu_e500 *vcpu_e500,
535 				     struct tlbe_ref *ref,
536 				     int esel)
537 {
538 	unsigned int sesel = vcpu_e500->host_tlb1_nv++;
539 
540 	if (unlikely(vcpu_e500->host_tlb1_nv >= tlb1_max_shadow_size()))
541 		vcpu_e500->host_tlb1_nv = 0;
542 
543 	if (vcpu_e500->h2g_tlb1_rmap[sesel]) {
544 		unsigned int idx = vcpu_e500->h2g_tlb1_rmap[sesel] - 1;
545 		vcpu_e500->g2h_tlb1_map[idx] &= ~(1ULL << sesel);
546 	}
547 
548 	vcpu_e500->gtlb_priv[1][esel].ref.flags |= E500_TLB_BITMAP;
549 	vcpu_e500->g2h_tlb1_map[esel] |= (u64)1 << sesel;
550 	vcpu_e500->h2g_tlb1_rmap[sesel] = esel + 1;
551 	WARN_ON(!(ref->flags & E500_TLB_VALID));
552 
553 	return sesel;
554 }
555 
556 /* Caller must ensure that the specified guest TLB entry is safe to insert into
557  * the shadow TLB. */
558 /* For both one-one and one-to-many */
559 static int kvmppc_e500_tlb1_map(struct kvmppc_vcpu_e500 *vcpu_e500,
560 		u64 gvaddr, gfn_t gfn, struct kvm_book3e_206_tlb_entry *gtlbe,
561 		struct kvm_book3e_206_tlb_entry *stlbe, int esel)
562 {
563 	struct tlbe_ref *ref = &vcpu_e500->gtlb_priv[1][esel].ref;
564 	int sesel;
565 	int r;
566 
567 	r = kvmppc_e500_shadow_map(vcpu_e500, gvaddr, gfn, gtlbe, 1, stlbe,
568 				   ref);
569 	if (r)
570 		return r;
571 
572 	/* Use TLB0 when we can only map a page with 4k */
573 	if (get_tlb_tsize(stlbe) == BOOK3E_PAGESZ_4K) {
574 		vcpu_e500->gtlb_priv[1][esel].ref.flags |= E500_TLB_TLB0;
575 		write_stlbe(vcpu_e500, gtlbe, stlbe, 0, 0);
576 		return 0;
577 	}
578 
579 	/* Otherwise map into TLB1 */
580 	sesel = kvmppc_e500_tlb1_map_tlb1(vcpu_e500, ref, esel);
581 	write_stlbe(vcpu_e500, gtlbe, stlbe, 1, sesel);
582 
583 	return 0;
584 }
585 
586 void kvmppc_mmu_map(struct kvm_vcpu *vcpu, u64 eaddr, gpa_t gpaddr,
587 		    unsigned int index)
588 {
589 	struct kvmppc_vcpu_e500 *vcpu_e500 = to_e500(vcpu);
590 	struct tlbe_priv *priv;
591 	struct kvm_book3e_206_tlb_entry *gtlbe, stlbe;
592 	int tlbsel = tlbsel_of(index);
593 	int esel = esel_of(index);
594 
595 	gtlbe = get_entry(vcpu_e500, tlbsel, esel);
596 
597 	switch (tlbsel) {
598 	case 0:
599 		priv = &vcpu_e500->gtlb_priv[tlbsel][esel];
600 
601 		/* Triggers after clear_tlb_privs or on initial mapping */
602 		if (!(priv->ref.flags & E500_TLB_VALID)) {
603 			kvmppc_e500_tlb0_map(vcpu_e500, esel, &stlbe);
604 		} else {
605 			kvmppc_e500_setup_stlbe(vcpu, gtlbe, BOOK3E_PAGESZ_4K,
606 						&priv->ref, eaddr, &stlbe);
607 			write_stlbe(vcpu_e500, gtlbe, &stlbe, 0, 0);
608 		}
609 		break;
610 
611 	case 1: {
612 		gfn_t gfn = gpaddr >> PAGE_SHIFT;
613 		kvmppc_e500_tlb1_map(vcpu_e500, eaddr, gfn, gtlbe, &stlbe,
614 				     esel);
615 		break;
616 	}
617 
618 	default:
619 		BUG();
620 		break;
621 	}
622 }
623 
624 #ifdef CONFIG_KVM_BOOKE_HV
625 int kvmppc_load_last_inst(struct kvm_vcpu *vcpu,
626 		enum instruction_fetch_type type, u32 *instr)
627 {
628 	gva_t geaddr;
629 	hpa_t addr;
630 	hfn_t pfn;
631 	hva_t eaddr;
632 	u32 mas1, mas2, mas3;
633 	u64 mas7_mas3;
634 	struct page *page;
635 	unsigned int addr_space, psize_shift;
636 	bool pr;
637 	unsigned long flags;
638 
639 	/* Search TLB for guest pc to get the real address */
640 	geaddr = kvmppc_get_pc(vcpu);
641 
642 	addr_space = (vcpu->arch.shared->msr & MSR_IS) >> MSR_IR_LG;
643 
644 	local_irq_save(flags);
645 	mtspr(SPRN_MAS6, (vcpu->arch.pid << MAS6_SPID_SHIFT) | addr_space);
646 	mtspr(SPRN_MAS5, MAS5_SGS | get_lpid(vcpu));
647 	asm volatile("tlbsx 0, %[geaddr]\n" : :
648 		     [geaddr] "r" (geaddr));
649 	mtspr(SPRN_MAS5, 0);
650 	mtspr(SPRN_MAS8, 0);
651 	mas1 = mfspr(SPRN_MAS1);
652 	mas2 = mfspr(SPRN_MAS2);
653 	mas3 = mfspr(SPRN_MAS3);
654 #ifdef CONFIG_64BIT
655 	mas7_mas3 = mfspr(SPRN_MAS7_MAS3);
656 #else
657 	mas7_mas3 = ((u64)mfspr(SPRN_MAS7) << 32) | mas3;
658 #endif
659 	local_irq_restore(flags);
660 
661 	/*
662 	 * If the TLB entry for guest pc was evicted, return to the guest.
663 	 * There are high chances to find a valid TLB entry next time.
664 	 */
665 	if (!(mas1 & MAS1_VALID))
666 		return EMULATE_AGAIN;
667 
668 	/*
669 	 * Another thread may rewrite the TLB entry in parallel, don't
670 	 * execute from the address if the execute permission is not set
671 	 */
672 	pr = vcpu->arch.shared->msr & MSR_PR;
673 	if (unlikely((pr && !(mas3 & MAS3_UX)) ||
674 		     (!pr && !(mas3 & MAS3_SX)))) {
675 		pr_err_ratelimited(
676 			"%s: Instruction emulation from guest address %08lx without execute permission\n",
677 			__func__, geaddr);
678 		return EMULATE_AGAIN;
679 	}
680 
681 	/*
682 	 * The real address will be mapped by a cacheable, memory coherent,
683 	 * write-back page. Check for mismatches when LRAT is used.
684 	 */
685 	if (has_feature(vcpu, VCPU_FTR_MMU_V2) &&
686 	    unlikely((mas2 & MAS2_I) || (mas2 & MAS2_W) || !(mas2 & MAS2_M))) {
687 		pr_err_ratelimited(
688 			"%s: Instruction emulation from guest address %08lx mismatches storage attributes\n",
689 			__func__, geaddr);
690 		return EMULATE_AGAIN;
691 	}
692 
693 	/* Get pfn */
694 	psize_shift = MAS1_GET_TSIZE(mas1) + 10;
695 	addr = (mas7_mas3 & (~0ULL << psize_shift)) |
696 	       (geaddr & ((1ULL << psize_shift) - 1ULL));
697 	pfn = addr >> PAGE_SHIFT;
698 
699 	/* Guard against emulation from devices area */
700 	if (unlikely(!page_is_ram(pfn))) {
701 		pr_err_ratelimited("%s: Instruction emulation from non-RAM host address %08llx is not supported\n",
702 			 __func__, addr);
703 		return EMULATE_AGAIN;
704 	}
705 
706 	/* Map a page and get guest's instruction */
707 	page = pfn_to_page(pfn);
708 	eaddr = (unsigned long)kmap_atomic(page);
709 	*instr = *(u32 *)(eaddr | (unsigned long)(addr & ~PAGE_MASK));
710 	kunmap_atomic((u32 *)eaddr);
711 
712 	return EMULATE_DONE;
713 }
714 #else
715 int kvmppc_load_last_inst(struct kvm_vcpu *vcpu,
716 		enum instruction_fetch_type type, u32 *instr)
717 {
718 	return EMULATE_AGAIN;
719 }
720 #endif
721 
722 /************* MMU Notifiers *************/
723 
724 static int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
725 {
726 	trace_kvm_unmap_hva(hva);
727 
728 	/*
729 	 * Flush all shadow tlb entries everywhere. This is slow, but
730 	 * we are 100% sure that we catch the to be unmapped page
731 	 */
732 	kvm_flush_remote_tlbs(kvm);
733 
734 	return 0;
735 }
736 
737 int kvm_unmap_hva_range(struct kvm *kvm, unsigned long start, unsigned long end)
738 {
739 	/* kvm_unmap_hva flushes everything anyways */
740 	kvm_unmap_hva(kvm, start);
741 
742 	return 0;
743 }
744 
745 int kvm_age_hva(struct kvm *kvm, unsigned long start, unsigned long end)
746 {
747 	/* XXX could be more clever ;) */
748 	return 0;
749 }
750 
751 int kvm_test_age_hva(struct kvm *kvm, unsigned long hva)
752 {
753 	/* XXX could be more clever ;) */
754 	return 0;
755 }
756 
757 int kvm_set_spte_hva(struct kvm *kvm, unsigned long hva, pte_t pte)
758 {
759 	/* The page will get remapped properly on its next fault */
760 	kvm_unmap_hva(kvm, hva);
761 	return 0;
762 }
763 
764 /*****************************************/
765 
766 int e500_mmu_host_init(struct kvmppc_vcpu_e500 *vcpu_e500)
767 {
768 	host_tlb_params[0].entries = mfspr(SPRN_TLB0CFG) & TLBnCFG_N_ENTRY;
769 	host_tlb_params[1].entries = mfspr(SPRN_TLB1CFG) & TLBnCFG_N_ENTRY;
770 
771 	/*
772 	 * This should never happen on real e500 hardware, but is
773 	 * architecturally possible -- e.g. in some weird nested
774 	 * virtualization case.
775 	 */
776 	if (host_tlb_params[0].entries == 0 ||
777 	    host_tlb_params[1].entries == 0) {
778 		pr_err("%s: need to know host tlb size\n", __func__);
779 		return -ENODEV;
780 	}
781 
782 	host_tlb_params[0].ways = (mfspr(SPRN_TLB0CFG) & TLBnCFG_ASSOC) >>
783 				  TLBnCFG_ASSOC_SHIFT;
784 	host_tlb_params[1].ways = host_tlb_params[1].entries;
785 
786 	if (!is_power_of_2(host_tlb_params[0].entries) ||
787 	    !is_power_of_2(host_tlb_params[0].ways) ||
788 	    host_tlb_params[0].entries < host_tlb_params[0].ways ||
789 	    host_tlb_params[0].ways == 0) {
790 		pr_err("%s: bad tlb0 host config: %u entries %u ways\n",
791 		       __func__, host_tlb_params[0].entries,
792 		       host_tlb_params[0].ways);
793 		return -ENODEV;
794 	}
795 
796 	host_tlb_params[0].sets =
797 		host_tlb_params[0].entries / host_tlb_params[0].ways;
798 	host_tlb_params[1].sets = 1;
799 	vcpu_e500->h2g_tlb1_rmap = kcalloc(host_tlb_params[1].entries,
800 					   sizeof(*vcpu_e500->h2g_tlb1_rmap),
801 					   GFP_KERNEL);
802 	if (!vcpu_e500->h2g_tlb1_rmap)
803 		return -EINVAL;
804 
805 	return 0;
806 }
807 
808 void e500_mmu_host_uninit(struct kvmppc_vcpu_e500 *vcpu_e500)
809 {
810 	kfree(vcpu_e500->h2g_tlb1_rmap);
811 }
812