1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  *
4  * Copyright 2010 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/highmem.h>
12 #include <linux/gfp.h>
13 #include <linux/slab.h>
14 #include <linux/hugetlb.h>
15 #include <linux/vmalloc.h>
16 #include <linux/srcu.h>
17 #include <linux/anon_inodes.h>
18 #include <linux/file.h>
19 #include <linux/debugfs.h>
20 
21 #include <asm/kvm_ppc.h>
22 #include <asm/kvm_book3s.h>
23 #include <asm/book3s/64/mmu-hash.h>
24 #include <asm/hvcall.h>
25 #include <asm/synch.h>
26 #include <asm/ppc-opcode.h>
27 #include <asm/cputable.h>
28 #include <asm/pte-walk.h>
29 
30 #include "book3s.h"
31 #include "trace_hv.h"
32 
33 //#define DEBUG_RESIZE_HPT	1
34 
35 #ifdef DEBUG_RESIZE_HPT
36 #define resize_hpt_debug(resize, ...)				\
37 	do {							\
38 		printk(KERN_DEBUG "RESIZE HPT %p: ", resize);	\
39 		printk(__VA_ARGS__);				\
40 	} while (0)
41 #else
42 #define resize_hpt_debug(resize, ...)				\
43 	do { } while (0)
44 #endif
45 
46 static long kvmppc_virtmode_do_h_enter(struct kvm *kvm, unsigned long flags,
47 				long pte_index, unsigned long pteh,
48 				unsigned long ptel, unsigned long *pte_idx_ret);
49 
50 struct kvm_resize_hpt {
51 	/* These fields read-only after init */
52 	struct kvm *kvm;
53 	struct work_struct work;
54 	u32 order;
55 
56 	/* These fields protected by kvm->arch.mmu_setup_lock */
57 
58 	/* Possible values and their usage:
59 	 *  <0     an error occurred during allocation,
60 	 *  -EBUSY allocation is in the progress,
61 	 *  0      allocation made successfuly.
62 	 */
63 	int error;
64 
65 	/* Private to the work thread, until error != -EBUSY,
66 	 * then protected by kvm->arch.mmu_setup_lock.
67 	 */
68 	struct kvm_hpt_info hpt;
69 };
70 
71 int kvmppc_allocate_hpt(struct kvm_hpt_info *info, u32 order)
72 {
73 	unsigned long hpt = 0;
74 	int cma = 0;
75 	struct page *page = NULL;
76 	struct revmap_entry *rev;
77 	unsigned long npte;
78 
79 	if ((order < PPC_MIN_HPT_ORDER) || (order > PPC_MAX_HPT_ORDER))
80 		return -EINVAL;
81 
82 	page = kvm_alloc_hpt_cma(1ul << (order - PAGE_SHIFT));
83 	if (page) {
84 		hpt = (unsigned long)pfn_to_kaddr(page_to_pfn(page));
85 		memset((void *)hpt, 0, (1ul << order));
86 		cma = 1;
87 	}
88 
89 	if (!hpt)
90 		hpt = __get_free_pages(GFP_KERNEL|__GFP_ZERO|__GFP_RETRY_MAYFAIL
91 				       |__GFP_NOWARN, order - PAGE_SHIFT);
92 
93 	if (!hpt)
94 		return -ENOMEM;
95 
96 	/* HPTEs are 2**4 bytes long */
97 	npte = 1ul << (order - 4);
98 
99 	/* Allocate reverse map array */
100 	rev = vmalloc(array_size(npte, sizeof(struct revmap_entry)));
101 	if (!rev) {
102 		if (cma)
103 			kvm_free_hpt_cma(page, 1 << (order - PAGE_SHIFT));
104 		else
105 			free_pages(hpt, order - PAGE_SHIFT);
106 		return -ENOMEM;
107 	}
108 
109 	info->order = order;
110 	info->virt = hpt;
111 	info->cma = cma;
112 	info->rev = rev;
113 
114 	return 0;
115 }
116 
117 void kvmppc_set_hpt(struct kvm *kvm, struct kvm_hpt_info *info)
118 {
119 	atomic64_set(&kvm->arch.mmio_update, 0);
120 	kvm->arch.hpt = *info;
121 	kvm->arch.sdr1 = __pa(info->virt) | (info->order - 18);
122 
123 	pr_debug("KVM guest htab at %lx (order %ld), LPID %x\n",
124 		 info->virt, (long)info->order, kvm->arch.lpid);
125 }
126 
127 long kvmppc_alloc_reset_hpt(struct kvm *kvm, int order)
128 {
129 	long err = -EBUSY;
130 	struct kvm_hpt_info info;
131 
132 	mutex_lock(&kvm->arch.mmu_setup_lock);
133 	if (kvm->arch.mmu_ready) {
134 		kvm->arch.mmu_ready = 0;
135 		/* order mmu_ready vs. vcpus_running */
136 		smp_mb();
137 		if (atomic_read(&kvm->arch.vcpus_running)) {
138 			kvm->arch.mmu_ready = 1;
139 			goto out;
140 		}
141 	}
142 	if (kvm_is_radix(kvm)) {
143 		err = kvmppc_switch_mmu_to_hpt(kvm);
144 		if (err)
145 			goto out;
146 	}
147 
148 	if (kvm->arch.hpt.order == order) {
149 		/* We already have a suitable HPT */
150 
151 		/* Set the entire HPT to 0, i.e. invalid HPTEs */
152 		memset((void *)kvm->arch.hpt.virt, 0, 1ul << order);
153 		/*
154 		 * Reset all the reverse-mapping chains for all memslots
155 		 */
156 		kvmppc_rmap_reset(kvm);
157 		err = 0;
158 		goto out;
159 	}
160 
161 	if (kvm->arch.hpt.virt) {
162 		kvmppc_free_hpt(&kvm->arch.hpt);
163 		kvmppc_rmap_reset(kvm);
164 	}
165 
166 	err = kvmppc_allocate_hpt(&info, order);
167 	if (err < 0)
168 		goto out;
169 	kvmppc_set_hpt(kvm, &info);
170 
171 out:
172 	if (err == 0)
173 		/* Ensure that each vcpu will flush its TLB on next entry. */
174 		cpumask_setall(&kvm->arch.need_tlb_flush);
175 
176 	mutex_unlock(&kvm->arch.mmu_setup_lock);
177 	return err;
178 }
179 
180 void kvmppc_free_hpt(struct kvm_hpt_info *info)
181 {
182 	vfree(info->rev);
183 	info->rev = NULL;
184 	if (info->cma)
185 		kvm_free_hpt_cma(virt_to_page(info->virt),
186 				 1 << (info->order - PAGE_SHIFT));
187 	else if (info->virt)
188 		free_pages(info->virt, info->order - PAGE_SHIFT);
189 	info->virt = 0;
190 	info->order = 0;
191 }
192 
193 /* Bits in first HPTE dword for pagesize 4k, 64k or 16M */
194 static inline unsigned long hpte0_pgsize_encoding(unsigned long pgsize)
195 {
196 	return (pgsize > 0x1000) ? HPTE_V_LARGE : 0;
197 }
198 
199 /* Bits in second HPTE dword for pagesize 4k, 64k or 16M */
200 static inline unsigned long hpte1_pgsize_encoding(unsigned long pgsize)
201 {
202 	return (pgsize == 0x10000) ? 0x1000 : 0;
203 }
204 
205 void kvmppc_map_vrma(struct kvm_vcpu *vcpu, struct kvm_memory_slot *memslot,
206 		     unsigned long porder)
207 {
208 	unsigned long i;
209 	unsigned long npages;
210 	unsigned long hp_v, hp_r;
211 	unsigned long addr, hash;
212 	unsigned long psize;
213 	unsigned long hp0, hp1;
214 	unsigned long idx_ret;
215 	long ret;
216 	struct kvm *kvm = vcpu->kvm;
217 
218 	psize = 1ul << porder;
219 	npages = memslot->npages >> (porder - PAGE_SHIFT);
220 
221 	/* VRMA can't be > 1TB */
222 	if (npages > 1ul << (40 - porder))
223 		npages = 1ul << (40 - porder);
224 	/* Can't use more than 1 HPTE per HPTEG */
225 	if (npages > kvmppc_hpt_mask(&kvm->arch.hpt) + 1)
226 		npages = kvmppc_hpt_mask(&kvm->arch.hpt) + 1;
227 
228 	hp0 = HPTE_V_1TB_SEG | (VRMA_VSID << (40 - 16)) |
229 		HPTE_V_BOLTED | hpte0_pgsize_encoding(psize);
230 	hp1 = hpte1_pgsize_encoding(psize) |
231 		HPTE_R_R | HPTE_R_C | HPTE_R_M | PP_RWXX;
232 
233 	for (i = 0; i < npages; ++i) {
234 		addr = i << porder;
235 		/* can't use hpt_hash since va > 64 bits */
236 		hash = (i ^ (VRMA_VSID ^ (VRMA_VSID << 25)))
237 			& kvmppc_hpt_mask(&kvm->arch.hpt);
238 		/*
239 		 * We assume that the hash table is empty and no
240 		 * vcpus are using it at this stage.  Since we create
241 		 * at most one HPTE per HPTEG, we just assume entry 7
242 		 * is available and use it.
243 		 */
244 		hash = (hash << 3) + 7;
245 		hp_v = hp0 | ((addr >> 16) & ~0x7fUL);
246 		hp_r = hp1 | addr;
247 		ret = kvmppc_virtmode_do_h_enter(kvm, H_EXACT, hash, hp_v, hp_r,
248 						 &idx_ret);
249 		if (ret != H_SUCCESS) {
250 			pr_err("KVM: map_vrma at %lx failed, ret=%ld\n",
251 			       addr, ret);
252 			break;
253 		}
254 	}
255 }
256 
257 int kvmppc_mmu_hv_init(void)
258 {
259 	unsigned long host_lpid, rsvd_lpid;
260 
261 	if (!mmu_has_feature(MMU_FTR_LOCKLESS_TLBIE))
262 		return -EINVAL;
263 
264 	host_lpid = 0;
265 	if (cpu_has_feature(CPU_FTR_HVMODE))
266 		host_lpid = mfspr(SPRN_LPID);
267 
268 	/* POWER8 and above have 12-bit LPIDs (10-bit in POWER7) */
269 	if (cpu_has_feature(CPU_FTR_ARCH_207S))
270 		rsvd_lpid = LPID_RSVD;
271 	else
272 		rsvd_lpid = LPID_RSVD_POWER7;
273 
274 	kvmppc_init_lpid(rsvd_lpid + 1);
275 
276 	kvmppc_claim_lpid(host_lpid);
277 	/* rsvd_lpid is reserved for use in partition switching */
278 	kvmppc_claim_lpid(rsvd_lpid);
279 
280 	return 0;
281 }
282 
283 static long kvmppc_virtmode_do_h_enter(struct kvm *kvm, unsigned long flags,
284 				long pte_index, unsigned long pteh,
285 				unsigned long ptel, unsigned long *pte_idx_ret)
286 {
287 	long ret;
288 
289 	preempt_disable();
290 	ret = kvmppc_do_h_enter(kvm, flags, pte_index, pteh, ptel,
291 				kvm->mm->pgd, false, pte_idx_ret);
292 	preempt_enable();
293 	if (ret == H_TOO_HARD) {
294 		/* this can't happen */
295 		pr_err("KVM: Oops, kvmppc_h_enter returned too hard!\n");
296 		ret = H_RESOURCE;	/* or something */
297 	}
298 	return ret;
299 
300 }
301 
302 static struct kvmppc_slb *kvmppc_mmu_book3s_hv_find_slbe(struct kvm_vcpu *vcpu,
303 							 gva_t eaddr)
304 {
305 	u64 mask;
306 	int i;
307 
308 	for (i = 0; i < vcpu->arch.slb_nr; i++) {
309 		if (!(vcpu->arch.slb[i].orige & SLB_ESID_V))
310 			continue;
311 
312 		if (vcpu->arch.slb[i].origv & SLB_VSID_B_1T)
313 			mask = ESID_MASK_1T;
314 		else
315 			mask = ESID_MASK;
316 
317 		if (((vcpu->arch.slb[i].orige ^ eaddr) & mask) == 0)
318 			return &vcpu->arch.slb[i];
319 	}
320 	return NULL;
321 }
322 
323 static unsigned long kvmppc_mmu_get_real_addr(unsigned long v, unsigned long r,
324 			unsigned long ea)
325 {
326 	unsigned long ra_mask;
327 
328 	ra_mask = kvmppc_actual_pgsz(v, r) - 1;
329 	return (r & HPTE_R_RPN & ~ra_mask) | (ea & ra_mask);
330 }
331 
332 static int kvmppc_mmu_book3s_64_hv_xlate(struct kvm_vcpu *vcpu, gva_t eaddr,
333 			struct kvmppc_pte *gpte, bool data, bool iswrite)
334 {
335 	struct kvm *kvm = vcpu->kvm;
336 	struct kvmppc_slb *slbe;
337 	unsigned long slb_v;
338 	unsigned long pp, key;
339 	unsigned long v, orig_v, gr;
340 	__be64 *hptep;
341 	long int index;
342 	int virtmode = vcpu->arch.shregs.msr & (data ? MSR_DR : MSR_IR);
343 
344 	if (kvm_is_radix(vcpu->kvm))
345 		return kvmppc_mmu_radix_xlate(vcpu, eaddr, gpte, data, iswrite);
346 
347 	/* Get SLB entry */
348 	if (virtmode) {
349 		slbe = kvmppc_mmu_book3s_hv_find_slbe(vcpu, eaddr);
350 		if (!slbe)
351 			return -EINVAL;
352 		slb_v = slbe->origv;
353 	} else {
354 		/* real mode access */
355 		slb_v = vcpu->kvm->arch.vrma_slb_v;
356 	}
357 
358 	preempt_disable();
359 	/* Find the HPTE in the hash table */
360 	index = kvmppc_hv_find_lock_hpte(kvm, eaddr, slb_v,
361 					 HPTE_V_VALID | HPTE_V_ABSENT);
362 	if (index < 0) {
363 		preempt_enable();
364 		return -ENOENT;
365 	}
366 	hptep = (__be64 *)(kvm->arch.hpt.virt + (index << 4));
367 	v = orig_v = be64_to_cpu(hptep[0]) & ~HPTE_V_HVLOCK;
368 	if (cpu_has_feature(CPU_FTR_ARCH_300))
369 		v = hpte_new_to_old_v(v, be64_to_cpu(hptep[1]));
370 	gr = kvm->arch.hpt.rev[index].guest_rpte;
371 
372 	unlock_hpte(hptep, orig_v);
373 	preempt_enable();
374 
375 	gpte->eaddr = eaddr;
376 	gpte->vpage = ((v & HPTE_V_AVPN) << 4) | ((eaddr >> 12) & 0xfff);
377 
378 	/* Get PP bits and key for permission check */
379 	pp = gr & (HPTE_R_PP0 | HPTE_R_PP);
380 	key = (vcpu->arch.shregs.msr & MSR_PR) ? SLB_VSID_KP : SLB_VSID_KS;
381 	key &= slb_v;
382 
383 	/* Calculate permissions */
384 	gpte->may_read = hpte_read_permission(pp, key);
385 	gpte->may_write = hpte_write_permission(pp, key);
386 	gpte->may_execute = gpte->may_read && !(gr & (HPTE_R_N | HPTE_R_G));
387 
388 	/* Storage key permission check for POWER7 */
389 	if (data && virtmode) {
390 		int amrfield = hpte_get_skey_perm(gr, vcpu->arch.amr);
391 		if (amrfield & 1)
392 			gpte->may_read = 0;
393 		if (amrfield & 2)
394 			gpte->may_write = 0;
395 	}
396 
397 	/* Get the guest physical address */
398 	gpte->raddr = kvmppc_mmu_get_real_addr(v, gr, eaddr);
399 	return 0;
400 }
401 
402 /*
403  * Quick test for whether an instruction is a load or a store.
404  * If the instruction is a load or a store, then this will indicate
405  * which it is, at least on server processors.  (Embedded processors
406  * have some external PID instructions that don't follow the rule
407  * embodied here.)  If the instruction isn't a load or store, then
408  * this doesn't return anything useful.
409  */
410 static int instruction_is_store(unsigned int instr)
411 {
412 	unsigned int mask;
413 
414 	mask = 0x10000000;
415 	if ((instr & 0xfc000000) == 0x7c000000)
416 		mask = 0x100;		/* major opcode 31 */
417 	return (instr & mask) != 0;
418 }
419 
420 int kvmppc_hv_emulate_mmio(struct kvm_vcpu *vcpu,
421 			   unsigned long gpa, gva_t ea, int is_store)
422 {
423 	u32 last_inst;
424 
425 	/*
426 	 * Fast path - check if the guest physical address corresponds to a
427 	 * device on the FAST_MMIO_BUS, if so we can avoid loading the
428 	 * instruction all together, then we can just handle it and return.
429 	 */
430 	if (is_store) {
431 		int idx, ret;
432 
433 		idx = srcu_read_lock(&vcpu->kvm->srcu);
434 		ret = kvm_io_bus_write(vcpu, KVM_FAST_MMIO_BUS, (gpa_t) gpa, 0,
435 				       NULL);
436 		srcu_read_unlock(&vcpu->kvm->srcu, idx);
437 		if (!ret) {
438 			kvmppc_set_pc(vcpu, kvmppc_get_pc(vcpu) + 4);
439 			return RESUME_GUEST;
440 		}
441 	}
442 
443 	/*
444 	 * If we fail, we just return to the guest and try executing it again.
445 	 */
446 	if (kvmppc_get_last_inst(vcpu, INST_GENERIC, &last_inst) !=
447 		EMULATE_DONE)
448 		return RESUME_GUEST;
449 
450 	/*
451 	 * WARNING: We do not know for sure whether the instruction we just
452 	 * read from memory is the same that caused the fault in the first
453 	 * place.  If the instruction we read is neither an load or a store,
454 	 * then it can't access memory, so we don't need to worry about
455 	 * enforcing access permissions.  So, assuming it is a load or
456 	 * store, we just check that its direction (load or store) is
457 	 * consistent with the original fault, since that's what we
458 	 * checked the access permissions against.  If there is a mismatch
459 	 * we just return and retry the instruction.
460 	 */
461 
462 	if (instruction_is_store(last_inst) != !!is_store)
463 		return RESUME_GUEST;
464 
465 	/*
466 	 * Emulated accesses are emulated by looking at the hash for
467 	 * translation once, then performing the access later. The
468 	 * translation could be invalidated in the meantime in which
469 	 * point performing the subsequent memory access on the old
470 	 * physical address could possibly be a security hole for the
471 	 * guest (but not the host).
472 	 *
473 	 * This is less of an issue for MMIO stores since they aren't
474 	 * globally visible. It could be an issue for MMIO loads to
475 	 * a certain extent but we'll ignore it for now.
476 	 */
477 
478 	vcpu->arch.paddr_accessed = gpa;
479 	vcpu->arch.vaddr_accessed = ea;
480 	return kvmppc_emulate_mmio(vcpu);
481 }
482 
483 int kvmppc_book3s_hv_page_fault(struct kvm_vcpu *vcpu,
484 				unsigned long ea, unsigned long dsisr)
485 {
486 	struct kvm *kvm = vcpu->kvm;
487 	unsigned long hpte[3], r;
488 	unsigned long hnow_v, hnow_r;
489 	__be64 *hptep;
490 	unsigned long mmu_seq, psize, pte_size;
491 	unsigned long gpa_base, gfn_base;
492 	unsigned long gpa, gfn, hva, pfn, hpa;
493 	struct kvm_memory_slot *memslot;
494 	unsigned long *rmap;
495 	struct revmap_entry *rev;
496 	struct page *page;
497 	long index, ret;
498 	bool is_ci;
499 	bool writing, write_ok;
500 	unsigned int shift;
501 	unsigned long rcbits;
502 	long mmio_update;
503 	pte_t pte, *ptep;
504 
505 	if (kvm_is_radix(kvm))
506 		return kvmppc_book3s_radix_page_fault(vcpu, ea, dsisr);
507 
508 	/*
509 	 * Real-mode code has already searched the HPT and found the
510 	 * entry we're interested in.  Lock the entry and check that
511 	 * it hasn't changed.  If it has, just return and re-execute the
512 	 * instruction.
513 	 */
514 	if (ea != vcpu->arch.pgfault_addr)
515 		return RESUME_GUEST;
516 
517 	if (vcpu->arch.pgfault_cache) {
518 		mmio_update = atomic64_read(&kvm->arch.mmio_update);
519 		if (mmio_update == vcpu->arch.pgfault_cache->mmio_update) {
520 			r = vcpu->arch.pgfault_cache->rpte;
521 			psize = kvmppc_actual_pgsz(vcpu->arch.pgfault_hpte[0],
522 						   r);
523 			gpa_base = r & HPTE_R_RPN & ~(psize - 1);
524 			gfn_base = gpa_base >> PAGE_SHIFT;
525 			gpa = gpa_base | (ea & (psize - 1));
526 			return kvmppc_hv_emulate_mmio(vcpu, gpa, ea,
527 						dsisr & DSISR_ISSTORE);
528 		}
529 	}
530 	index = vcpu->arch.pgfault_index;
531 	hptep = (__be64 *)(kvm->arch.hpt.virt + (index << 4));
532 	rev = &kvm->arch.hpt.rev[index];
533 	preempt_disable();
534 	while (!try_lock_hpte(hptep, HPTE_V_HVLOCK))
535 		cpu_relax();
536 	hpte[0] = be64_to_cpu(hptep[0]) & ~HPTE_V_HVLOCK;
537 	hpte[1] = be64_to_cpu(hptep[1]);
538 	hpte[2] = r = rev->guest_rpte;
539 	unlock_hpte(hptep, hpte[0]);
540 	preempt_enable();
541 
542 	if (cpu_has_feature(CPU_FTR_ARCH_300)) {
543 		hpte[0] = hpte_new_to_old_v(hpte[0], hpte[1]);
544 		hpte[1] = hpte_new_to_old_r(hpte[1]);
545 	}
546 	if (hpte[0] != vcpu->arch.pgfault_hpte[0] ||
547 	    hpte[1] != vcpu->arch.pgfault_hpte[1])
548 		return RESUME_GUEST;
549 
550 	/* Translate the logical address and get the page */
551 	psize = kvmppc_actual_pgsz(hpte[0], r);
552 	gpa_base = r & HPTE_R_RPN & ~(psize - 1);
553 	gfn_base = gpa_base >> PAGE_SHIFT;
554 	gpa = gpa_base | (ea & (psize - 1));
555 	gfn = gpa >> PAGE_SHIFT;
556 	memslot = gfn_to_memslot(kvm, gfn);
557 
558 	trace_kvm_page_fault_enter(vcpu, hpte, memslot, ea, dsisr);
559 
560 	/* No memslot means it's an emulated MMIO region */
561 	if (!memslot || (memslot->flags & KVM_MEMSLOT_INVALID))
562 		return kvmppc_hv_emulate_mmio(vcpu, gpa, ea,
563 					      dsisr & DSISR_ISSTORE);
564 
565 	/*
566 	 * This should never happen, because of the slot_is_aligned()
567 	 * check in kvmppc_do_h_enter().
568 	 */
569 	if (gfn_base < memslot->base_gfn)
570 		return -EFAULT;
571 
572 	/* used to check for invalidations in progress */
573 	mmu_seq = kvm->mmu_notifier_seq;
574 	smp_rmb();
575 
576 	ret = -EFAULT;
577 	page = NULL;
578 	writing = (dsisr & DSISR_ISSTORE) != 0;
579 	/* If writing != 0, then the HPTE must allow writing, if we get here */
580 	write_ok = writing;
581 	hva = gfn_to_hva_memslot(memslot, gfn);
582 
583 	/*
584 	 * Do a fast check first, since __gfn_to_pfn_memslot doesn't
585 	 * do it with !atomic && !async, which is how we call it.
586 	 * We always ask for write permission since the common case
587 	 * is that the page is writable.
588 	 */
589 	if (get_user_page_fast_only(hva, FOLL_WRITE, &page)) {
590 		write_ok = true;
591 	} else {
592 		/* Call KVM generic code to do the slow-path check */
593 		pfn = __gfn_to_pfn_memslot(memslot, gfn, false, NULL,
594 					   writing, &write_ok, NULL);
595 		if (is_error_noslot_pfn(pfn))
596 			return -EFAULT;
597 		page = NULL;
598 		if (pfn_valid(pfn)) {
599 			page = pfn_to_page(pfn);
600 			if (PageReserved(page))
601 				page = NULL;
602 		}
603 	}
604 
605 	/*
606 	 * Read the PTE from the process' radix tree and use that
607 	 * so we get the shift and attribute bits.
608 	 */
609 	spin_lock(&kvm->mmu_lock);
610 	ptep = find_kvm_host_pte(kvm, mmu_seq, hva, &shift);
611 	pte = __pte(0);
612 	if (ptep)
613 		pte = READ_ONCE(*ptep);
614 	spin_unlock(&kvm->mmu_lock);
615 	/*
616 	 * If the PTE disappeared temporarily due to a THP
617 	 * collapse, just return and let the guest try again.
618 	 */
619 	if (!pte_present(pte)) {
620 		if (page)
621 			put_page(page);
622 		return RESUME_GUEST;
623 	}
624 	hpa = pte_pfn(pte) << PAGE_SHIFT;
625 	pte_size = PAGE_SIZE;
626 	if (shift)
627 		pte_size = 1ul << shift;
628 	is_ci = pte_ci(pte);
629 
630 	if (psize > pte_size)
631 		goto out_put;
632 	if (pte_size > psize)
633 		hpa |= hva & (pte_size - psize);
634 
635 	/* Check WIMG vs. the actual page we're accessing */
636 	if (!hpte_cache_flags_ok(r, is_ci)) {
637 		if (is_ci)
638 			goto out_put;
639 		/*
640 		 * Allow guest to map emulated device memory as
641 		 * uncacheable, but actually make it cacheable.
642 		 */
643 		r = (r & ~(HPTE_R_W|HPTE_R_I|HPTE_R_G)) | HPTE_R_M;
644 	}
645 
646 	/*
647 	 * Set the HPTE to point to hpa.
648 	 * Since the hpa is at PAGE_SIZE granularity, make sure we
649 	 * don't mask out lower-order bits if psize < PAGE_SIZE.
650 	 */
651 	if (psize < PAGE_SIZE)
652 		psize = PAGE_SIZE;
653 	r = (r & HPTE_R_KEY_HI) | (r & ~(HPTE_R_PP0 - psize)) | hpa;
654 	if (hpte_is_writable(r) && !write_ok)
655 		r = hpte_make_readonly(r);
656 	ret = RESUME_GUEST;
657 	preempt_disable();
658 	while (!try_lock_hpte(hptep, HPTE_V_HVLOCK))
659 		cpu_relax();
660 	hnow_v = be64_to_cpu(hptep[0]);
661 	hnow_r = be64_to_cpu(hptep[1]);
662 	if (cpu_has_feature(CPU_FTR_ARCH_300)) {
663 		hnow_v = hpte_new_to_old_v(hnow_v, hnow_r);
664 		hnow_r = hpte_new_to_old_r(hnow_r);
665 	}
666 
667 	/*
668 	 * If the HPT is being resized, don't update the HPTE,
669 	 * instead let the guest retry after the resize operation is complete.
670 	 * The synchronization for mmu_ready test vs. set is provided
671 	 * by the HPTE lock.
672 	 */
673 	if (!kvm->arch.mmu_ready)
674 		goto out_unlock;
675 
676 	if ((hnow_v & ~HPTE_V_HVLOCK) != hpte[0] || hnow_r != hpte[1] ||
677 	    rev->guest_rpte != hpte[2])
678 		/* HPTE has been changed under us; let the guest retry */
679 		goto out_unlock;
680 	hpte[0] = (hpte[0] & ~HPTE_V_ABSENT) | HPTE_V_VALID;
681 
682 	/* Always put the HPTE in the rmap chain for the page base address */
683 	rmap = &memslot->arch.rmap[gfn_base - memslot->base_gfn];
684 	lock_rmap(rmap);
685 
686 	/* Check if we might have been invalidated; let the guest retry if so */
687 	ret = RESUME_GUEST;
688 	if (mmu_notifier_retry(vcpu->kvm, mmu_seq)) {
689 		unlock_rmap(rmap);
690 		goto out_unlock;
691 	}
692 
693 	/* Only set R/C in real HPTE if set in both *rmap and guest_rpte */
694 	rcbits = *rmap >> KVMPPC_RMAP_RC_SHIFT;
695 	r &= rcbits | ~(HPTE_R_R | HPTE_R_C);
696 
697 	if (be64_to_cpu(hptep[0]) & HPTE_V_VALID) {
698 		/* HPTE was previously valid, so we need to invalidate it */
699 		unlock_rmap(rmap);
700 		hptep[0] |= cpu_to_be64(HPTE_V_ABSENT);
701 		kvmppc_invalidate_hpte(kvm, hptep, index);
702 		/* don't lose previous R and C bits */
703 		r |= be64_to_cpu(hptep[1]) & (HPTE_R_R | HPTE_R_C);
704 	} else {
705 		kvmppc_add_revmap_chain(kvm, rev, rmap, index, 0);
706 	}
707 
708 	if (cpu_has_feature(CPU_FTR_ARCH_300)) {
709 		r = hpte_old_to_new_r(hpte[0], r);
710 		hpte[0] = hpte_old_to_new_v(hpte[0]);
711 	}
712 	hptep[1] = cpu_to_be64(r);
713 	eieio();
714 	__unlock_hpte(hptep, hpte[0]);
715 	asm volatile("ptesync" : : : "memory");
716 	preempt_enable();
717 	if (page && hpte_is_writable(r))
718 		set_page_dirty_lock(page);
719 
720  out_put:
721 	trace_kvm_page_fault_exit(vcpu, hpte, ret);
722 
723 	if (page)
724 		put_page(page);
725 	return ret;
726 
727  out_unlock:
728 	__unlock_hpte(hptep, be64_to_cpu(hptep[0]));
729 	preempt_enable();
730 	goto out_put;
731 }
732 
733 void kvmppc_rmap_reset(struct kvm *kvm)
734 {
735 	struct kvm_memslots *slots;
736 	struct kvm_memory_slot *memslot;
737 	int srcu_idx, bkt;
738 
739 	srcu_idx = srcu_read_lock(&kvm->srcu);
740 	slots = kvm_memslots(kvm);
741 	kvm_for_each_memslot(memslot, bkt, slots) {
742 		/* Mutual exclusion with kvm_unmap_hva_range etc. */
743 		spin_lock(&kvm->mmu_lock);
744 		/*
745 		 * This assumes it is acceptable to lose reference and
746 		 * change bits across a reset.
747 		 */
748 		memset(memslot->arch.rmap, 0,
749 		       memslot->npages * sizeof(*memslot->arch.rmap));
750 		spin_unlock(&kvm->mmu_lock);
751 	}
752 	srcu_read_unlock(&kvm->srcu, srcu_idx);
753 }
754 
755 /* Must be called with both HPTE and rmap locked */
756 static void kvmppc_unmap_hpte(struct kvm *kvm, unsigned long i,
757 			      struct kvm_memory_slot *memslot,
758 			      unsigned long *rmapp, unsigned long gfn)
759 {
760 	__be64 *hptep = (__be64 *) (kvm->arch.hpt.virt + (i << 4));
761 	struct revmap_entry *rev = kvm->arch.hpt.rev;
762 	unsigned long j, h;
763 	unsigned long ptel, psize, rcbits;
764 
765 	j = rev[i].forw;
766 	if (j == i) {
767 		/* chain is now empty */
768 		*rmapp &= ~(KVMPPC_RMAP_PRESENT | KVMPPC_RMAP_INDEX);
769 	} else {
770 		/* remove i from chain */
771 		h = rev[i].back;
772 		rev[h].forw = j;
773 		rev[j].back = h;
774 		rev[i].forw = rev[i].back = i;
775 		*rmapp = (*rmapp & ~KVMPPC_RMAP_INDEX) | j;
776 	}
777 
778 	/* Now check and modify the HPTE */
779 	ptel = rev[i].guest_rpte;
780 	psize = kvmppc_actual_pgsz(be64_to_cpu(hptep[0]), ptel);
781 	if ((be64_to_cpu(hptep[0]) & HPTE_V_VALID) &&
782 	    hpte_rpn(ptel, psize) == gfn) {
783 		hptep[0] |= cpu_to_be64(HPTE_V_ABSENT);
784 		kvmppc_invalidate_hpte(kvm, hptep, i);
785 		hptep[1] &= ~cpu_to_be64(HPTE_R_KEY_HI | HPTE_R_KEY_LO);
786 		/* Harvest R and C */
787 		rcbits = be64_to_cpu(hptep[1]) & (HPTE_R_R | HPTE_R_C);
788 		*rmapp |= rcbits << KVMPPC_RMAP_RC_SHIFT;
789 		if ((rcbits & HPTE_R_C) && memslot->dirty_bitmap)
790 			kvmppc_update_dirty_map(memslot, gfn, psize);
791 		if (rcbits & ~rev[i].guest_rpte) {
792 			rev[i].guest_rpte = ptel | rcbits;
793 			note_hpte_modification(kvm, &rev[i]);
794 		}
795 	}
796 }
797 
798 static void kvm_unmap_rmapp(struct kvm *kvm, struct kvm_memory_slot *memslot,
799 			    unsigned long gfn)
800 {
801 	unsigned long i;
802 	__be64 *hptep;
803 	unsigned long *rmapp;
804 
805 	rmapp = &memslot->arch.rmap[gfn - memslot->base_gfn];
806 	for (;;) {
807 		lock_rmap(rmapp);
808 		if (!(*rmapp & KVMPPC_RMAP_PRESENT)) {
809 			unlock_rmap(rmapp);
810 			break;
811 		}
812 
813 		/*
814 		 * To avoid an ABBA deadlock with the HPTE lock bit,
815 		 * we can't spin on the HPTE lock while holding the
816 		 * rmap chain lock.
817 		 */
818 		i = *rmapp & KVMPPC_RMAP_INDEX;
819 		hptep = (__be64 *) (kvm->arch.hpt.virt + (i << 4));
820 		if (!try_lock_hpte(hptep, HPTE_V_HVLOCK)) {
821 			/* unlock rmap before spinning on the HPTE lock */
822 			unlock_rmap(rmapp);
823 			while (be64_to_cpu(hptep[0]) & HPTE_V_HVLOCK)
824 				cpu_relax();
825 			continue;
826 		}
827 
828 		kvmppc_unmap_hpte(kvm, i, memslot, rmapp, gfn);
829 		unlock_rmap(rmapp);
830 		__unlock_hpte(hptep, be64_to_cpu(hptep[0]));
831 	}
832 }
833 
834 bool kvm_unmap_gfn_range_hv(struct kvm *kvm, struct kvm_gfn_range *range)
835 {
836 	gfn_t gfn;
837 
838 	if (kvm_is_radix(kvm)) {
839 		for (gfn = range->start; gfn < range->end; gfn++)
840 			kvm_unmap_radix(kvm, range->slot, gfn);
841 	} else {
842 		for (gfn = range->start; gfn < range->end; gfn++)
843 			kvm_unmap_rmapp(kvm, range->slot, gfn);
844 	}
845 
846 	return false;
847 }
848 
849 void kvmppc_core_flush_memslot_hv(struct kvm *kvm,
850 				  struct kvm_memory_slot *memslot)
851 {
852 	unsigned long gfn;
853 	unsigned long n;
854 	unsigned long *rmapp;
855 
856 	gfn = memslot->base_gfn;
857 	rmapp = memslot->arch.rmap;
858 	if (kvm_is_radix(kvm)) {
859 		kvmppc_radix_flush_memslot(kvm, memslot);
860 		return;
861 	}
862 
863 	for (n = memslot->npages; n; --n, ++gfn) {
864 		/*
865 		 * Testing the present bit without locking is OK because
866 		 * the memslot has been marked invalid already, and hence
867 		 * no new HPTEs referencing this page can be created,
868 		 * thus the present bit can't go from 0 to 1.
869 		 */
870 		if (*rmapp & KVMPPC_RMAP_PRESENT)
871 			kvm_unmap_rmapp(kvm, memslot, gfn);
872 		++rmapp;
873 	}
874 }
875 
876 static bool kvm_age_rmapp(struct kvm *kvm, struct kvm_memory_slot *memslot,
877 			  unsigned long gfn)
878 {
879 	struct revmap_entry *rev = kvm->arch.hpt.rev;
880 	unsigned long head, i, j;
881 	__be64 *hptep;
882 	int ret = 0;
883 	unsigned long *rmapp;
884 
885 	rmapp = &memslot->arch.rmap[gfn - memslot->base_gfn];
886  retry:
887 	lock_rmap(rmapp);
888 	if (*rmapp & KVMPPC_RMAP_REFERENCED) {
889 		*rmapp &= ~KVMPPC_RMAP_REFERENCED;
890 		ret = 1;
891 	}
892 	if (!(*rmapp & KVMPPC_RMAP_PRESENT)) {
893 		unlock_rmap(rmapp);
894 		return ret;
895 	}
896 
897 	i = head = *rmapp & KVMPPC_RMAP_INDEX;
898 	do {
899 		hptep = (__be64 *) (kvm->arch.hpt.virt + (i << 4));
900 		j = rev[i].forw;
901 
902 		/* If this HPTE isn't referenced, ignore it */
903 		if (!(be64_to_cpu(hptep[1]) & HPTE_R_R))
904 			continue;
905 
906 		if (!try_lock_hpte(hptep, HPTE_V_HVLOCK)) {
907 			/* unlock rmap before spinning on the HPTE lock */
908 			unlock_rmap(rmapp);
909 			while (be64_to_cpu(hptep[0]) & HPTE_V_HVLOCK)
910 				cpu_relax();
911 			goto retry;
912 		}
913 
914 		/* Now check and modify the HPTE */
915 		if ((be64_to_cpu(hptep[0]) & HPTE_V_VALID) &&
916 		    (be64_to_cpu(hptep[1]) & HPTE_R_R)) {
917 			kvmppc_clear_ref_hpte(kvm, hptep, i);
918 			if (!(rev[i].guest_rpte & HPTE_R_R)) {
919 				rev[i].guest_rpte |= HPTE_R_R;
920 				note_hpte_modification(kvm, &rev[i]);
921 			}
922 			ret = 1;
923 		}
924 		__unlock_hpte(hptep, be64_to_cpu(hptep[0]));
925 	} while ((i = j) != head);
926 
927 	unlock_rmap(rmapp);
928 	return ret;
929 }
930 
931 bool kvm_age_gfn_hv(struct kvm *kvm, struct kvm_gfn_range *range)
932 {
933 	gfn_t gfn;
934 	bool ret = false;
935 
936 	if (kvm_is_radix(kvm)) {
937 		for (gfn = range->start; gfn < range->end; gfn++)
938 			ret |= kvm_age_radix(kvm, range->slot, gfn);
939 	} else {
940 		for (gfn = range->start; gfn < range->end; gfn++)
941 			ret |= kvm_age_rmapp(kvm, range->slot, gfn);
942 	}
943 
944 	return ret;
945 }
946 
947 static bool kvm_test_age_rmapp(struct kvm *kvm, struct kvm_memory_slot *memslot,
948 			       unsigned long gfn)
949 {
950 	struct revmap_entry *rev = kvm->arch.hpt.rev;
951 	unsigned long head, i, j;
952 	unsigned long *hp;
953 	bool ret = true;
954 	unsigned long *rmapp;
955 
956 	rmapp = &memslot->arch.rmap[gfn - memslot->base_gfn];
957 	if (*rmapp & KVMPPC_RMAP_REFERENCED)
958 		return true;
959 
960 	lock_rmap(rmapp);
961 	if (*rmapp & KVMPPC_RMAP_REFERENCED)
962 		goto out;
963 
964 	if (*rmapp & KVMPPC_RMAP_PRESENT) {
965 		i = head = *rmapp & KVMPPC_RMAP_INDEX;
966 		do {
967 			hp = (unsigned long *)(kvm->arch.hpt.virt + (i << 4));
968 			j = rev[i].forw;
969 			if (be64_to_cpu(hp[1]) & HPTE_R_R)
970 				goto out;
971 		} while ((i = j) != head);
972 	}
973 	ret = false;
974 
975  out:
976 	unlock_rmap(rmapp);
977 	return ret;
978 }
979 
980 bool kvm_test_age_gfn_hv(struct kvm *kvm, struct kvm_gfn_range *range)
981 {
982 	WARN_ON(range->start + 1 != range->end);
983 
984 	if (kvm_is_radix(kvm))
985 		return kvm_test_age_radix(kvm, range->slot, range->start);
986 	else
987 		return kvm_test_age_rmapp(kvm, range->slot, range->start);
988 }
989 
990 bool kvm_set_spte_gfn_hv(struct kvm *kvm, struct kvm_gfn_range *range)
991 {
992 	WARN_ON(range->start + 1 != range->end);
993 
994 	if (kvm_is_radix(kvm))
995 		kvm_unmap_radix(kvm, range->slot, range->start);
996 	else
997 		kvm_unmap_rmapp(kvm, range->slot, range->start);
998 
999 	return false;
1000 }
1001 
1002 static int vcpus_running(struct kvm *kvm)
1003 {
1004 	return atomic_read(&kvm->arch.vcpus_running) != 0;
1005 }
1006 
1007 /*
1008  * Returns the number of system pages that are dirty.
1009  * This can be more than 1 if we find a huge-page HPTE.
1010  */
1011 static int kvm_test_clear_dirty_npages(struct kvm *kvm, unsigned long *rmapp)
1012 {
1013 	struct revmap_entry *rev = kvm->arch.hpt.rev;
1014 	unsigned long head, i, j;
1015 	unsigned long n;
1016 	unsigned long v, r;
1017 	__be64 *hptep;
1018 	int npages_dirty = 0;
1019 
1020  retry:
1021 	lock_rmap(rmapp);
1022 	if (!(*rmapp & KVMPPC_RMAP_PRESENT)) {
1023 		unlock_rmap(rmapp);
1024 		return npages_dirty;
1025 	}
1026 
1027 	i = head = *rmapp & KVMPPC_RMAP_INDEX;
1028 	do {
1029 		unsigned long hptep1;
1030 		hptep = (__be64 *) (kvm->arch.hpt.virt + (i << 4));
1031 		j = rev[i].forw;
1032 
1033 		/*
1034 		 * Checking the C (changed) bit here is racy since there
1035 		 * is no guarantee about when the hardware writes it back.
1036 		 * If the HPTE is not writable then it is stable since the
1037 		 * page can't be written to, and we would have done a tlbie
1038 		 * (which forces the hardware to complete any writeback)
1039 		 * when making the HPTE read-only.
1040 		 * If vcpus are running then this call is racy anyway
1041 		 * since the page could get dirtied subsequently, so we
1042 		 * expect there to be a further call which would pick up
1043 		 * any delayed C bit writeback.
1044 		 * Otherwise we need to do the tlbie even if C==0 in
1045 		 * order to pick up any delayed writeback of C.
1046 		 */
1047 		hptep1 = be64_to_cpu(hptep[1]);
1048 		if (!(hptep1 & HPTE_R_C) &&
1049 		    (!hpte_is_writable(hptep1) || vcpus_running(kvm)))
1050 			continue;
1051 
1052 		if (!try_lock_hpte(hptep, HPTE_V_HVLOCK)) {
1053 			/* unlock rmap before spinning on the HPTE lock */
1054 			unlock_rmap(rmapp);
1055 			while (hptep[0] & cpu_to_be64(HPTE_V_HVLOCK))
1056 				cpu_relax();
1057 			goto retry;
1058 		}
1059 
1060 		/* Now check and modify the HPTE */
1061 		if (!(hptep[0] & cpu_to_be64(HPTE_V_VALID))) {
1062 			__unlock_hpte(hptep, be64_to_cpu(hptep[0]));
1063 			continue;
1064 		}
1065 
1066 		/* need to make it temporarily absent so C is stable */
1067 		hptep[0] |= cpu_to_be64(HPTE_V_ABSENT);
1068 		kvmppc_invalidate_hpte(kvm, hptep, i);
1069 		v = be64_to_cpu(hptep[0]);
1070 		r = be64_to_cpu(hptep[1]);
1071 		if (r & HPTE_R_C) {
1072 			hptep[1] = cpu_to_be64(r & ~HPTE_R_C);
1073 			if (!(rev[i].guest_rpte & HPTE_R_C)) {
1074 				rev[i].guest_rpte |= HPTE_R_C;
1075 				note_hpte_modification(kvm, &rev[i]);
1076 			}
1077 			n = kvmppc_actual_pgsz(v, r);
1078 			n = (n + PAGE_SIZE - 1) >> PAGE_SHIFT;
1079 			if (n > npages_dirty)
1080 				npages_dirty = n;
1081 			eieio();
1082 		}
1083 		v &= ~HPTE_V_ABSENT;
1084 		v |= HPTE_V_VALID;
1085 		__unlock_hpte(hptep, v);
1086 	} while ((i = j) != head);
1087 
1088 	unlock_rmap(rmapp);
1089 	return npages_dirty;
1090 }
1091 
1092 void kvmppc_harvest_vpa_dirty(struct kvmppc_vpa *vpa,
1093 			      struct kvm_memory_slot *memslot,
1094 			      unsigned long *map)
1095 {
1096 	unsigned long gfn;
1097 
1098 	if (!vpa->dirty || !vpa->pinned_addr)
1099 		return;
1100 	gfn = vpa->gpa >> PAGE_SHIFT;
1101 	if (gfn < memslot->base_gfn ||
1102 	    gfn >= memslot->base_gfn + memslot->npages)
1103 		return;
1104 
1105 	vpa->dirty = false;
1106 	if (map)
1107 		__set_bit_le(gfn - memslot->base_gfn, map);
1108 }
1109 
1110 long kvmppc_hv_get_dirty_log_hpt(struct kvm *kvm,
1111 			struct kvm_memory_slot *memslot, unsigned long *map)
1112 {
1113 	unsigned long i;
1114 	unsigned long *rmapp;
1115 
1116 	preempt_disable();
1117 	rmapp = memslot->arch.rmap;
1118 	for (i = 0; i < memslot->npages; ++i) {
1119 		int npages = kvm_test_clear_dirty_npages(kvm, rmapp);
1120 		/*
1121 		 * Note that if npages > 0 then i must be a multiple of npages,
1122 		 * since we always put huge-page HPTEs in the rmap chain
1123 		 * corresponding to their page base address.
1124 		 */
1125 		if (npages)
1126 			set_dirty_bits(map, i, npages);
1127 		++rmapp;
1128 	}
1129 	preempt_enable();
1130 	return 0;
1131 }
1132 
1133 void *kvmppc_pin_guest_page(struct kvm *kvm, unsigned long gpa,
1134 			    unsigned long *nb_ret)
1135 {
1136 	struct kvm_memory_slot *memslot;
1137 	unsigned long gfn = gpa >> PAGE_SHIFT;
1138 	struct page *page, *pages[1];
1139 	int npages;
1140 	unsigned long hva, offset;
1141 	int srcu_idx;
1142 
1143 	srcu_idx = srcu_read_lock(&kvm->srcu);
1144 	memslot = gfn_to_memslot(kvm, gfn);
1145 	if (!memslot || (memslot->flags & KVM_MEMSLOT_INVALID))
1146 		goto err;
1147 	hva = gfn_to_hva_memslot(memslot, gfn);
1148 	npages = get_user_pages_fast(hva, 1, FOLL_WRITE, pages);
1149 	if (npages < 1)
1150 		goto err;
1151 	page = pages[0];
1152 	srcu_read_unlock(&kvm->srcu, srcu_idx);
1153 
1154 	offset = gpa & (PAGE_SIZE - 1);
1155 	if (nb_ret)
1156 		*nb_ret = PAGE_SIZE - offset;
1157 	return page_address(page) + offset;
1158 
1159  err:
1160 	srcu_read_unlock(&kvm->srcu, srcu_idx);
1161 	return NULL;
1162 }
1163 
1164 void kvmppc_unpin_guest_page(struct kvm *kvm, void *va, unsigned long gpa,
1165 			     bool dirty)
1166 {
1167 	struct page *page = virt_to_page(va);
1168 	struct kvm_memory_slot *memslot;
1169 	unsigned long gfn;
1170 	int srcu_idx;
1171 
1172 	put_page(page);
1173 
1174 	if (!dirty)
1175 		return;
1176 
1177 	/* We need to mark this page dirty in the memslot dirty_bitmap, if any */
1178 	gfn = gpa >> PAGE_SHIFT;
1179 	srcu_idx = srcu_read_lock(&kvm->srcu);
1180 	memslot = gfn_to_memslot(kvm, gfn);
1181 	if (memslot && memslot->dirty_bitmap)
1182 		set_bit_le(gfn - memslot->base_gfn, memslot->dirty_bitmap);
1183 	srcu_read_unlock(&kvm->srcu, srcu_idx);
1184 }
1185 
1186 /*
1187  * HPT resizing
1188  */
1189 static int resize_hpt_allocate(struct kvm_resize_hpt *resize)
1190 {
1191 	int rc;
1192 
1193 	rc = kvmppc_allocate_hpt(&resize->hpt, resize->order);
1194 	if (rc < 0)
1195 		return rc;
1196 
1197 	resize_hpt_debug(resize, "resize_hpt_allocate(): HPT @ 0x%lx\n",
1198 			 resize->hpt.virt);
1199 
1200 	return 0;
1201 }
1202 
1203 static unsigned long resize_hpt_rehash_hpte(struct kvm_resize_hpt *resize,
1204 					    unsigned long idx)
1205 {
1206 	struct kvm *kvm = resize->kvm;
1207 	struct kvm_hpt_info *old = &kvm->arch.hpt;
1208 	struct kvm_hpt_info *new = &resize->hpt;
1209 	unsigned long old_hash_mask = (1ULL << (old->order - 7)) - 1;
1210 	unsigned long new_hash_mask = (1ULL << (new->order - 7)) - 1;
1211 	__be64 *hptep, *new_hptep;
1212 	unsigned long vpte, rpte, guest_rpte;
1213 	int ret;
1214 	struct revmap_entry *rev;
1215 	unsigned long apsize, avpn, pteg, hash;
1216 	unsigned long new_idx, new_pteg, replace_vpte;
1217 	int pshift;
1218 
1219 	hptep = (__be64 *)(old->virt + (idx << 4));
1220 
1221 	/* Guest is stopped, so new HPTEs can't be added or faulted
1222 	 * in, only unmapped or altered by host actions.  So, it's
1223 	 * safe to check this before we take the HPTE lock */
1224 	vpte = be64_to_cpu(hptep[0]);
1225 	if (!(vpte & HPTE_V_VALID) && !(vpte & HPTE_V_ABSENT))
1226 		return 0; /* nothing to do */
1227 
1228 	while (!try_lock_hpte(hptep, HPTE_V_HVLOCK))
1229 		cpu_relax();
1230 
1231 	vpte = be64_to_cpu(hptep[0]);
1232 
1233 	ret = 0;
1234 	if (!(vpte & HPTE_V_VALID) && !(vpte & HPTE_V_ABSENT))
1235 		/* Nothing to do */
1236 		goto out;
1237 
1238 	if (cpu_has_feature(CPU_FTR_ARCH_300)) {
1239 		rpte = be64_to_cpu(hptep[1]);
1240 		vpte = hpte_new_to_old_v(vpte, rpte);
1241 	}
1242 
1243 	/* Unmap */
1244 	rev = &old->rev[idx];
1245 	guest_rpte = rev->guest_rpte;
1246 
1247 	ret = -EIO;
1248 	apsize = kvmppc_actual_pgsz(vpte, guest_rpte);
1249 	if (!apsize)
1250 		goto out;
1251 
1252 	if (vpte & HPTE_V_VALID) {
1253 		unsigned long gfn = hpte_rpn(guest_rpte, apsize);
1254 		int srcu_idx = srcu_read_lock(&kvm->srcu);
1255 		struct kvm_memory_slot *memslot =
1256 			__gfn_to_memslot(kvm_memslots(kvm), gfn);
1257 
1258 		if (memslot) {
1259 			unsigned long *rmapp;
1260 			rmapp = &memslot->arch.rmap[gfn - memslot->base_gfn];
1261 
1262 			lock_rmap(rmapp);
1263 			kvmppc_unmap_hpte(kvm, idx, memslot, rmapp, gfn);
1264 			unlock_rmap(rmapp);
1265 		}
1266 
1267 		srcu_read_unlock(&kvm->srcu, srcu_idx);
1268 	}
1269 
1270 	/* Reload PTE after unmap */
1271 	vpte = be64_to_cpu(hptep[0]);
1272 	BUG_ON(vpte & HPTE_V_VALID);
1273 	BUG_ON(!(vpte & HPTE_V_ABSENT));
1274 
1275 	ret = 0;
1276 	if (!(vpte & HPTE_V_BOLTED))
1277 		goto out;
1278 
1279 	rpte = be64_to_cpu(hptep[1]);
1280 
1281 	if (cpu_has_feature(CPU_FTR_ARCH_300)) {
1282 		vpte = hpte_new_to_old_v(vpte, rpte);
1283 		rpte = hpte_new_to_old_r(rpte);
1284 	}
1285 
1286 	pshift = kvmppc_hpte_base_page_shift(vpte, rpte);
1287 	avpn = HPTE_V_AVPN_VAL(vpte) & ~(((1ul << pshift) - 1) >> 23);
1288 	pteg = idx / HPTES_PER_GROUP;
1289 	if (vpte & HPTE_V_SECONDARY)
1290 		pteg = ~pteg;
1291 
1292 	if (!(vpte & HPTE_V_1TB_SEG)) {
1293 		unsigned long offset, vsid;
1294 
1295 		/* We only have 28 - 23 bits of offset in avpn */
1296 		offset = (avpn & 0x1f) << 23;
1297 		vsid = avpn >> 5;
1298 		/* We can find more bits from the pteg value */
1299 		if (pshift < 23)
1300 			offset |= ((vsid ^ pteg) & old_hash_mask) << pshift;
1301 
1302 		hash = vsid ^ (offset >> pshift);
1303 	} else {
1304 		unsigned long offset, vsid;
1305 
1306 		/* We only have 40 - 23 bits of seg_off in avpn */
1307 		offset = (avpn & 0x1ffff) << 23;
1308 		vsid = avpn >> 17;
1309 		if (pshift < 23)
1310 			offset |= ((vsid ^ (vsid << 25) ^ pteg) & old_hash_mask) << pshift;
1311 
1312 		hash = vsid ^ (vsid << 25) ^ (offset >> pshift);
1313 	}
1314 
1315 	new_pteg = hash & new_hash_mask;
1316 	if (vpte & HPTE_V_SECONDARY)
1317 		new_pteg = ~hash & new_hash_mask;
1318 
1319 	new_idx = new_pteg * HPTES_PER_GROUP + (idx % HPTES_PER_GROUP);
1320 	new_hptep = (__be64 *)(new->virt + (new_idx << 4));
1321 
1322 	replace_vpte = be64_to_cpu(new_hptep[0]);
1323 	if (cpu_has_feature(CPU_FTR_ARCH_300)) {
1324 		unsigned long replace_rpte = be64_to_cpu(new_hptep[1]);
1325 		replace_vpte = hpte_new_to_old_v(replace_vpte, replace_rpte);
1326 	}
1327 
1328 	if (replace_vpte & (HPTE_V_VALID | HPTE_V_ABSENT)) {
1329 		BUG_ON(new->order >= old->order);
1330 
1331 		if (replace_vpte & HPTE_V_BOLTED) {
1332 			if (vpte & HPTE_V_BOLTED)
1333 				/* Bolted collision, nothing we can do */
1334 				ret = -ENOSPC;
1335 			/* Discard the new HPTE */
1336 			goto out;
1337 		}
1338 
1339 		/* Discard the previous HPTE */
1340 	}
1341 
1342 	if (cpu_has_feature(CPU_FTR_ARCH_300)) {
1343 		rpte = hpte_old_to_new_r(vpte, rpte);
1344 		vpte = hpte_old_to_new_v(vpte);
1345 	}
1346 
1347 	new_hptep[1] = cpu_to_be64(rpte);
1348 	new->rev[new_idx].guest_rpte = guest_rpte;
1349 	/* No need for a barrier, since new HPT isn't active */
1350 	new_hptep[0] = cpu_to_be64(vpte);
1351 	unlock_hpte(new_hptep, vpte);
1352 
1353 out:
1354 	unlock_hpte(hptep, vpte);
1355 	return ret;
1356 }
1357 
1358 static int resize_hpt_rehash(struct kvm_resize_hpt *resize)
1359 {
1360 	struct kvm *kvm = resize->kvm;
1361 	unsigned  long i;
1362 	int rc;
1363 
1364 	for (i = 0; i < kvmppc_hpt_npte(&kvm->arch.hpt); i++) {
1365 		rc = resize_hpt_rehash_hpte(resize, i);
1366 		if (rc != 0)
1367 			return rc;
1368 	}
1369 
1370 	return 0;
1371 }
1372 
1373 static void resize_hpt_pivot(struct kvm_resize_hpt *resize)
1374 {
1375 	struct kvm *kvm = resize->kvm;
1376 	struct kvm_hpt_info hpt_tmp;
1377 
1378 	/* Exchange the pending tables in the resize structure with
1379 	 * the active tables */
1380 
1381 	resize_hpt_debug(resize, "resize_hpt_pivot()\n");
1382 
1383 	spin_lock(&kvm->mmu_lock);
1384 	asm volatile("ptesync" : : : "memory");
1385 
1386 	hpt_tmp = kvm->arch.hpt;
1387 	kvmppc_set_hpt(kvm, &resize->hpt);
1388 	resize->hpt = hpt_tmp;
1389 
1390 	spin_unlock(&kvm->mmu_lock);
1391 
1392 	synchronize_srcu_expedited(&kvm->srcu);
1393 
1394 	if (cpu_has_feature(CPU_FTR_ARCH_300))
1395 		kvmppc_setup_partition_table(kvm);
1396 
1397 	resize_hpt_debug(resize, "resize_hpt_pivot() done\n");
1398 }
1399 
1400 static void resize_hpt_release(struct kvm *kvm, struct kvm_resize_hpt *resize)
1401 {
1402 	if (WARN_ON(!mutex_is_locked(&kvm->arch.mmu_setup_lock)))
1403 		return;
1404 
1405 	if (!resize)
1406 		return;
1407 
1408 	if (resize->error != -EBUSY) {
1409 		if (resize->hpt.virt)
1410 			kvmppc_free_hpt(&resize->hpt);
1411 		kfree(resize);
1412 	}
1413 
1414 	if (kvm->arch.resize_hpt == resize)
1415 		kvm->arch.resize_hpt = NULL;
1416 }
1417 
1418 static void resize_hpt_prepare_work(struct work_struct *work)
1419 {
1420 	struct kvm_resize_hpt *resize = container_of(work,
1421 						     struct kvm_resize_hpt,
1422 						     work);
1423 	struct kvm *kvm = resize->kvm;
1424 	int err = 0;
1425 
1426 	if (WARN_ON(resize->error != -EBUSY))
1427 		return;
1428 
1429 	mutex_lock(&kvm->arch.mmu_setup_lock);
1430 
1431 	/* Request is still current? */
1432 	if (kvm->arch.resize_hpt == resize) {
1433 		/* We may request large allocations here:
1434 		 * do not sleep with kvm->arch.mmu_setup_lock held for a while.
1435 		 */
1436 		mutex_unlock(&kvm->arch.mmu_setup_lock);
1437 
1438 		resize_hpt_debug(resize, "resize_hpt_prepare_work(): order = %d\n",
1439 				 resize->order);
1440 
1441 		err = resize_hpt_allocate(resize);
1442 
1443 		/* We have strict assumption about -EBUSY
1444 		 * when preparing for HPT resize.
1445 		 */
1446 		if (WARN_ON(err == -EBUSY))
1447 			err = -EINPROGRESS;
1448 
1449 		mutex_lock(&kvm->arch.mmu_setup_lock);
1450 		/* It is possible that kvm->arch.resize_hpt != resize
1451 		 * after we grab kvm->arch.mmu_setup_lock again.
1452 		 */
1453 	}
1454 
1455 	resize->error = err;
1456 
1457 	if (kvm->arch.resize_hpt != resize)
1458 		resize_hpt_release(kvm, resize);
1459 
1460 	mutex_unlock(&kvm->arch.mmu_setup_lock);
1461 }
1462 
1463 long kvm_vm_ioctl_resize_hpt_prepare(struct kvm *kvm,
1464 				     struct kvm_ppc_resize_hpt *rhpt)
1465 {
1466 	unsigned long flags = rhpt->flags;
1467 	unsigned long shift = rhpt->shift;
1468 	struct kvm_resize_hpt *resize;
1469 	int ret;
1470 
1471 	if (flags != 0 || kvm_is_radix(kvm))
1472 		return -EINVAL;
1473 
1474 	if (shift && ((shift < 18) || (shift > 46)))
1475 		return -EINVAL;
1476 
1477 	mutex_lock(&kvm->arch.mmu_setup_lock);
1478 
1479 	resize = kvm->arch.resize_hpt;
1480 
1481 	if (resize) {
1482 		if (resize->order == shift) {
1483 			/* Suitable resize in progress? */
1484 			ret = resize->error;
1485 			if (ret == -EBUSY)
1486 				ret = 100; /* estimated time in ms */
1487 			else if (ret)
1488 				resize_hpt_release(kvm, resize);
1489 
1490 			goto out;
1491 		}
1492 
1493 		/* not suitable, cancel it */
1494 		resize_hpt_release(kvm, resize);
1495 	}
1496 
1497 	ret = 0;
1498 	if (!shift)
1499 		goto out; /* nothing to do */
1500 
1501 	/* start new resize */
1502 
1503 	resize = kzalloc(sizeof(*resize), GFP_KERNEL);
1504 	if (!resize) {
1505 		ret = -ENOMEM;
1506 		goto out;
1507 	}
1508 
1509 	resize->error = -EBUSY;
1510 	resize->order = shift;
1511 	resize->kvm = kvm;
1512 	INIT_WORK(&resize->work, resize_hpt_prepare_work);
1513 	kvm->arch.resize_hpt = resize;
1514 
1515 	schedule_work(&resize->work);
1516 
1517 	ret = 100; /* estimated time in ms */
1518 
1519 out:
1520 	mutex_unlock(&kvm->arch.mmu_setup_lock);
1521 	return ret;
1522 }
1523 
1524 static void resize_hpt_boot_vcpu(void *opaque)
1525 {
1526 	/* Nothing to do, just force a KVM exit */
1527 }
1528 
1529 long kvm_vm_ioctl_resize_hpt_commit(struct kvm *kvm,
1530 				    struct kvm_ppc_resize_hpt *rhpt)
1531 {
1532 	unsigned long flags = rhpt->flags;
1533 	unsigned long shift = rhpt->shift;
1534 	struct kvm_resize_hpt *resize;
1535 	long ret;
1536 
1537 	if (flags != 0 || kvm_is_radix(kvm))
1538 		return -EINVAL;
1539 
1540 	if (shift && ((shift < 18) || (shift > 46)))
1541 		return -EINVAL;
1542 
1543 	mutex_lock(&kvm->arch.mmu_setup_lock);
1544 
1545 	resize = kvm->arch.resize_hpt;
1546 
1547 	/* This shouldn't be possible */
1548 	ret = -EIO;
1549 	if (WARN_ON(!kvm->arch.mmu_ready))
1550 		goto out_no_hpt;
1551 
1552 	/* Stop VCPUs from running while we mess with the HPT */
1553 	kvm->arch.mmu_ready = 0;
1554 	smp_mb();
1555 
1556 	/* Boot all CPUs out of the guest so they re-read
1557 	 * mmu_ready */
1558 	on_each_cpu(resize_hpt_boot_vcpu, NULL, 1);
1559 
1560 	ret = -ENXIO;
1561 	if (!resize || (resize->order != shift))
1562 		goto out;
1563 
1564 	ret = resize->error;
1565 	if (ret)
1566 		goto out;
1567 
1568 	ret = resize_hpt_rehash(resize);
1569 	if (ret)
1570 		goto out;
1571 
1572 	resize_hpt_pivot(resize);
1573 
1574 out:
1575 	/* Let VCPUs run again */
1576 	kvm->arch.mmu_ready = 1;
1577 	smp_mb();
1578 out_no_hpt:
1579 	resize_hpt_release(kvm, resize);
1580 	mutex_unlock(&kvm->arch.mmu_setup_lock);
1581 	return ret;
1582 }
1583 
1584 /*
1585  * Functions for reading and writing the hash table via reads and
1586  * writes on a file descriptor.
1587  *
1588  * Reads return the guest view of the hash table, which has to be
1589  * pieced together from the real hash table and the guest_rpte
1590  * values in the revmap array.
1591  *
1592  * On writes, each HPTE written is considered in turn, and if it
1593  * is valid, it is written to the HPT as if an H_ENTER with the
1594  * exact flag set was done.  When the invalid count is non-zero
1595  * in the header written to the stream, the kernel will make
1596  * sure that that many HPTEs are invalid, and invalidate them
1597  * if not.
1598  */
1599 
1600 struct kvm_htab_ctx {
1601 	unsigned long	index;
1602 	unsigned long	flags;
1603 	struct kvm	*kvm;
1604 	int		first_pass;
1605 };
1606 
1607 #define HPTE_SIZE	(2 * sizeof(unsigned long))
1608 
1609 /*
1610  * Returns 1 if this HPT entry has been modified or has pending
1611  * R/C bit changes.
1612  */
1613 static int hpte_dirty(struct revmap_entry *revp, __be64 *hptp)
1614 {
1615 	unsigned long rcbits_unset;
1616 
1617 	if (revp->guest_rpte & HPTE_GR_MODIFIED)
1618 		return 1;
1619 
1620 	/* Also need to consider changes in reference and changed bits */
1621 	rcbits_unset = ~revp->guest_rpte & (HPTE_R_R | HPTE_R_C);
1622 	if ((be64_to_cpu(hptp[0]) & HPTE_V_VALID) &&
1623 	    (be64_to_cpu(hptp[1]) & rcbits_unset))
1624 		return 1;
1625 
1626 	return 0;
1627 }
1628 
1629 static long record_hpte(unsigned long flags, __be64 *hptp,
1630 			unsigned long *hpte, struct revmap_entry *revp,
1631 			int want_valid, int first_pass)
1632 {
1633 	unsigned long v, r, hr;
1634 	unsigned long rcbits_unset;
1635 	int ok = 1;
1636 	int valid, dirty;
1637 
1638 	/* Unmodified entries are uninteresting except on the first pass */
1639 	dirty = hpte_dirty(revp, hptp);
1640 	if (!first_pass && !dirty)
1641 		return 0;
1642 
1643 	valid = 0;
1644 	if (be64_to_cpu(hptp[0]) & (HPTE_V_VALID | HPTE_V_ABSENT)) {
1645 		valid = 1;
1646 		if ((flags & KVM_GET_HTAB_BOLTED_ONLY) &&
1647 		    !(be64_to_cpu(hptp[0]) & HPTE_V_BOLTED))
1648 			valid = 0;
1649 	}
1650 	if (valid != want_valid)
1651 		return 0;
1652 
1653 	v = r = 0;
1654 	if (valid || dirty) {
1655 		/* lock the HPTE so it's stable and read it */
1656 		preempt_disable();
1657 		while (!try_lock_hpte(hptp, HPTE_V_HVLOCK))
1658 			cpu_relax();
1659 		v = be64_to_cpu(hptp[0]);
1660 		hr = be64_to_cpu(hptp[1]);
1661 		if (cpu_has_feature(CPU_FTR_ARCH_300)) {
1662 			v = hpte_new_to_old_v(v, hr);
1663 			hr = hpte_new_to_old_r(hr);
1664 		}
1665 
1666 		/* re-evaluate valid and dirty from synchronized HPTE value */
1667 		valid = !!(v & HPTE_V_VALID);
1668 		dirty = !!(revp->guest_rpte & HPTE_GR_MODIFIED);
1669 
1670 		/* Harvest R and C into guest view if necessary */
1671 		rcbits_unset = ~revp->guest_rpte & (HPTE_R_R | HPTE_R_C);
1672 		if (valid && (rcbits_unset & hr)) {
1673 			revp->guest_rpte |= (hr &
1674 				(HPTE_R_R | HPTE_R_C)) | HPTE_GR_MODIFIED;
1675 			dirty = 1;
1676 		}
1677 
1678 		if (v & HPTE_V_ABSENT) {
1679 			v &= ~HPTE_V_ABSENT;
1680 			v |= HPTE_V_VALID;
1681 			valid = 1;
1682 		}
1683 		if ((flags & KVM_GET_HTAB_BOLTED_ONLY) && !(v & HPTE_V_BOLTED))
1684 			valid = 0;
1685 
1686 		r = revp->guest_rpte;
1687 		/* only clear modified if this is the right sort of entry */
1688 		if (valid == want_valid && dirty) {
1689 			r &= ~HPTE_GR_MODIFIED;
1690 			revp->guest_rpte = r;
1691 		}
1692 		unlock_hpte(hptp, be64_to_cpu(hptp[0]));
1693 		preempt_enable();
1694 		if (!(valid == want_valid && (first_pass || dirty)))
1695 			ok = 0;
1696 	}
1697 	hpte[0] = cpu_to_be64(v);
1698 	hpte[1] = cpu_to_be64(r);
1699 	return ok;
1700 }
1701 
1702 static ssize_t kvm_htab_read(struct file *file, char __user *buf,
1703 			     size_t count, loff_t *ppos)
1704 {
1705 	struct kvm_htab_ctx *ctx = file->private_data;
1706 	struct kvm *kvm = ctx->kvm;
1707 	struct kvm_get_htab_header hdr;
1708 	__be64 *hptp;
1709 	struct revmap_entry *revp;
1710 	unsigned long i, nb, nw;
1711 	unsigned long __user *lbuf;
1712 	struct kvm_get_htab_header __user *hptr;
1713 	unsigned long flags;
1714 	int first_pass;
1715 	unsigned long hpte[2];
1716 
1717 	if (!access_ok(buf, count))
1718 		return -EFAULT;
1719 	if (kvm_is_radix(kvm))
1720 		return 0;
1721 
1722 	first_pass = ctx->first_pass;
1723 	flags = ctx->flags;
1724 
1725 	i = ctx->index;
1726 	hptp = (__be64 *)(kvm->arch.hpt.virt + (i * HPTE_SIZE));
1727 	revp = kvm->arch.hpt.rev + i;
1728 	lbuf = (unsigned long __user *)buf;
1729 
1730 	nb = 0;
1731 	while (nb + sizeof(hdr) + HPTE_SIZE < count) {
1732 		/* Initialize header */
1733 		hptr = (struct kvm_get_htab_header __user *)buf;
1734 		hdr.n_valid = 0;
1735 		hdr.n_invalid = 0;
1736 		nw = nb;
1737 		nb += sizeof(hdr);
1738 		lbuf = (unsigned long __user *)(buf + sizeof(hdr));
1739 
1740 		/* Skip uninteresting entries, i.e. clean on not-first pass */
1741 		if (!first_pass) {
1742 			while (i < kvmppc_hpt_npte(&kvm->arch.hpt) &&
1743 			       !hpte_dirty(revp, hptp)) {
1744 				++i;
1745 				hptp += 2;
1746 				++revp;
1747 			}
1748 		}
1749 		hdr.index = i;
1750 
1751 		/* Grab a series of valid entries */
1752 		while (i < kvmppc_hpt_npte(&kvm->arch.hpt) &&
1753 		       hdr.n_valid < 0xffff &&
1754 		       nb + HPTE_SIZE < count &&
1755 		       record_hpte(flags, hptp, hpte, revp, 1, first_pass)) {
1756 			/* valid entry, write it out */
1757 			++hdr.n_valid;
1758 			if (__put_user(hpte[0], lbuf) ||
1759 			    __put_user(hpte[1], lbuf + 1))
1760 				return -EFAULT;
1761 			nb += HPTE_SIZE;
1762 			lbuf += 2;
1763 			++i;
1764 			hptp += 2;
1765 			++revp;
1766 		}
1767 		/* Now skip invalid entries while we can */
1768 		while (i < kvmppc_hpt_npte(&kvm->arch.hpt) &&
1769 		       hdr.n_invalid < 0xffff &&
1770 		       record_hpte(flags, hptp, hpte, revp, 0, first_pass)) {
1771 			/* found an invalid entry */
1772 			++hdr.n_invalid;
1773 			++i;
1774 			hptp += 2;
1775 			++revp;
1776 		}
1777 
1778 		if (hdr.n_valid || hdr.n_invalid) {
1779 			/* write back the header */
1780 			if (__copy_to_user(hptr, &hdr, sizeof(hdr)))
1781 				return -EFAULT;
1782 			nw = nb;
1783 			buf = (char __user *)lbuf;
1784 		} else {
1785 			nb = nw;
1786 		}
1787 
1788 		/* Check if we've wrapped around the hash table */
1789 		if (i >= kvmppc_hpt_npte(&kvm->arch.hpt)) {
1790 			i = 0;
1791 			ctx->first_pass = 0;
1792 			break;
1793 		}
1794 	}
1795 
1796 	ctx->index = i;
1797 
1798 	return nb;
1799 }
1800 
1801 static ssize_t kvm_htab_write(struct file *file, const char __user *buf,
1802 			      size_t count, loff_t *ppos)
1803 {
1804 	struct kvm_htab_ctx *ctx = file->private_data;
1805 	struct kvm *kvm = ctx->kvm;
1806 	struct kvm_get_htab_header hdr;
1807 	unsigned long i, j;
1808 	unsigned long v, r;
1809 	unsigned long __user *lbuf;
1810 	__be64 *hptp;
1811 	unsigned long tmp[2];
1812 	ssize_t nb;
1813 	long int err, ret;
1814 	int mmu_ready;
1815 	int pshift;
1816 
1817 	if (!access_ok(buf, count))
1818 		return -EFAULT;
1819 	if (kvm_is_radix(kvm))
1820 		return -EINVAL;
1821 
1822 	/* lock out vcpus from running while we're doing this */
1823 	mutex_lock(&kvm->arch.mmu_setup_lock);
1824 	mmu_ready = kvm->arch.mmu_ready;
1825 	if (mmu_ready) {
1826 		kvm->arch.mmu_ready = 0;	/* temporarily */
1827 		/* order mmu_ready vs. vcpus_running */
1828 		smp_mb();
1829 		if (atomic_read(&kvm->arch.vcpus_running)) {
1830 			kvm->arch.mmu_ready = 1;
1831 			mutex_unlock(&kvm->arch.mmu_setup_lock);
1832 			return -EBUSY;
1833 		}
1834 	}
1835 
1836 	err = 0;
1837 	for (nb = 0; nb + sizeof(hdr) <= count; ) {
1838 		err = -EFAULT;
1839 		if (__copy_from_user(&hdr, buf, sizeof(hdr)))
1840 			break;
1841 
1842 		err = 0;
1843 		if (nb + hdr.n_valid * HPTE_SIZE > count)
1844 			break;
1845 
1846 		nb += sizeof(hdr);
1847 		buf += sizeof(hdr);
1848 
1849 		err = -EINVAL;
1850 		i = hdr.index;
1851 		if (i >= kvmppc_hpt_npte(&kvm->arch.hpt) ||
1852 		    i + hdr.n_valid + hdr.n_invalid > kvmppc_hpt_npte(&kvm->arch.hpt))
1853 			break;
1854 
1855 		hptp = (__be64 *)(kvm->arch.hpt.virt + (i * HPTE_SIZE));
1856 		lbuf = (unsigned long __user *)buf;
1857 		for (j = 0; j < hdr.n_valid; ++j) {
1858 			__be64 hpte_v;
1859 			__be64 hpte_r;
1860 
1861 			err = -EFAULT;
1862 			if (__get_user(hpte_v, lbuf) ||
1863 			    __get_user(hpte_r, lbuf + 1))
1864 				goto out;
1865 			v = be64_to_cpu(hpte_v);
1866 			r = be64_to_cpu(hpte_r);
1867 			err = -EINVAL;
1868 			if (!(v & HPTE_V_VALID))
1869 				goto out;
1870 			pshift = kvmppc_hpte_base_page_shift(v, r);
1871 			if (pshift <= 0)
1872 				goto out;
1873 			lbuf += 2;
1874 			nb += HPTE_SIZE;
1875 
1876 			if (be64_to_cpu(hptp[0]) & (HPTE_V_VALID | HPTE_V_ABSENT))
1877 				kvmppc_do_h_remove(kvm, 0, i, 0, tmp);
1878 			err = -EIO;
1879 			ret = kvmppc_virtmode_do_h_enter(kvm, H_EXACT, i, v, r,
1880 							 tmp);
1881 			if (ret != H_SUCCESS) {
1882 				pr_err("kvm_htab_write ret %ld i=%ld v=%lx "
1883 				       "r=%lx\n", ret, i, v, r);
1884 				goto out;
1885 			}
1886 			if (!mmu_ready && is_vrma_hpte(v)) {
1887 				unsigned long senc, lpcr;
1888 
1889 				senc = slb_pgsize_encoding(1ul << pshift);
1890 				kvm->arch.vrma_slb_v = senc | SLB_VSID_B_1T |
1891 					(VRMA_VSID << SLB_VSID_SHIFT_1T);
1892 				if (!cpu_has_feature(CPU_FTR_ARCH_300)) {
1893 					lpcr = senc << (LPCR_VRMASD_SH - 4);
1894 					kvmppc_update_lpcr(kvm, lpcr,
1895 							   LPCR_VRMASD);
1896 				} else {
1897 					kvmppc_setup_partition_table(kvm);
1898 				}
1899 				mmu_ready = 1;
1900 			}
1901 			++i;
1902 			hptp += 2;
1903 		}
1904 
1905 		for (j = 0; j < hdr.n_invalid; ++j) {
1906 			if (be64_to_cpu(hptp[0]) & (HPTE_V_VALID | HPTE_V_ABSENT))
1907 				kvmppc_do_h_remove(kvm, 0, i, 0, tmp);
1908 			++i;
1909 			hptp += 2;
1910 		}
1911 		err = 0;
1912 	}
1913 
1914  out:
1915 	/* Order HPTE updates vs. mmu_ready */
1916 	smp_wmb();
1917 	kvm->arch.mmu_ready = mmu_ready;
1918 	mutex_unlock(&kvm->arch.mmu_setup_lock);
1919 
1920 	if (err)
1921 		return err;
1922 	return nb;
1923 }
1924 
1925 static int kvm_htab_release(struct inode *inode, struct file *filp)
1926 {
1927 	struct kvm_htab_ctx *ctx = filp->private_data;
1928 
1929 	filp->private_data = NULL;
1930 	if (!(ctx->flags & KVM_GET_HTAB_WRITE))
1931 		atomic_dec(&ctx->kvm->arch.hpte_mod_interest);
1932 	kvm_put_kvm(ctx->kvm);
1933 	kfree(ctx);
1934 	return 0;
1935 }
1936 
1937 static const struct file_operations kvm_htab_fops = {
1938 	.read		= kvm_htab_read,
1939 	.write		= kvm_htab_write,
1940 	.llseek		= default_llseek,
1941 	.release	= kvm_htab_release,
1942 };
1943 
1944 int kvm_vm_ioctl_get_htab_fd(struct kvm *kvm, struct kvm_get_htab_fd *ghf)
1945 {
1946 	int ret;
1947 	struct kvm_htab_ctx *ctx;
1948 	int rwflag;
1949 
1950 	/* reject flags we don't recognize */
1951 	if (ghf->flags & ~(KVM_GET_HTAB_BOLTED_ONLY | KVM_GET_HTAB_WRITE))
1952 		return -EINVAL;
1953 	ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1954 	if (!ctx)
1955 		return -ENOMEM;
1956 	kvm_get_kvm(kvm);
1957 	ctx->kvm = kvm;
1958 	ctx->index = ghf->start_index;
1959 	ctx->flags = ghf->flags;
1960 	ctx->first_pass = 1;
1961 
1962 	rwflag = (ghf->flags & KVM_GET_HTAB_WRITE) ? O_WRONLY : O_RDONLY;
1963 	ret = anon_inode_getfd("kvm-htab", &kvm_htab_fops, ctx, rwflag | O_CLOEXEC);
1964 	if (ret < 0) {
1965 		kfree(ctx);
1966 		kvm_put_kvm_no_destroy(kvm);
1967 		return ret;
1968 	}
1969 
1970 	if (rwflag == O_RDONLY) {
1971 		mutex_lock(&kvm->slots_lock);
1972 		atomic_inc(&kvm->arch.hpte_mod_interest);
1973 		/* make sure kvmppc_do_h_enter etc. see the increment */
1974 		synchronize_srcu_expedited(&kvm->srcu);
1975 		mutex_unlock(&kvm->slots_lock);
1976 	}
1977 
1978 	return ret;
1979 }
1980 
1981 struct debugfs_htab_state {
1982 	struct kvm	*kvm;
1983 	struct mutex	mutex;
1984 	unsigned long	hpt_index;
1985 	int		chars_left;
1986 	int		buf_index;
1987 	char		buf[64];
1988 };
1989 
1990 static int debugfs_htab_open(struct inode *inode, struct file *file)
1991 {
1992 	struct kvm *kvm = inode->i_private;
1993 	struct debugfs_htab_state *p;
1994 
1995 	p = kzalloc(sizeof(*p), GFP_KERNEL);
1996 	if (!p)
1997 		return -ENOMEM;
1998 
1999 	kvm_get_kvm(kvm);
2000 	p->kvm = kvm;
2001 	mutex_init(&p->mutex);
2002 	file->private_data = p;
2003 
2004 	return nonseekable_open(inode, file);
2005 }
2006 
2007 static int debugfs_htab_release(struct inode *inode, struct file *file)
2008 {
2009 	struct debugfs_htab_state *p = file->private_data;
2010 
2011 	kvm_put_kvm(p->kvm);
2012 	kfree(p);
2013 	return 0;
2014 }
2015 
2016 static ssize_t debugfs_htab_read(struct file *file, char __user *buf,
2017 				 size_t len, loff_t *ppos)
2018 {
2019 	struct debugfs_htab_state *p = file->private_data;
2020 	ssize_t ret, r;
2021 	unsigned long i, n;
2022 	unsigned long v, hr, gr;
2023 	struct kvm *kvm;
2024 	__be64 *hptp;
2025 
2026 	kvm = p->kvm;
2027 	if (kvm_is_radix(kvm))
2028 		return 0;
2029 
2030 	ret = mutex_lock_interruptible(&p->mutex);
2031 	if (ret)
2032 		return ret;
2033 
2034 	if (p->chars_left) {
2035 		n = p->chars_left;
2036 		if (n > len)
2037 			n = len;
2038 		r = copy_to_user(buf, p->buf + p->buf_index, n);
2039 		n -= r;
2040 		p->chars_left -= n;
2041 		p->buf_index += n;
2042 		buf += n;
2043 		len -= n;
2044 		ret = n;
2045 		if (r) {
2046 			if (!n)
2047 				ret = -EFAULT;
2048 			goto out;
2049 		}
2050 	}
2051 
2052 	i = p->hpt_index;
2053 	hptp = (__be64 *)(kvm->arch.hpt.virt + (i * HPTE_SIZE));
2054 	for (; len != 0 && i < kvmppc_hpt_npte(&kvm->arch.hpt);
2055 	     ++i, hptp += 2) {
2056 		if (!(be64_to_cpu(hptp[0]) & (HPTE_V_VALID | HPTE_V_ABSENT)))
2057 			continue;
2058 
2059 		/* lock the HPTE so it's stable and read it */
2060 		preempt_disable();
2061 		while (!try_lock_hpte(hptp, HPTE_V_HVLOCK))
2062 			cpu_relax();
2063 		v = be64_to_cpu(hptp[0]) & ~HPTE_V_HVLOCK;
2064 		hr = be64_to_cpu(hptp[1]);
2065 		gr = kvm->arch.hpt.rev[i].guest_rpte;
2066 		unlock_hpte(hptp, v);
2067 		preempt_enable();
2068 
2069 		if (!(v & (HPTE_V_VALID | HPTE_V_ABSENT)))
2070 			continue;
2071 
2072 		n = scnprintf(p->buf, sizeof(p->buf),
2073 			      "%6lx %.16lx %.16lx %.16lx\n",
2074 			      i, v, hr, gr);
2075 		p->chars_left = n;
2076 		if (n > len)
2077 			n = len;
2078 		r = copy_to_user(buf, p->buf, n);
2079 		n -= r;
2080 		p->chars_left -= n;
2081 		p->buf_index = n;
2082 		buf += n;
2083 		len -= n;
2084 		ret += n;
2085 		if (r) {
2086 			if (!ret)
2087 				ret = -EFAULT;
2088 			goto out;
2089 		}
2090 	}
2091 	p->hpt_index = i;
2092 
2093  out:
2094 	mutex_unlock(&p->mutex);
2095 	return ret;
2096 }
2097 
2098 static ssize_t debugfs_htab_write(struct file *file, const char __user *buf,
2099 			   size_t len, loff_t *ppos)
2100 {
2101 	return -EACCES;
2102 }
2103 
2104 static const struct file_operations debugfs_htab_fops = {
2105 	.owner	 = THIS_MODULE,
2106 	.open	 = debugfs_htab_open,
2107 	.release = debugfs_htab_release,
2108 	.read	 = debugfs_htab_read,
2109 	.write	 = debugfs_htab_write,
2110 	.llseek	 = generic_file_llseek,
2111 };
2112 
2113 void kvmppc_mmu_debugfs_init(struct kvm *kvm)
2114 {
2115 	debugfs_create_file("htab", 0400, kvm->arch.debugfs_dir, kvm,
2116 			    &debugfs_htab_fops);
2117 }
2118 
2119 void kvmppc_mmu_book3s_hv_init(struct kvm_vcpu *vcpu)
2120 {
2121 	struct kvmppc_mmu *mmu = &vcpu->arch.mmu;
2122 
2123 	vcpu->arch.slb_nr = 32;		/* POWER7/POWER8 */
2124 
2125 	mmu->xlate = kvmppc_mmu_book3s_64_hv_xlate;
2126 
2127 	vcpu->arch.hflags |= BOOK3S_HFLAG_SLB;
2128 }
2129