xref: /linux/arch/powerpc/kvm/book3s_64_mmu_radix.c (revision 4b660dbd9ee2059850fd30e0df420ca7a38a1856)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  *
4  * Copyright 2016 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
5  */
6 
7 #include <linux/types.h>
8 #include <linux/string.h>
9 #include <linux/kvm.h>
10 #include <linux/kvm_host.h>
11 #include <linux/anon_inodes.h>
12 #include <linux/file.h>
13 #include <linux/debugfs.h>
14 #include <linux/pgtable.h>
15 
16 #include <asm/kvm_ppc.h>
17 #include <asm/kvm_book3s.h>
18 #include "book3s_hv.h"
19 #include <asm/page.h>
20 #include <asm/mmu.h>
21 #include <asm/pgalloc.h>
22 #include <asm/pte-walk.h>
23 #include <asm/ultravisor.h>
24 #include <asm/kvm_book3s_uvmem.h>
25 #include <asm/plpar_wrappers.h>
26 #include <asm/firmware.h>
27 
28 /*
29  * Supported radix tree geometry.
30  * Like p9, we support either 5 or 9 bits at the first (lowest) level,
31  * for a page size of 64k or 4k.
32  */
33 static int p9_supported_radix_bits[4] = { 5, 9, 9, 13 };
34 
35 unsigned long __kvmhv_copy_tofrom_guest_radix(int lpid, int pid,
36 					      gva_t eaddr, void *to, void *from,
37 					      unsigned long n)
38 {
39 	int old_pid, old_lpid;
40 	unsigned long quadrant, ret = n;
41 	bool is_load = !!to;
42 
43 	if (kvmhv_is_nestedv2())
44 		return H_UNSUPPORTED;
45 
46 	/* Can't access quadrants 1 or 2 in non-HV mode, call the HV to do it */
47 	if (kvmhv_on_pseries())
48 		return plpar_hcall_norets(H_COPY_TOFROM_GUEST, lpid, pid, eaddr,
49 					  (to != NULL) ? __pa(to): 0,
50 					  (from != NULL) ? __pa(from): 0, n);
51 
52 	if (eaddr & (0xFFFUL << 52))
53 		return ret;
54 
55 	quadrant = 1;
56 	if (!pid)
57 		quadrant = 2;
58 	if (is_load)
59 		from = (void *) (eaddr | (quadrant << 62));
60 	else
61 		to = (void *) (eaddr | (quadrant << 62));
62 
63 	preempt_disable();
64 
65 	asm volatile("hwsync" ::: "memory");
66 	isync();
67 	/* switch the lpid first to avoid running host with unallocated pid */
68 	old_lpid = mfspr(SPRN_LPID);
69 	if (old_lpid != lpid)
70 		mtspr(SPRN_LPID, lpid);
71 	if (quadrant == 1) {
72 		old_pid = mfspr(SPRN_PID);
73 		if (old_pid != pid)
74 			mtspr(SPRN_PID, pid);
75 	}
76 	isync();
77 
78 	pagefault_disable();
79 	if (is_load)
80 		ret = __copy_from_user_inatomic(to, (const void __user *)from, n);
81 	else
82 		ret = __copy_to_user_inatomic((void __user *)to, from, n);
83 	pagefault_enable();
84 
85 	asm volatile("hwsync" ::: "memory");
86 	isync();
87 	/* switch the pid first to avoid running host with unallocated pid */
88 	if (quadrant == 1 && pid != old_pid)
89 		mtspr(SPRN_PID, old_pid);
90 	if (lpid != old_lpid)
91 		mtspr(SPRN_LPID, old_lpid);
92 	isync();
93 
94 	preempt_enable();
95 
96 	return ret;
97 }
98 
99 static long kvmhv_copy_tofrom_guest_radix(struct kvm_vcpu *vcpu, gva_t eaddr,
100 					  void *to, void *from, unsigned long n)
101 {
102 	int lpid = vcpu->kvm->arch.lpid;
103 	int pid;
104 
105 	/* This would cause a data segment intr so don't allow the access */
106 	if (eaddr & (0x3FFUL << 52))
107 		return -EINVAL;
108 
109 	/* Should we be using the nested lpid */
110 	if (vcpu->arch.nested)
111 		lpid = vcpu->arch.nested->shadow_lpid;
112 
113 	/* If accessing quadrant 3 then pid is expected to be 0 */
114 	if (((eaddr >> 62) & 0x3) == 0x3)
115 		pid = 0;
116 	else
117 		pid = kvmppc_get_pid(vcpu);
118 
119 	eaddr &= ~(0xFFFUL << 52);
120 
121 	return __kvmhv_copy_tofrom_guest_radix(lpid, pid, eaddr, to, from, n);
122 }
123 
124 long kvmhv_copy_from_guest_radix(struct kvm_vcpu *vcpu, gva_t eaddr, void *to,
125 				 unsigned long n)
126 {
127 	long ret;
128 
129 	ret = kvmhv_copy_tofrom_guest_radix(vcpu, eaddr, to, NULL, n);
130 	if (ret > 0)
131 		memset(to + (n - ret), 0, ret);
132 
133 	return ret;
134 }
135 
136 long kvmhv_copy_to_guest_radix(struct kvm_vcpu *vcpu, gva_t eaddr, void *from,
137 			       unsigned long n)
138 {
139 	return kvmhv_copy_tofrom_guest_radix(vcpu, eaddr, NULL, from, n);
140 }
141 
142 int kvmppc_mmu_walk_radix_tree(struct kvm_vcpu *vcpu, gva_t eaddr,
143 			       struct kvmppc_pte *gpte, u64 root,
144 			       u64 *pte_ret_p)
145 {
146 	struct kvm *kvm = vcpu->kvm;
147 	int ret, level, ps;
148 	unsigned long rts, bits, offset, index;
149 	u64 pte, base, gpa;
150 	__be64 rpte;
151 
152 	rts = ((root & RTS1_MASK) >> (RTS1_SHIFT - 3)) |
153 		((root & RTS2_MASK) >> RTS2_SHIFT);
154 	bits = root & RPDS_MASK;
155 	base = root & RPDB_MASK;
156 
157 	offset = rts + 31;
158 
159 	/* Current implementations only support 52-bit space */
160 	if (offset != 52)
161 		return -EINVAL;
162 
163 	/* Walk each level of the radix tree */
164 	for (level = 3; level >= 0; --level) {
165 		u64 addr;
166 		/* Check a valid size */
167 		if (level && bits != p9_supported_radix_bits[level])
168 			return -EINVAL;
169 		if (level == 0 && !(bits == 5 || bits == 9))
170 			return -EINVAL;
171 		offset -= bits;
172 		index = (eaddr >> offset) & ((1UL << bits) - 1);
173 		/* Check that low bits of page table base are zero */
174 		if (base & ((1UL << (bits + 3)) - 1))
175 			return -EINVAL;
176 		/* Read the entry from guest memory */
177 		addr = base + (index * sizeof(rpte));
178 
179 		kvm_vcpu_srcu_read_lock(vcpu);
180 		ret = kvm_read_guest(kvm, addr, &rpte, sizeof(rpte));
181 		kvm_vcpu_srcu_read_unlock(vcpu);
182 		if (ret) {
183 			if (pte_ret_p)
184 				*pte_ret_p = addr;
185 			return ret;
186 		}
187 		pte = __be64_to_cpu(rpte);
188 		if (!(pte & _PAGE_PRESENT))
189 			return -ENOENT;
190 		/* Check if a leaf entry */
191 		if (pte & _PAGE_PTE)
192 			break;
193 		/* Get ready to walk the next level */
194 		base = pte & RPDB_MASK;
195 		bits = pte & RPDS_MASK;
196 	}
197 
198 	/* Need a leaf at lowest level; 512GB pages not supported */
199 	if (level < 0 || level == 3)
200 		return -EINVAL;
201 
202 	/* We found a valid leaf PTE */
203 	/* Offset is now log base 2 of the page size */
204 	gpa = pte & 0x01fffffffffff000ul;
205 	if (gpa & ((1ul << offset) - 1))
206 		return -EINVAL;
207 	gpa |= eaddr & ((1ul << offset) - 1);
208 	for (ps = MMU_PAGE_4K; ps < MMU_PAGE_COUNT; ++ps)
209 		if (offset == mmu_psize_defs[ps].shift)
210 			break;
211 	gpte->page_size = ps;
212 	gpte->page_shift = offset;
213 
214 	gpte->eaddr = eaddr;
215 	gpte->raddr = gpa;
216 
217 	/* Work out permissions */
218 	gpte->may_read = !!(pte & _PAGE_READ);
219 	gpte->may_write = !!(pte & _PAGE_WRITE);
220 	gpte->may_execute = !!(pte & _PAGE_EXEC);
221 
222 	gpte->rc = pte & (_PAGE_ACCESSED | _PAGE_DIRTY);
223 
224 	if (pte_ret_p)
225 		*pte_ret_p = pte;
226 
227 	return 0;
228 }
229 
230 /*
231  * Used to walk a partition or process table radix tree in guest memory
232  * Note: We exploit the fact that a partition table and a process
233  * table have the same layout, a partition-scoped page table and a
234  * process-scoped page table have the same layout, and the 2nd
235  * doubleword of a partition table entry has the same layout as
236  * the PTCR register.
237  */
238 int kvmppc_mmu_radix_translate_table(struct kvm_vcpu *vcpu, gva_t eaddr,
239 				     struct kvmppc_pte *gpte, u64 table,
240 				     int table_index, u64 *pte_ret_p)
241 {
242 	struct kvm *kvm = vcpu->kvm;
243 	int ret;
244 	unsigned long size, ptbl, root;
245 	struct prtb_entry entry;
246 
247 	if ((table & PRTS_MASK) > 24)
248 		return -EINVAL;
249 	size = 1ul << ((table & PRTS_MASK) + 12);
250 
251 	/* Is the table big enough to contain this entry? */
252 	if ((table_index * sizeof(entry)) >= size)
253 		return -EINVAL;
254 
255 	/* Read the table to find the root of the radix tree */
256 	ptbl = (table & PRTB_MASK) + (table_index * sizeof(entry));
257 	kvm_vcpu_srcu_read_lock(vcpu);
258 	ret = kvm_read_guest(kvm, ptbl, &entry, sizeof(entry));
259 	kvm_vcpu_srcu_read_unlock(vcpu);
260 	if (ret)
261 		return ret;
262 
263 	/* Root is stored in the first double word */
264 	root = be64_to_cpu(entry.prtb0);
265 
266 	return kvmppc_mmu_walk_radix_tree(vcpu, eaddr, gpte, root, pte_ret_p);
267 }
268 
269 int kvmppc_mmu_radix_xlate(struct kvm_vcpu *vcpu, gva_t eaddr,
270 			   struct kvmppc_pte *gpte, bool data, bool iswrite)
271 {
272 	u32 pid;
273 	u64 pte;
274 	int ret;
275 
276 	/* Work out effective PID */
277 	switch (eaddr >> 62) {
278 	case 0:
279 		pid = kvmppc_get_pid(vcpu);
280 		break;
281 	case 3:
282 		pid = 0;
283 		break;
284 	default:
285 		return -EINVAL;
286 	}
287 
288 	ret = kvmppc_mmu_radix_translate_table(vcpu, eaddr, gpte,
289 				vcpu->kvm->arch.process_table, pid, &pte);
290 	if (ret)
291 		return ret;
292 
293 	/* Check privilege (applies only to process scoped translations) */
294 	if (kvmppc_get_msr(vcpu) & MSR_PR) {
295 		if (pte & _PAGE_PRIVILEGED) {
296 			gpte->may_read = 0;
297 			gpte->may_write = 0;
298 			gpte->may_execute = 0;
299 		}
300 	} else {
301 		if (!(pte & _PAGE_PRIVILEGED)) {
302 			/* Check AMR/IAMR to see if strict mode is in force */
303 			if (kvmppc_get_amr_hv(vcpu) & (1ul << 62))
304 				gpte->may_read = 0;
305 			if (kvmppc_get_amr_hv(vcpu) & (1ul << 63))
306 				gpte->may_write = 0;
307 			if (vcpu->arch.iamr & (1ul << 62))
308 				gpte->may_execute = 0;
309 		}
310 	}
311 
312 	return 0;
313 }
314 
315 void kvmppc_radix_tlbie_page(struct kvm *kvm, unsigned long addr,
316 			     unsigned int pshift, u64 lpid)
317 {
318 	unsigned long psize = PAGE_SIZE;
319 	int psi;
320 	long rc;
321 	unsigned long rb;
322 
323 	if (pshift)
324 		psize = 1UL << pshift;
325 	else
326 		pshift = PAGE_SHIFT;
327 
328 	addr &= ~(psize - 1);
329 
330 	if (!kvmhv_on_pseries()) {
331 		radix__flush_tlb_lpid_page(lpid, addr, psize);
332 		return;
333 	}
334 
335 	psi = shift_to_mmu_psize(pshift);
336 
337 	if (!firmware_has_feature(FW_FEATURE_RPT_INVALIDATE)) {
338 		rb = addr | (mmu_get_ap(psi) << PPC_BITLSHIFT(58));
339 		rc = plpar_hcall_norets(H_TLB_INVALIDATE, H_TLBIE_P1_ENC(0, 0, 1),
340 					lpid, rb);
341 	} else {
342 		rc = pseries_rpt_invalidate(lpid, H_RPTI_TARGET_CMMU,
343 					    H_RPTI_TYPE_NESTED |
344 					    H_RPTI_TYPE_TLB,
345 					    psize_to_rpti_pgsize(psi),
346 					    addr, addr + psize);
347 	}
348 
349 	if (rc)
350 		pr_err("KVM: TLB page invalidation hcall failed, rc=%ld\n", rc);
351 }
352 
353 static void kvmppc_radix_flush_pwc(struct kvm *kvm, u64 lpid)
354 {
355 	long rc;
356 
357 	if (!kvmhv_on_pseries()) {
358 		radix__flush_pwc_lpid(lpid);
359 		return;
360 	}
361 
362 	if (!firmware_has_feature(FW_FEATURE_RPT_INVALIDATE))
363 		rc = plpar_hcall_norets(H_TLB_INVALIDATE, H_TLBIE_P1_ENC(1, 0, 1),
364 					lpid, TLBIEL_INVAL_SET_LPID);
365 	else
366 		rc = pseries_rpt_invalidate(lpid, H_RPTI_TARGET_CMMU,
367 					    H_RPTI_TYPE_NESTED |
368 					    H_RPTI_TYPE_PWC, H_RPTI_PAGE_ALL,
369 					    0, -1UL);
370 	if (rc)
371 		pr_err("KVM: TLB PWC invalidation hcall failed, rc=%ld\n", rc);
372 }
373 
374 static unsigned long kvmppc_radix_update_pte(struct kvm *kvm, pte_t *ptep,
375 				      unsigned long clr, unsigned long set,
376 				      unsigned long addr, unsigned int shift)
377 {
378 	return __radix_pte_update(ptep, clr, set);
379 }
380 
381 static void kvmppc_radix_set_pte_at(struct kvm *kvm, unsigned long addr,
382 			     pte_t *ptep, pte_t pte)
383 {
384 	radix__set_pte_at(kvm->mm, addr, ptep, pte, 0);
385 }
386 
387 static struct kmem_cache *kvm_pte_cache;
388 static struct kmem_cache *kvm_pmd_cache;
389 
390 static pte_t *kvmppc_pte_alloc(void)
391 {
392 	pte_t *pte;
393 
394 	pte = kmem_cache_alloc(kvm_pte_cache, GFP_KERNEL);
395 	/* pmd_populate() will only reference _pa(pte). */
396 	kmemleak_ignore(pte);
397 
398 	return pte;
399 }
400 
401 static void kvmppc_pte_free(pte_t *ptep)
402 {
403 	kmem_cache_free(kvm_pte_cache, ptep);
404 }
405 
406 static pmd_t *kvmppc_pmd_alloc(void)
407 {
408 	pmd_t *pmd;
409 
410 	pmd = kmem_cache_alloc(kvm_pmd_cache, GFP_KERNEL);
411 	/* pud_populate() will only reference _pa(pmd). */
412 	kmemleak_ignore(pmd);
413 
414 	return pmd;
415 }
416 
417 static void kvmppc_pmd_free(pmd_t *pmdp)
418 {
419 	kmem_cache_free(kvm_pmd_cache, pmdp);
420 }
421 
422 /* Called with kvm->mmu_lock held */
423 void kvmppc_unmap_pte(struct kvm *kvm, pte_t *pte, unsigned long gpa,
424 		      unsigned int shift,
425 		      const struct kvm_memory_slot *memslot,
426 		      u64 lpid)
427 
428 {
429 	unsigned long old;
430 	unsigned long gfn = gpa >> PAGE_SHIFT;
431 	unsigned long page_size = PAGE_SIZE;
432 	unsigned long hpa;
433 
434 	old = kvmppc_radix_update_pte(kvm, pte, ~0UL, 0, gpa, shift);
435 	kvmppc_radix_tlbie_page(kvm, gpa, shift, lpid);
436 
437 	/* The following only applies to L1 entries */
438 	if (lpid != kvm->arch.lpid)
439 		return;
440 
441 	if (!memslot) {
442 		memslot = gfn_to_memslot(kvm, gfn);
443 		if (!memslot)
444 			return;
445 	}
446 	if (shift) { /* 1GB or 2MB page */
447 		page_size = 1ul << shift;
448 		if (shift == PMD_SHIFT)
449 			kvm->stat.num_2M_pages--;
450 		else if (shift == PUD_SHIFT)
451 			kvm->stat.num_1G_pages--;
452 	}
453 
454 	gpa &= ~(page_size - 1);
455 	hpa = old & PTE_RPN_MASK;
456 	kvmhv_remove_nest_rmap_range(kvm, memslot, gpa, hpa, page_size);
457 
458 	if ((old & _PAGE_DIRTY) && memslot->dirty_bitmap)
459 		kvmppc_update_dirty_map(memslot, gfn, page_size);
460 }
461 
462 /*
463  * kvmppc_free_p?d are used to free existing page tables, and recursively
464  * descend and clear and free children.
465  * Callers are responsible for flushing the PWC.
466  *
467  * When page tables are being unmapped/freed as part of page fault path
468  * (full == false), valid ptes are generally not expected; however, there
469  * is one situation where they arise, which is when dirty page logging is
470  * turned off for a memslot while the VM is running.  The new memslot
471  * becomes visible to page faults before the memslot commit function
472  * gets to flush the memslot, which can lead to a 2MB page mapping being
473  * installed for a guest physical address where there are already 64kB
474  * (or 4kB) mappings (of sub-pages of the same 2MB page).
475  */
476 static void kvmppc_unmap_free_pte(struct kvm *kvm, pte_t *pte, bool full,
477 				  u64 lpid)
478 {
479 	if (full) {
480 		memset(pte, 0, sizeof(long) << RADIX_PTE_INDEX_SIZE);
481 	} else {
482 		pte_t *p = pte;
483 		unsigned long it;
484 
485 		for (it = 0; it < PTRS_PER_PTE; ++it, ++p) {
486 			if (pte_val(*p) == 0)
487 				continue;
488 			kvmppc_unmap_pte(kvm, p,
489 					 pte_pfn(*p) << PAGE_SHIFT,
490 					 PAGE_SHIFT, NULL, lpid);
491 		}
492 	}
493 
494 	kvmppc_pte_free(pte);
495 }
496 
497 static void kvmppc_unmap_free_pmd(struct kvm *kvm, pmd_t *pmd, bool full,
498 				  u64 lpid)
499 {
500 	unsigned long im;
501 	pmd_t *p = pmd;
502 
503 	for (im = 0; im < PTRS_PER_PMD; ++im, ++p) {
504 		if (!pmd_present(*p))
505 			continue;
506 		if (pmd_leaf(*p)) {
507 			if (full) {
508 				pmd_clear(p);
509 			} else {
510 				WARN_ON_ONCE(1);
511 				kvmppc_unmap_pte(kvm, (pte_t *)p,
512 					 pte_pfn(*(pte_t *)p) << PAGE_SHIFT,
513 					 PMD_SHIFT, NULL, lpid);
514 			}
515 		} else {
516 			pte_t *pte;
517 
518 			pte = pte_offset_kernel(p, 0);
519 			kvmppc_unmap_free_pte(kvm, pte, full, lpid);
520 			pmd_clear(p);
521 		}
522 	}
523 	kvmppc_pmd_free(pmd);
524 }
525 
526 static void kvmppc_unmap_free_pud(struct kvm *kvm, pud_t *pud,
527 				  u64 lpid)
528 {
529 	unsigned long iu;
530 	pud_t *p = pud;
531 
532 	for (iu = 0; iu < PTRS_PER_PUD; ++iu, ++p) {
533 		if (!pud_present(*p))
534 			continue;
535 		if (pud_leaf(*p)) {
536 			pud_clear(p);
537 		} else {
538 			pmd_t *pmd;
539 
540 			pmd = pmd_offset(p, 0);
541 			kvmppc_unmap_free_pmd(kvm, pmd, true, lpid);
542 			pud_clear(p);
543 		}
544 	}
545 	pud_free(kvm->mm, pud);
546 }
547 
548 void kvmppc_free_pgtable_radix(struct kvm *kvm, pgd_t *pgd, u64 lpid)
549 {
550 	unsigned long ig;
551 
552 	for (ig = 0; ig < PTRS_PER_PGD; ++ig, ++pgd) {
553 		p4d_t *p4d = p4d_offset(pgd, 0);
554 		pud_t *pud;
555 
556 		if (!p4d_present(*p4d))
557 			continue;
558 		pud = pud_offset(p4d, 0);
559 		kvmppc_unmap_free_pud(kvm, pud, lpid);
560 		p4d_clear(p4d);
561 	}
562 }
563 
564 void kvmppc_free_radix(struct kvm *kvm)
565 {
566 	if (kvm->arch.pgtable) {
567 		kvmppc_free_pgtable_radix(kvm, kvm->arch.pgtable,
568 					  kvm->arch.lpid);
569 		pgd_free(kvm->mm, kvm->arch.pgtable);
570 		kvm->arch.pgtable = NULL;
571 	}
572 }
573 
574 static void kvmppc_unmap_free_pmd_entry_table(struct kvm *kvm, pmd_t *pmd,
575 					unsigned long gpa, u64 lpid)
576 {
577 	pte_t *pte = pte_offset_kernel(pmd, 0);
578 
579 	/*
580 	 * Clearing the pmd entry then flushing the PWC ensures that the pte
581 	 * page no longer be cached by the MMU, so can be freed without
582 	 * flushing the PWC again.
583 	 */
584 	pmd_clear(pmd);
585 	kvmppc_radix_flush_pwc(kvm, lpid);
586 
587 	kvmppc_unmap_free_pte(kvm, pte, false, lpid);
588 }
589 
590 static void kvmppc_unmap_free_pud_entry_table(struct kvm *kvm, pud_t *pud,
591 					unsigned long gpa, u64 lpid)
592 {
593 	pmd_t *pmd = pmd_offset(pud, 0);
594 
595 	/*
596 	 * Clearing the pud entry then flushing the PWC ensures that the pmd
597 	 * page and any children pte pages will no longer be cached by the MMU,
598 	 * so can be freed without flushing the PWC again.
599 	 */
600 	pud_clear(pud);
601 	kvmppc_radix_flush_pwc(kvm, lpid);
602 
603 	kvmppc_unmap_free_pmd(kvm, pmd, false, lpid);
604 }
605 
606 /*
607  * There are a number of bits which may differ between different faults to
608  * the same partition scope entry. RC bits, in the course of cleaning and
609  * aging. And the write bit can change, either the access could have been
610  * upgraded, or a read fault could happen concurrently with a write fault
611  * that sets those bits first.
612  */
613 #define PTE_BITS_MUST_MATCH (~(_PAGE_WRITE | _PAGE_DIRTY | _PAGE_ACCESSED))
614 
615 int kvmppc_create_pte(struct kvm *kvm, pgd_t *pgtable, pte_t pte,
616 		      unsigned long gpa, unsigned int level,
617 		      unsigned long mmu_seq, u64 lpid,
618 		      unsigned long *rmapp, struct rmap_nested **n_rmap)
619 {
620 	pgd_t *pgd;
621 	p4d_t *p4d;
622 	pud_t *pud, *new_pud = NULL;
623 	pmd_t *pmd, *new_pmd = NULL;
624 	pte_t *ptep, *new_ptep = NULL;
625 	int ret;
626 
627 	/* Traverse the guest's 2nd-level tree, allocate new levels needed */
628 	pgd = pgtable + pgd_index(gpa);
629 	p4d = p4d_offset(pgd, gpa);
630 
631 	pud = NULL;
632 	if (p4d_present(*p4d))
633 		pud = pud_offset(p4d, gpa);
634 	else
635 		new_pud = pud_alloc_one(kvm->mm, gpa);
636 
637 	pmd = NULL;
638 	if (pud && pud_present(*pud) && !pud_leaf(*pud))
639 		pmd = pmd_offset(pud, gpa);
640 	else if (level <= 1)
641 		new_pmd = kvmppc_pmd_alloc();
642 
643 	if (level == 0 && !(pmd && pmd_present(*pmd) && !pmd_leaf(*pmd)))
644 		new_ptep = kvmppc_pte_alloc();
645 
646 	/* Check if we might have been invalidated; let the guest retry if so */
647 	spin_lock(&kvm->mmu_lock);
648 	ret = -EAGAIN;
649 	if (mmu_invalidate_retry(kvm, mmu_seq))
650 		goto out_unlock;
651 
652 	/* Now traverse again under the lock and change the tree */
653 	ret = -ENOMEM;
654 	if (p4d_none(*p4d)) {
655 		if (!new_pud)
656 			goto out_unlock;
657 		p4d_populate(kvm->mm, p4d, new_pud);
658 		new_pud = NULL;
659 	}
660 	pud = pud_offset(p4d, gpa);
661 	if (pud_leaf(*pud)) {
662 		unsigned long hgpa = gpa & PUD_MASK;
663 
664 		/* Check if we raced and someone else has set the same thing */
665 		if (level == 2) {
666 			if (pud_raw(*pud) == pte_raw(pte)) {
667 				ret = 0;
668 				goto out_unlock;
669 			}
670 			/* Valid 1GB page here already, add our extra bits */
671 			WARN_ON_ONCE((pud_val(*pud) ^ pte_val(pte)) &
672 							PTE_BITS_MUST_MATCH);
673 			kvmppc_radix_update_pte(kvm, (pte_t *)pud,
674 					      0, pte_val(pte), hgpa, PUD_SHIFT);
675 			ret = 0;
676 			goto out_unlock;
677 		}
678 		/*
679 		 * If we raced with another CPU which has just put
680 		 * a 1GB pte in after we saw a pmd page, try again.
681 		 */
682 		if (!new_pmd) {
683 			ret = -EAGAIN;
684 			goto out_unlock;
685 		}
686 		/* Valid 1GB page here already, remove it */
687 		kvmppc_unmap_pte(kvm, (pte_t *)pud, hgpa, PUD_SHIFT, NULL,
688 				 lpid);
689 	}
690 	if (level == 2) {
691 		if (!pud_none(*pud)) {
692 			/*
693 			 * There's a page table page here, but we wanted to
694 			 * install a large page, so remove and free the page
695 			 * table page.
696 			 */
697 			kvmppc_unmap_free_pud_entry_table(kvm, pud, gpa, lpid);
698 		}
699 		kvmppc_radix_set_pte_at(kvm, gpa, (pte_t *)pud, pte);
700 		if (rmapp && n_rmap)
701 			kvmhv_insert_nest_rmap(kvm, rmapp, n_rmap);
702 		ret = 0;
703 		goto out_unlock;
704 	}
705 	if (pud_none(*pud)) {
706 		if (!new_pmd)
707 			goto out_unlock;
708 		pud_populate(kvm->mm, pud, new_pmd);
709 		new_pmd = NULL;
710 	}
711 	pmd = pmd_offset(pud, gpa);
712 	if (pmd_leaf(*pmd)) {
713 		unsigned long lgpa = gpa & PMD_MASK;
714 
715 		/* Check if we raced and someone else has set the same thing */
716 		if (level == 1) {
717 			if (pmd_raw(*pmd) == pte_raw(pte)) {
718 				ret = 0;
719 				goto out_unlock;
720 			}
721 			/* Valid 2MB page here already, add our extra bits */
722 			WARN_ON_ONCE((pmd_val(*pmd) ^ pte_val(pte)) &
723 							PTE_BITS_MUST_MATCH);
724 			kvmppc_radix_update_pte(kvm, pmdp_ptep(pmd),
725 					0, pte_val(pte), lgpa, PMD_SHIFT);
726 			ret = 0;
727 			goto out_unlock;
728 		}
729 
730 		/*
731 		 * If we raced with another CPU which has just put
732 		 * a 2MB pte in after we saw a pte page, try again.
733 		 */
734 		if (!new_ptep) {
735 			ret = -EAGAIN;
736 			goto out_unlock;
737 		}
738 		/* Valid 2MB page here already, remove it */
739 		kvmppc_unmap_pte(kvm, pmdp_ptep(pmd), lgpa, PMD_SHIFT, NULL,
740 				 lpid);
741 	}
742 	if (level == 1) {
743 		if (!pmd_none(*pmd)) {
744 			/*
745 			 * There's a page table page here, but we wanted to
746 			 * install a large page, so remove and free the page
747 			 * table page.
748 			 */
749 			kvmppc_unmap_free_pmd_entry_table(kvm, pmd, gpa, lpid);
750 		}
751 		kvmppc_radix_set_pte_at(kvm, gpa, pmdp_ptep(pmd), pte);
752 		if (rmapp && n_rmap)
753 			kvmhv_insert_nest_rmap(kvm, rmapp, n_rmap);
754 		ret = 0;
755 		goto out_unlock;
756 	}
757 	if (pmd_none(*pmd)) {
758 		if (!new_ptep)
759 			goto out_unlock;
760 		pmd_populate(kvm->mm, pmd, new_ptep);
761 		new_ptep = NULL;
762 	}
763 	ptep = pte_offset_kernel(pmd, gpa);
764 	if (pte_present(*ptep)) {
765 		/* Check if someone else set the same thing */
766 		if (pte_raw(*ptep) == pte_raw(pte)) {
767 			ret = 0;
768 			goto out_unlock;
769 		}
770 		/* Valid page here already, add our extra bits */
771 		WARN_ON_ONCE((pte_val(*ptep) ^ pte_val(pte)) &
772 							PTE_BITS_MUST_MATCH);
773 		kvmppc_radix_update_pte(kvm, ptep, 0, pte_val(pte), gpa, 0);
774 		ret = 0;
775 		goto out_unlock;
776 	}
777 	kvmppc_radix_set_pte_at(kvm, gpa, ptep, pte);
778 	if (rmapp && n_rmap)
779 		kvmhv_insert_nest_rmap(kvm, rmapp, n_rmap);
780 	ret = 0;
781 
782  out_unlock:
783 	spin_unlock(&kvm->mmu_lock);
784 	if (new_pud)
785 		pud_free(kvm->mm, new_pud);
786 	if (new_pmd)
787 		kvmppc_pmd_free(new_pmd);
788 	if (new_ptep)
789 		kvmppc_pte_free(new_ptep);
790 	return ret;
791 }
792 
793 bool kvmppc_hv_handle_set_rc(struct kvm *kvm, bool nested, bool writing,
794 			     unsigned long gpa, u64 lpid)
795 {
796 	unsigned long pgflags;
797 	unsigned int shift;
798 	pte_t *ptep;
799 
800 	/*
801 	 * Need to set an R or C bit in the 2nd-level tables;
802 	 * since we are just helping out the hardware here,
803 	 * it is sufficient to do what the hardware does.
804 	 */
805 	pgflags = _PAGE_ACCESSED;
806 	if (writing)
807 		pgflags |= _PAGE_DIRTY;
808 
809 	if (nested)
810 		ptep = find_kvm_nested_guest_pte(kvm, lpid, gpa, &shift);
811 	else
812 		ptep = find_kvm_secondary_pte(kvm, gpa, &shift);
813 
814 	if (ptep && pte_present(*ptep) && (!writing || pte_write(*ptep))) {
815 		kvmppc_radix_update_pte(kvm, ptep, 0, pgflags, gpa, shift);
816 		return true;
817 	}
818 	return false;
819 }
820 
821 int kvmppc_book3s_instantiate_page(struct kvm_vcpu *vcpu,
822 				   unsigned long gpa,
823 				   struct kvm_memory_slot *memslot,
824 				   bool writing, bool kvm_ro,
825 				   pte_t *inserted_pte, unsigned int *levelp)
826 {
827 	struct kvm *kvm = vcpu->kvm;
828 	struct page *page = NULL;
829 	unsigned long mmu_seq;
830 	unsigned long hva, gfn = gpa >> PAGE_SHIFT;
831 	bool upgrade_write = false;
832 	bool *upgrade_p = &upgrade_write;
833 	pte_t pte, *ptep;
834 	unsigned int shift, level;
835 	int ret;
836 	bool large_enable;
837 
838 	/* used to check for invalidations in progress */
839 	mmu_seq = kvm->mmu_invalidate_seq;
840 	smp_rmb();
841 
842 	/*
843 	 * Do a fast check first, since __gfn_to_pfn_memslot doesn't
844 	 * do it with !atomic && !async, which is how we call it.
845 	 * We always ask for write permission since the common case
846 	 * is that the page is writable.
847 	 */
848 	hva = gfn_to_hva_memslot(memslot, gfn);
849 	if (!kvm_ro && get_user_page_fast_only(hva, FOLL_WRITE, &page)) {
850 		upgrade_write = true;
851 	} else {
852 		unsigned long pfn;
853 
854 		/* Call KVM generic code to do the slow-path check */
855 		pfn = __gfn_to_pfn_memslot(memslot, gfn, false, false, NULL,
856 					   writing, upgrade_p, NULL);
857 		if (is_error_noslot_pfn(pfn))
858 			return -EFAULT;
859 		page = NULL;
860 		if (pfn_valid(pfn)) {
861 			page = pfn_to_page(pfn);
862 			if (PageReserved(page))
863 				page = NULL;
864 		}
865 	}
866 
867 	/*
868 	 * Read the PTE from the process' radix tree and use that
869 	 * so we get the shift and attribute bits.
870 	 */
871 	spin_lock(&kvm->mmu_lock);
872 	ptep = find_kvm_host_pte(kvm, mmu_seq, hva, &shift);
873 	pte = __pte(0);
874 	if (ptep)
875 		pte = READ_ONCE(*ptep);
876 	spin_unlock(&kvm->mmu_lock);
877 	/*
878 	 * If the PTE disappeared temporarily due to a THP
879 	 * collapse, just return and let the guest try again.
880 	 */
881 	if (!pte_present(pte)) {
882 		if (page)
883 			put_page(page);
884 		return RESUME_GUEST;
885 	}
886 
887 	/* If we're logging dirty pages, always map single pages */
888 	large_enable = !(memslot->flags & KVM_MEM_LOG_DIRTY_PAGES);
889 
890 	/* Get pte level from shift/size */
891 	if (large_enable && shift == PUD_SHIFT &&
892 	    (gpa & (PUD_SIZE - PAGE_SIZE)) ==
893 	    (hva & (PUD_SIZE - PAGE_SIZE))) {
894 		level = 2;
895 	} else if (large_enable && shift == PMD_SHIFT &&
896 		   (gpa & (PMD_SIZE - PAGE_SIZE)) ==
897 		   (hva & (PMD_SIZE - PAGE_SIZE))) {
898 		level = 1;
899 	} else {
900 		level = 0;
901 		if (shift > PAGE_SHIFT) {
902 			/*
903 			 * If the pte maps more than one page, bring over
904 			 * bits from the virtual address to get the real
905 			 * address of the specific single page we want.
906 			 */
907 			unsigned long rpnmask = (1ul << shift) - PAGE_SIZE;
908 			pte = __pte(pte_val(pte) | (hva & rpnmask));
909 		}
910 	}
911 
912 	pte = __pte(pte_val(pte) | _PAGE_EXEC | _PAGE_ACCESSED);
913 	if (writing || upgrade_write) {
914 		if (pte_val(pte) & _PAGE_WRITE)
915 			pte = __pte(pte_val(pte) | _PAGE_DIRTY);
916 	} else {
917 		pte = __pte(pte_val(pte) & ~(_PAGE_WRITE | _PAGE_DIRTY));
918 	}
919 
920 	/* Allocate space in the tree and write the PTE */
921 	ret = kvmppc_create_pte(kvm, kvm->arch.pgtable, pte, gpa, level,
922 				mmu_seq, kvm->arch.lpid, NULL, NULL);
923 	if (inserted_pte)
924 		*inserted_pte = pte;
925 	if (levelp)
926 		*levelp = level;
927 
928 	if (page) {
929 		if (!ret && (pte_val(pte) & _PAGE_WRITE))
930 			set_page_dirty_lock(page);
931 		put_page(page);
932 	}
933 
934 	/* Increment number of large pages if we (successfully) inserted one */
935 	if (!ret) {
936 		if (level == 1)
937 			kvm->stat.num_2M_pages++;
938 		else if (level == 2)
939 			kvm->stat.num_1G_pages++;
940 	}
941 
942 	return ret;
943 }
944 
945 int kvmppc_book3s_radix_page_fault(struct kvm_vcpu *vcpu,
946 				   unsigned long ea, unsigned long dsisr)
947 {
948 	struct kvm *kvm = vcpu->kvm;
949 	unsigned long gpa, gfn;
950 	struct kvm_memory_slot *memslot;
951 	long ret;
952 	bool writing = !!(dsisr & DSISR_ISSTORE);
953 	bool kvm_ro = false;
954 
955 	/* Check for unusual errors */
956 	if (dsisr & DSISR_UNSUPP_MMU) {
957 		pr_err("KVM: Got unsupported MMU fault\n");
958 		return -EFAULT;
959 	}
960 	if (dsisr & DSISR_BADACCESS) {
961 		/* Reflect to the guest as DSI */
962 		pr_err("KVM: Got radix HV page fault with DSISR=%lx\n", dsisr);
963 		kvmppc_core_queue_data_storage(vcpu,
964 				kvmppc_get_msr(vcpu) & SRR1_PREFIXED,
965 				ea, dsisr);
966 		return RESUME_GUEST;
967 	}
968 
969 	/* Translate the logical address */
970 	gpa = vcpu->arch.fault_gpa & ~0xfffUL;
971 	gpa &= ~0xF000000000000000ul;
972 	gfn = gpa >> PAGE_SHIFT;
973 	if (!(dsisr & DSISR_PRTABLE_FAULT))
974 		gpa |= ea & 0xfff;
975 
976 	if (kvm->arch.secure_guest & KVMPPC_SECURE_INIT_DONE)
977 		return kvmppc_send_page_to_uv(kvm, gfn);
978 
979 	/* Get the corresponding memslot */
980 	memslot = gfn_to_memslot(kvm, gfn);
981 
982 	/* No memslot means it's an emulated MMIO region */
983 	if (!memslot || (memslot->flags & KVM_MEMSLOT_INVALID)) {
984 		if (dsisr & (DSISR_PRTABLE_FAULT | DSISR_BADACCESS |
985 			     DSISR_SET_RC)) {
986 			/*
987 			 * Bad address in guest page table tree, or other
988 			 * unusual error - reflect it to the guest as DSI.
989 			 */
990 			kvmppc_core_queue_data_storage(vcpu,
991 					kvmppc_get_msr(vcpu) & SRR1_PREFIXED,
992 					ea, dsisr);
993 			return RESUME_GUEST;
994 		}
995 		return kvmppc_hv_emulate_mmio(vcpu, gpa, ea, writing);
996 	}
997 
998 	if (memslot->flags & KVM_MEM_READONLY) {
999 		if (writing) {
1000 			/* give the guest a DSI */
1001 			kvmppc_core_queue_data_storage(vcpu,
1002 					kvmppc_get_msr(vcpu) & SRR1_PREFIXED,
1003 					ea, DSISR_ISSTORE | DSISR_PROTFAULT);
1004 			return RESUME_GUEST;
1005 		}
1006 		kvm_ro = true;
1007 	}
1008 
1009 	/* Failed to set the reference/change bits */
1010 	if (dsisr & DSISR_SET_RC) {
1011 		spin_lock(&kvm->mmu_lock);
1012 		if (kvmppc_hv_handle_set_rc(kvm, false, writing,
1013 					    gpa, kvm->arch.lpid))
1014 			dsisr &= ~DSISR_SET_RC;
1015 		spin_unlock(&kvm->mmu_lock);
1016 
1017 		if (!(dsisr & (DSISR_BAD_FAULT_64S | DSISR_NOHPTE |
1018 			       DSISR_PROTFAULT | DSISR_SET_RC)))
1019 			return RESUME_GUEST;
1020 	}
1021 
1022 	/* Try to insert a pte */
1023 	ret = kvmppc_book3s_instantiate_page(vcpu, gpa, memslot, writing,
1024 					     kvm_ro, NULL, NULL);
1025 
1026 	if (ret == 0 || ret == -EAGAIN)
1027 		ret = RESUME_GUEST;
1028 	return ret;
1029 }
1030 
1031 /* Called with kvm->mmu_lock held */
1032 void kvm_unmap_radix(struct kvm *kvm, struct kvm_memory_slot *memslot,
1033 		     unsigned long gfn)
1034 {
1035 	pte_t *ptep;
1036 	unsigned long gpa = gfn << PAGE_SHIFT;
1037 	unsigned int shift;
1038 
1039 	if (kvm->arch.secure_guest & KVMPPC_SECURE_INIT_DONE) {
1040 		uv_page_inval(kvm->arch.lpid, gpa, PAGE_SHIFT);
1041 		return;
1042 	}
1043 
1044 	ptep = find_kvm_secondary_pte(kvm, gpa, &shift);
1045 	if (ptep && pte_present(*ptep))
1046 		kvmppc_unmap_pte(kvm, ptep, gpa, shift, memslot,
1047 				 kvm->arch.lpid);
1048 }
1049 
1050 /* Called with kvm->mmu_lock held */
1051 bool kvm_age_radix(struct kvm *kvm, struct kvm_memory_slot *memslot,
1052 		   unsigned long gfn)
1053 {
1054 	pte_t *ptep;
1055 	unsigned long gpa = gfn << PAGE_SHIFT;
1056 	unsigned int shift;
1057 	bool ref = false;
1058 	unsigned long old, *rmapp;
1059 
1060 	if (kvm->arch.secure_guest & KVMPPC_SECURE_INIT_DONE)
1061 		return ref;
1062 
1063 	ptep = find_kvm_secondary_pte(kvm, gpa, &shift);
1064 	if (ptep && pte_present(*ptep) && pte_young(*ptep)) {
1065 		old = kvmppc_radix_update_pte(kvm, ptep, _PAGE_ACCESSED, 0,
1066 					      gpa, shift);
1067 		/* XXX need to flush tlb here? */
1068 		/* Also clear bit in ptes in shadow pgtable for nested guests */
1069 		rmapp = &memslot->arch.rmap[gfn - memslot->base_gfn];
1070 		kvmhv_update_nest_rmap_rc_list(kvm, rmapp, _PAGE_ACCESSED, 0,
1071 					       old & PTE_RPN_MASK,
1072 					       1UL << shift);
1073 		ref = true;
1074 	}
1075 	return ref;
1076 }
1077 
1078 /* Called with kvm->mmu_lock held */
1079 bool kvm_test_age_radix(struct kvm *kvm, struct kvm_memory_slot *memslot,
1080 			unsigned long gfn)
1081 
1082 {
1083 	pte_t *ptep;
1084 	unsigned long gpa = gfn << PAGE_SHIFT;
1085 	unsigned int shift;
1086 	bool ref = false;
1087 
1088 	if (kvm->arch.secure_guest & KVMPPC_SECURE_INIT_DONE)
1089 		return ref;
1090 
1091 	ptep = find_kvm_secondary_pte(kvm, gpa, &shift);
1092 	if (ptep && pte_present(*ptep) && pte_young(*ptep))
1093 		ref = true;
1094 	return ref;
1095 }
1096 
1097 /* Returns the number of PAGE_SIZE pages that are dirty */
1098 static int kvm_radix_test_clear_dirty(struct kvm *kvm,
1099 				struct kvm_memory_slot *memslot, int pagenum)
1100 {
1101 	unsigned long gfn = memslot->base_gfn + pagenum;
1102 	unsigned long gpa = gfn << PAGE_SHIFT;
1103 	pte_t *ptep, pte;
1104 	unsigned int shift;
1105 	int ret = 0;
1106 	unsigned long old, *rmapp;
1107 
1108 	if (kvm->arch.secure_guest & KVMPPC_SECURE_INIT_DONE)
1109 		return ret;
1110 
1111 	/*
1112 	 * For performance reasons we don't hold kvm->mmu_lock while walking the
1113 	 * partition scoped table.
1114 	 */
1115 	ptep = find_kvm_secondary_pte_unlocked(kvm, gpa, &shift);
1116 	if (!ptep)
1117 		return 0;
1118 
1119 	pte = READ_ONCE(*ptep);
1120 	if (pte_present(pte) && pte_dirty(pte)) {
1121 		spin_lock(&kvm->mmu_lock);
1122 		/*
1123 		 * Recheck the pte again
1124 		 */
1125 		if (pte_val(pte) != pte_val(*ptep)) {
1126 			/*
1127 			 * We have KVM_MEM_LOG_DIRTY_PAGES enabled. Hence we can
1128 			 * only find PAGE_SIZE pte entries here. We can continue
1129 			 * to use the pte addr returned by above page table
1130 			 * walk.
1131 			 */
1132 			if (!pte_present(*ptep) || !pte_dirty(*ptep)) {
1133 				spin_unlock(&kvm->mmu_lock);
1134 				return 0;
1135 			}
1136 		}
1137 
1138 		ret = 1;
1139 		VM_BUG_ON(shift);
1140 		old = kvmppc_radix_update_pte(kvm, ptep, _PAGE_DIRTY, 0,
1141 					      gpa, shift);
1142 		kvmppc_radix_tlbie_page(kvm, gpa, shift, kvm->arch.lpid);
1143 		/* Also clear bit in ptes in shadow pgtable for nested guests */
1144 		rmapp = &memslot->arch.rmap[gfn - memslot->base_gfn];
1145 		kvmhv_update_nest_rmap_rc_list(kvm, rmapp, _PAGE_DIRTY, 0,
1146 					       old & PTE_RPN_MASK,
1147 					       1UL << shift);
1148 		spin_unlock(&kvm->mmu_lock);
1149 	}
1150 	return ret;
1151 }
1152 
1153 long kvmppc_hv_get_dirty_log_radix(struct kvm *kvm,
1154 			struct kvm_memory_slot *memslot, unsigned long *map)
1155 {
1156 	unsigned long i, j;
1157 	int npages;
1158 
1159 	for (i = 0; i < memslot->npages; i = j) {
1160 		npages = kvm_radix_test_clear_dirty(kvm, memslot, i);
1161 
1162 		/*
1163 		 * Note that if npages > 0 then i must be a multiple of npages,
1164 		 * since huge pages are only used to back the guest at guest
1165 		 * real addresses that are a multiple of their size.
1166 		 * Since we have at most one PTE covering any given guest
1167 		 * real address, if npages > 1 we can skip to i + npages.
1168 		 */
1169 		j = i + 1;
1170 		if (npages) {
1171 			set_dirty_bits(map, i, npages);
1172 			j = i + npages;
1173 		}
1174 	}
1175 	return 0;
1176 }
1177 
1178 void kvmppc_radix_flush_memslot(struct kvm *kvm,
1179 				const struct kvm_memory_slot *memslot)
1180 {
1181 	unsigned long n;
1182 	pte_t *ptep;
1183 	unsigned long gpa;
1184 	unsigned int shift;
1185 
1186 	if (kvm->arch.secure_guest & KVMPPC_SECURE_INIT_START)
1187 		kvmppc_uvmem_drop_pages(memslot, kvm, true);
1188 
1189 	if (kvm->arch.secure_guest & KVMPPC_SECURE_INIT_DONE)
1190 		return;
1191 
1192 	gpa = memslot->base_gfn << PAGE_SHIFT;
1193 	spin_lock(&kvm->mmu_lock);
1194 	for (n = memslot->npages; n; --n) {
1195 		ptep = find_kvm_secondary_pte(kvm, gpa, &shift);
1196 		if (ptep && pte_present(*ptep))
1197 			kvmppc_unmap_pte(kvm, ptep, gpa, shift, memslot,
1198 					 kvm->arch.lpid);
1199 		gpa += PAGE_SIZE;
1200 	}
1201 	/*
1202 	 * Increase the mmu notifier sequence number to prevent any page
1203 	 * fault that read the memslot earlier from writing a PTE.
1204 	 */
1205 	kvm->mmu_invalidate_seq++;
1206 	spin_unlock(&kvm->mmu_lock);
1207 }
1208 
1209 static void add_rmmu_ap_encoding(struct kvm_ppc_rmmu_info *info,
1210 				 int psize, int *indexp)
1211 {
1212 	if (!mmu_psize_defs[psize].shift)
1213 		return;
1214 	info->ap_encodings[*indexp] = mmu_psize_defs[psize].shift |
1215 		(mmu_psize_defs[psize].ap << 29);
1216 	++(*indexp);
1217 }
1218 
1219 int kvmhv_get_rmmu_info(struct kvm *kvm, struct kvm_ppc_rmmu_info *info)
1220 {
1221 	int i;
1222 
1223 	if (!radix_enabled())
1224 		return -EINVAL;
1225 	memset(info, 0, sizeof(*info));
1226 
1227 	/* 4k page size */
1228 	info->geometries[0].page_shift = 12;
1229 	info->geometries[0].level_bits[0] = 9;
1230 	for (i = 1; i < 4; ++i)
1231 		info->geometries[0].level_bits[i] = p9_supported_radix_bits[i];
1232 	/* 64k page size */
1233 	info->geometries[1].page_shift = 16;
1234 	for (i = 0; i < 4; ++i)
1235 		info->geometries[1].level_bits[i] = p9_supported_radix_bits[i];
1236 
1237 	i = 0;
1238 	add_rmmu_ap_encoding(info, MMU_PAGE_4K, &i);
1239 	add_rmmu_ap_encoding(info, MMU_PAGE_64K, &i);
1240 	add_rmmu_ap_encoding(info, MMU_PAGE_2M, &i);
1241 	add_rmmu_ap_encoding(info, MMU_PAGE_1G, &i);
1242 
1243 	return 0;
1244 }
1245 
1246 int kvmppc_init_vm_radix(struct kvm *kvm)
1247 {
1248 	kvm->arch.pgtable = pgd_alloc(kvm->mm);
1249 	if (!kvm->arch.pgtable)
1250 		return -ENOMEM;
1251 	return 0;
1252 }
1253 
1254 static void pte_ctor(void *addr)
1255 {
1256 	memset(addr, 0, RADIX_PTE_TABLE_SIZE);
1257 }
1258 
1259 static void pmd_ctor(void *addr)
1260 {
1261 	memset(addr, 0, RADIX_PMD_TABLE_SIZE);
1262 }
1263 
1264 struct debugfs_radix_state {
1265 	struct kvm	*kvm;
1266 	struct mutex	mutex;
1267 	unsigned long	gpa;
1268 	int		lpid;
1269 	int		chars_left;
1270 	int		buf_index;
1271 	char		buf[128];
1272 	u8		hdr;
1273 };
1274 
1275 static int debugfs_radix_open(struct inode *inode, struct file *file)
1276 {
1277 	struct kvm *kvm = inode->i_private;
1278 	struct debugfs_radix_state *p;
1279 
1280 	p = kzalloc(sizeof(*p), GFP_KERNEL);
1281 	if (!p)
1282 		return -ENOMEM;
1283 
1284 	kvm_get_kvm(kvm);
1285 	p->kvm = kvm;
1286 	mutex_init(&p->mutex);
1287 	file->private_data = p;
1288 
1289 	return nonseekable_open(inode, file);
1290 }
1291 
1292 static int debugfs_radix_release(struct inode *inode, struct file *file)
1293 {
1294 	struct debugfs_radix_state *p = file->private_data;
1295 
1296 	kvm_put_kvm(p->kvm);
1297 	kfree(p);
1298 	return 0;
1299 }
1300 
1301 static ssize_t debugfs_radix_read(struct file *file, char __user *buf,
1302 				 size_t len, loff_t *ppos)
1303 {
1304 	struct debugfs_radix_state *p = file->private_data;
1305 	ssize_t ret, r;
1306 	unsigned long n;
1307 	struct kvm *kvm;
1308 	unsigned long gpa;
1309 	pgd_t *pgt;
1310 	struct kvm_nested_guest *nested;
1311 	pgd_t *pgdp;
1312 	p4d_t p4d, *p4dp;
1313 	pud_t pud, *pudp;
1314 	pmd_t pmd, *pmdp;
1315 	pte_t *ptep;
1316 	int shift;
1317 	unsigned long pte;
1318 
1319 	kvm = p->kvm;
1320 	if (!kvm_is_radix(kvm))
1321 		return 0;
1322 
1323 	ret = mutex_lock_interruptible(&p->mutex);
1324 	if (ret)
1325 		return ret;
1326 
1327 	if (p->chars_left) {
1328 		n = p->chars_left;
1329 		if (n > len)
1330 			n = len;
1331 		r = copy_to_user(buf, p->buf + p->buf_index, n);
1332 		n -= r;
1333 		p->chars_left -= n;
1334 		p->buf_index += n;
1335 		buf += n;
1336 		len -= n;
1337 		ret = n;
1338 		if (r) {
1339 			if (!n)
1340 				ret = -EFAULT;
1341 			goto out;
1342 		}
1343 	}
1344 
1345 	gpa = p->gpa;
1346 	nested = NULL;
1347 	pgt = NULL;
1348 	while (len != 0 && p->lpid >= 0) {
1349 		if (gpa >= RADIX_PGTABLE_RANGE) {
1350 			gpa = 0;
1351 			pgt = NULL;
1352 			if (nested) {
1353 				kvmhv_put_nested(nested);
1354 				nested = NULL;
1355 			}
1356 			p->lpid = kvmhv_nested_next_lpid(kvm, p->lpid);
1357 			p->hdr = 0;
1358 			if (p->lpid < 0)
1359 				break;
1360 		}
1361 		if (!pgt) {
1362 			if (p->lpid == 0) {
1363 				pgt = kvm->arch.pgtable;
1364 			} else {
1365 				nested = kvmhv_get_nested(kvm, p->lpid, false);
1366 				if (!nested) {
1367 					gpa = RADIX_PGTABLE_RANGE;
1368 					continue;
1369 				}
1370 				pgt = nested->shadow_pgtable;
1371 			}
1372 		}
1373 		n = 0;
1374 		if (!p->hdr) {
1375 			if (p->lpid > 0)
1376 				n = scnprintf(p->buf, sizeof(p->buf),
1377 					      "\nNested LPID %d: ", p->lpid);
1378 			n += scnprintf(p->buf + n, sizeof(p->buf) - n,
1379 				      "pgdir: %lx\n", (unsigned long)pgt);
1380 			p->hdr = 1;
1381 			goto copy;
1382 		}
1383 
1384 		pgdp = pgt + pgd_index(gpa);
1385 		p4dp = p4d_offset(pgdp, gpa);
1386 		p4d = READ_ONCE(*p4dp);
1387 		if (!(p4d_val(p4d) & _PAGE_PRESENT)) {
1388 			gpa = (gpa & P4D_MASK) + P4D_SIZE;
1389 			continue;
1390 		}
1391 
1392 		pudp = pud_offset(&p4d, gpa);
1393 		pud = READ_ONCE(*pudp);
1394 		if (!(pud_val(pud) & _PAGE_PRESENT)) {
1395 			gpa = (gpa & PUD_MASK) + PUD_SIZE;
1396 			continue;
1397 		}
1398 		if (pud_val(pud) & _PAGE_PTE) {
1399 			pte = pud_val(pud);
1400 			shift = PUD_SHIFT;
1401 			goto leaf;
1402 		}
1403 
1404 		pmdp = pmd_offset(&pud, gpa);
1405 		pmd = READ_ONCE(*pmdp);
1406 		if (!(pmd_val(pmd) & _PAGE_PRESENT)) {
1407 			gpa = (gpa & PMD_MASK) + PMD_SIZE;
1408 			continue;
1409 		}
1410 		if (pmd_val(pmd) & _PAGE_PTE) {
1411 			pte = pmd_val(pmd);
1412 			shift = PMD_SHIFT;
1413 			goto leaf;
1414 		}
1415 
1416 		ptep = pte_offset_kernel(&pmd, gpa);
1417 		pte = pte_val(READ_ONCE(*ptep));
1418 		if (!(pte & _PAGE_PRESENT)) {
1419 			gpa += PAGE_SIZE;
1420 			continue;
1421 		}
1422 		shift = PAGE_SHIFT;
1423 	leaf:
1424 		n = scnprintf(p->buf, sizeof(p->buf),
1425 			      " %lx: %lx %d\n", gpa, pte, shift);
1426 		gpa += 1ul << shift;
1427 	copy:
1428 		p->chars_left = n;
1429 		if (n > len)
1430 			n = len;
1431 		r = copy_to_user(buf, p->buf, n);
1432 		n -= r;
1433 		p->chars_left -= n;
1434 		p->buf_index = n;
1435 		buf += n;
1436 		len -= n;
1437 		ret += n;
1438 		if (r) {
1439 			if (!ret)
1440 				ret = -EFAULT;
1441 			break;
1442 		}
1443 	}
1444 	p->gpa = gpa;
1445 	if (nested)
1446 		kvmhv_put_nested(nested);
1447 
1448  out:
1449 	mutex_unlock(&p->mutex);
1450 	return ret;
1451 }
1452 
1453 static ssize_t debugfs_radix_write(struct file *file, const char __user *buf,
1454 			   size_t len, loff_t *ppos)
1455 {
1456 	return -EACCES;
1457 }
1458 
1459 static const struct file_operations debugfs_radix_fops = {
1460 	.owner	 = THIS_MODULE,
1461 	.open	 = debugfs_radix_open,
1462 	.release = debugfs_radix_release,
1463 	.read	 = debugfs_radix_read,
1464 	.write	 = debugfs_radix_write,
1465 	.llseek	 = generic_file_llseek,
1466 };
1467 
1468 void kvmhv_radix_debugfs_init(struct kvm *kvm)
1469 {
1470 	debugfs_create_file("radix", 0400, kvm->debugfs_dentry, kvm,
1471 			    &debugfs_radix_fops);
1472 }
1473 
1474 int kvmppc_radix_init(void)
1475 {
1476 	unsigned long size = sizeof(void *) << RADIX_PTE_INDEX_SIZE;
1477 
1478 	kvm_pte_cache = kmem_cache_create("kvm-pte", size, size, 0, pte_ctor);
1479 	if (!kvm_pte_cache)
1480 		return -ENOMEM;
1481 
1482 	size = sizeof(void *) << RADIX_PMD_INDEX_SIZE;
1483 
1484 	kvm_pmd_cache = kmem_cache_create("kvm-pmd", size, size, 0, pmd_ctor);
1485 	if (!kvm_pmd_cache) {
1486 		kmem_cache_destroy(kvm_pte_cache);
1487 		return -ENOMEM;
1488 	}
1489 
1490 	return 0;
1491 }
1492 
1493 void kvmppc_radix_exit(void)
1494 {
1495 	kmem_cache_destroy(kvm_pte_cache);
1496 	kmem_cache_destroy(kvm_pmd_cache);
1497 }
1498