xref: /linux/arch/s390/kvm/vsie.c (revision 0b92515916799ca2711c7c46abccb96c275cc88a)
1 /*
2  * kvm nested virtualization support for s390x
3  *
4  * Copyright IBM Corp. 2016
5  *
6  * This program is free software; you can redistribute it and/or modify
7  * it under the terms of the GNU General Public License (version 2 only)
8  * as published by the Free Software Foundation.
9  *
10  *    Author(s): David Hildenbrand <dahi@linux.vnet.ibm.com>
11  */
12 #include <linux/vmalloc.h>
13 #include <linux/kvm_host.h>
14 #include <linux/bug.h>
15 #include <linux/list.h>
16 #include <linux/bitmap.h>
17 #include <asm/gmap.h>
18 #include <asm/mmu_context.h>
19 #include <asm/sclp.h>
20 #include <asm/nmi.h>
21 #include <asm/dis.h>
22 #include "kvm-s390.h"
23 #include "gaccess.h"
24 
25 struct vsie_page {
26 	struct kvm_s390_sie_block scb_s;	/* 0x0000 */
27 	/* the pinned originial scb */
28 	struct kvm_s390_sie_block *scb_o;	/* 0x0200 */
29 	/* the shadow gmap in use by the vsie_page */
30 	struct gmap *gmap;			/* 0x0208 */
31 	/* address of the last reported fault to guest2 */
32 	unsigned long fault_addr;		/* 0x0210 */
33 	__u8 reserved[0x0700 - 0x0218];		/* 0x0218 */
34 	struct kvm_s390_crypto_cb crycb;	/* 0x0700 */
35 	__u8 fac[S390_ARCH_FAC_LIST_SIZE_BYTE];	/* 0x0800 */
36 } __packed;
37 
38 /* trigger a validity icpt for the given scb */
39 static int set_validity_icpt(struct kvm_s390_sie_block *scb,
40 			     __u16 reason_code)
41 {
42 	scb->ipa = 0x1000;
43 	scb->ipb = ((__u32) reason_code) << 16;
44 	scb->icptcode = ICPT_VALIDITY;
45 	return 1;
46 }
47 
48 /* mark the prefix as unmapped, this will block the VSIE */
49 static void prefix_unmapped(struct vsie_page *vsie_page)
50 {
51 	atomic_or(PROG_REQUEST, &vsie_page->scb_s.prog20);
52 }
53 
54 /* mark the prefix as unmapped and wait until the VSIE has been left */
55 static void prefix_unmapped_sync(struct vsie_page *vsie_page)
56 {
57 	prefix_unmapped(vsie_page);
58 	if (vsie_page->scb_s.prog0c & PROG_IN_SIE)
59 		atomic_or(CPUSTAT_STOP_INT, &vsie_page->scb_s.cpuflags);
60 	while (vsie_page->scb_s.prog0c & PROG_IN_SIE)
61 		cpu_relax();
62 }
63 
64 /* mark the prefix as mapped, this will allow the VSIE to run */
65 static void prefix_mapped(struct vsie_page *vsie_page)
66 {
67 	atomic_andnot(PROG_REQUEST, &vsie_page->scb_s.prog20);
68 }
69 
70 /* test if the prefix is mapped into the gmap shadow */
71 static int prefix_is_mapped(struct vsie_page *vsie_page)
72 {
73 	return !(atomic_read(&vsie_page->scb_s.prog20) & PROG_REQUEST);
74 }
75 
76 /* copy the updated intervention request bits into the shadow scb */
77 static void update_intervention_requests(struct vsie_page *vsie_page)
78 {
79 	const int bits = CPUSTAT_STOP_INT | CPUSTAT_IO_INT | CPUSTAT_EXT_INT;
80 	int cpuflags;
81 
82 	cpuflags = atomic_read(&vsie_page->scb_o->cpuflags);
83 	atomic_andnot(bits, &vsie_page->scb_s.cpuflags);
84 	atomic_or(cpuflags & bits, &vsie_page->scb_s.cpuflags);
85 }
86 
87 /* shadow (filter and validate) the cpuflags  */
88 static int prepare_cpuflags(struct kvm_vcpu *vcpu, struct vsie_page *vsie_page)
89 {
90 	struct kvm_s390_sie_block *scb_s = &vsie_page->scb_s;
91 	struct kvm_s390_sie_block *scb_o = vsie_page->scb_o;
92 	int newflags, cpuflags = atomic_read(&scb_o->cpuflags);
93 
94 	/* we don't allow ESA/390 guests */
95 	if (!(cpuflags & CPUSTAT_ZARCH))
96 		return set_validity_icpt(scb_s, 0x0001U);
97 
98 	if (cpuflags & (CPUSTAT_RRF | CPUSTAT_MCDS))
99 		return set_validity_icpt(scb_s, 0x0001U);
100 	else if (cpuflags & (CPUSTAT_SLSV | CPUSTAT_SLSR))
101 		return set_validity_icpt(scb_s, 0x0007U);
102 
103 	/* intervention requests will be set later */
104 	newflags = CPUSTAT_ZARCH;
105 	if (cpuflags & CPUSTAT_GED && test_kvm_facility(vcpu->kvm, 8))
106 		newflags |= CPUSTAT_GED;
107 	if (cpuflags & CPUSTAT_GED2 && test_kvm_facility(vcpu->kvm, 78)) {
108 		if (cpuflags & CPUSTAT_GED)
109 			return set_validity_icpt(scb_s, 0x0001U);
110 		newflags |= CPUSTAT_GED2;
111 	}
112 	if (test_kvm_cpu_feat(vcpu->kvm, KVM_S390_VM_CPU_FEAT_GPERE))
113 		newflags |= cpuflags & CPUSTAT_P;
114 	if (test_kvm_cpu_feat(vcpu->kvm, KVM_S390_VM_CPU_FEAT_GSLS))
115 		newflags |= cpuflags & CPUSTAT_SM;
116 	if (test_kvm_cpu_feat(vcpu->kvm, KVM_S390_VM_CPU_FEAT_IBS))
117 		newflags |= cpuflags & CPUSTAT_IBS;
118 
119 	atomic_set(&scb_s->cpuflags, newflags);
120 	return 0;
121 }
122 
123 /*
124  * Create a shadow copy of the crycb block and setup key wrapping, if
125  * requested for guest 3 and enabled for guest 2.
126  *
127  * We only accept format-1 (no AP in g2), but convert it into format-2
128  * There is nothing to do for format-0.
129  *
130  * Returns: - 0 if shadowed or nothing to do
131  *          - > 0 if control has to be given to guest 2
132  */
133 static int shadow_crycb(struct kvm_vcpu *vcpu, struct vsie_page *vsie_page)
134 {
135 	struct kvm_s390_sie_block *scb_s = &vsie_page->scb_s;
136 	struct kvm_s390_sie_block *scb_o = vsie_page->scb_o;
137 	u32 crycb_addr = scb_o->crycbd & 0x7ffffff8U;
138 	unsigned long *b1, *b2;
139 	u8 ecb3_flags;
140 
141 	scb_s->crycbd = 0;
142 	if (!(scb_o->crycbd & vcpu->arch.sie_block->crycbd & CRYCB_FORMAT1))
143 		return 0;
144 	/* format-1 is supported with message-security-assist extension 3 */
145 	if (!test_kvm_facility(vcpu->kvm, 76))
146 		return 0;
147 	/* we may only allow it if enabled for guest 2 */
148 	ecb3_flags = scb_o->ecb3 & vcpu->arch.sie_block->ecb3 &
149 		     (ECB3_AES | ECB3_DEA);
150 	if (!ecb3_flags)
151 		return 0;
152 
153 	if ((crycb_addr & PAGE_MASK) != ((crycb_addr + 128) & PAGE_MASK))
154 		return set_validity_icpt(scb_s, 0x003CU);
155 	else if (!crycb_addr)
156 		return set_validity_icpt(scb_s, 0x0039U);
157 
158 	/* copy only the wrapping keys */
159 	if (read_guest_real(vcpu, crycb_addr + 72, &vsie_page->crycb, 56))
160 		return set_validity_icpt(scb_s, 0x0035U);
161 
162 	scb_s->ecb3 |= ecb3_flags;
163 	scb_s->crycbd = ((__u32)(__u64) &vsie_page->crycb) | CRYCB_FORMAT1 |
164 			CRYCB_FORMAT2;
165 
166 	/* xor both blocks in one run */
167 	b1 = (unsigned long *) vsie_page->crycb.dea_wrapping_key_mask;
168 	b2 = (unsigned long *)
169 			    vcpu->kvm->arch.crypto.crycb->dea_wrapping_key_mask;
170 	/* as 56%8 == 0, bitmap_xor won't overwrite any data */
171 	bitmap_xor(b1, b1, b2, BITS_PER_BYTE * 56);
172 	return 0;
173 }
174 
175 /* shadow (round up/down) the ibc to avoid validity icpt */
176 static void prepare_ibc(struct kvm_vcpu *vcpu, struct vsie_page *vsie_page)
177 {
178 	struct kvm_s390_sie_block *scb_s = &vsie_page->scb_s;
179 	struct kvm_s390_sie_block *scb_o = vsie_page->scb_o;
180 	__u64 min_ibc = (sclp.ibc >> 16) & 0x0fffU;
181 
182 	scb_s->ibc = 0;
183 	/* ibc installed in g2 and requested for g3 */
184 	if (vcpu->kvm->arch.model.ibc && (scb_o->ibc & 0x0fffU)) {
185 		scb_s->ibc = scb_o->ibc & 0x0fffU;
186 		/* takte care of the minimum ibc level of the machine */
187 		if (scb_s->ibc < min_ibc)
188 			scb_s->ibc = min_ibc;
189 		/* take care of the maximum ibc level set for the guest */
190 		if (scb_s->ibc > vcpu->kvm->arch.model.ibc)
191 			scb_s->ibc = vcpu->kvm->arch.model.ibc;
192 	}
193 }
194 
195 /* unshadow the scb, copying parameters back to the real scb */
196 static void unshadow_scb(struct kvm_vcpu *vcpu, struct vsie_page *vsie_page)
197 {
198 	struct kvm_s390_sie_block *scb_s = &vsie_page->scb_s;
199 	struct kvm_s390_sie_block *scb_o = vsie_page->scb_o;
200 
201 	/* interception */
202 	scb_o->icptcode = scb_s->icptcode;
203 	scb_o->icptstatus = scb_s->icptstatus;
204 	scb_o->ipa = scb_s->ipa;
205 	scb_o->ipb = scb_s->ipb;
206 	scb_o->gbea = scb_s->gbea;
207 
208 	/* timer */
209 	scb_o->cputm = scb_s->cputm;
210 	scb_o->ckc = scb_s->ckc;
211 	scb_o->todpr = scb_s->todpr;
212 
213 	/* guest state */
214 	scb_o->gpsw = scb_s->gpsw;
215 	scb_o->gg14 = scb_s->gg14;
216 	scb_o->gg15 = scb_s->gg15;
217 	memcpy(scb_o->gcr, scb_s->gcr, 128);
218 	scb_o->pp = scb_s->pp;
219 
220 	/* interrupt intercept */
221 	switch (scb_s->icptcode) {
222 	case ICPT_PROGI:
223 	case ICPT_INSTPROGI:
224 	case ICPT_EXTINT:
225 		memcpy((void *)((u64)scb_o + 0xc0),
226 		       (void *)((u64)scb_s + 0xc0), 0xf0 - 0xc0);
227 		break;
228 	case ICPT_PARTEXEC:
229 		/* MVPG only */
230 		memcpy((void *)((u64)scb_o + 0xc0),
231 		       (void *)((u64)scb_s + 0xc0), 0xd0 - 0xc0);
232 		break;
233 	}
234 
235 	if (scb_s->ihcpu != 0xffffU)
236 		scb_o->ihcpu = scb_s->ihcpu;
237 }
238 
239 /*
240  * Setup the shadow scb by copying and checking the relevant parts of the g2
241  * provided scb.
242  *
243  * Returns: - 0 if the scb has been shadowed
244  *          - > 0 if control has to be given to guest 2
245  */
246 static int shadow_scb(struct kvm_vcpu *vcpu, struct vsie_page *vsie_page)
247 {
248 	struct kvm_s390_sie_block *scb_o = vsie_page->scb_o;
249 	struct kvm_s390_sie_block *scb_s = &vsie_page->scb_s;
250 	bool had_tx = scb_s->ecb & 0x10U;
251 	unsigned long new_mso = 0;
252 	int rc;
253 
254 	/* make sure we don't have any leftovers when reusing the scb */
255 	scb_s->icptcode = 0;
256 	scb_s->eca = 0;
257 	scb_s->ecb = 0;
258 	scb_s->ecb2 = 0;
259 	scb_s->ecb3 = 0;
260 	scb_s->ecd = 0;
261 	scb_s->fac = 0;
262 
263 	rc = prepare_cpuflags(vcpu, vsie_page);
264 	if (rc)
265 		goto out;
266 
267 	/* timer */
268 	scb_s->cputm = scb_o->cputm;
269 	scb_s->ckc = scb_o->ckc;
270 	scb_s->todpr = scb_o->todpr;
271 	scb_s->epoch = scb_o->epoch;
272 
273 	/* guest state */
274 	scb_s->gpsw = scb_o->gpsw;
275 	scb_s->gg14 = scb_o->gg14;
276 	scb_s->gg15 = scb_o->gg15;
277 	memcpy(scb_s->gcr, scb_o->gcr, 128);
278 	scb_s->pp = scb_o->pp;
279 
280 	/* interception / execution handling */
281 	scb_s->gbea = scb_o->gbea;
282 	scb_s->lctl = scb_o->lctl;
283 	scb_s->svcc = scb_o->svcc;
284 	scb_s->ictl = scb_o->ictl;
285 	/*
286 	 * SKEY handling functions can't deal with false setting of PTE invalid
287 	 * bits. Therefore we cannot provide interpretation and would later
288 	 * have to provide own emulation handlers.
289 	 */
290 	scb_s->ictl |= ICTL_ISKE | ICTL_SSKE | ICTL_RRBE;
291 	scb_s->icpua = scb_o->icpua;
292 
293 	if (!(atomic_read(&scb_s->cpuflags) & CPUSTAT_SM))
294 		new_mso = scb_o->mso & 0xfffffffffff00000UL;
295 	/* if the hva of the prefix changes, we have to remap the prefix */
296 	if (scb_s->mso != new_mso || scb_s->prefix != scb_o->prefix)
297 		prefix_unmapped(vsie_page);
298 	 /* SIE will do mso/msl validity and exception checks for us */
299 	scb_s->msl = scb_o->msl & 0xfffffffffff00000UL;
300 	scb_s->mso = new_mso;
301 	scb_s->prefix = scb_o->prefix;
302 
303 	/* We have to definetly flush the tlb if this scb never ran */
304 	if (scb_s->ihcpu != 0xffffU)
305 		scb_s->ihcpu = scb_o->ihcpu;
306 
307 	/* MVPG and Protection Exception Interpretation are always available */
308 	scb_s->eca |= scb_o->eca & 0x01002000U;
309 	/* Host-protection-interruption introduced with ESOP */
310 	if (test_kvm_cpu_feat(vcpu->kvm, KVM_S390_VM_CPU_FEAT_ESOP))
311 		scb_s->ecb |= scb_o->ecb & 0x02U;
312 	/* transactional execution */
313 	if (test_kvm_facility(vcpu->kvm, 73)) {
314 		/* remap the prefix is tx is toggled on */
315 		if ((scb_o->ecb & 0x10U) && !had_tx)
316 			prefix_unmapped(vsie_page);
317 		scb_s->ecb |= scb_o->ecb & 0x10U;
318 	}
319 	/* SIMD */
320 	if (test_kvm_facility(vcpu->kvm, 129)) {
321 		scb_s->eca |= scb_o->eca & 0x00020000U;
322 		scb_s->ecd |= scb_o->ecd & 0x20000000U;
323 	}
324 	/* Run-time-Instrumentation */
325 	if (test_kvm_facility(vcpu->kvm, 64))
326 		scb_s->ecb3 |= scb_o->ecb3 & 0x01U;
327 	if (test_kvm_cpu_feat(vcpu->kvm, KVM_S390_VM_CPU_FEAT_SIIF))
328 		scb_s->eca |= scb_o->eca & 0x00000001U;
329 	if (test_kvm_cpu_feat(vcpu->kvm, KVM_S390_VM_CPU_FEAT_IB))
330 		scb_s->eca |= scb_o->eca & 0x40000000U;
331 	if (test_kvm_cpu_feat(vcpu->kvm, KVM_S390_VM_CPU_FEAT_CEI))
332 		scb_s->eca |= scb_o->eca & 0x80000000U;
333 
334 	prepare_ibc(vcpu, vsie_page);
335 	rc = shadow_crycb(vcpu, vsie_page);
336 out:
337 	if (rc)
338 		unshadow_scb(vcpu, vsie_page);
339 	return rc;
340 }
341 
342 void kvm_s390_vsie_gmap_notifier(struct gmap *gmap, unsigned long start,
343 				 unsigned long end)
344 {
345 	struct kvm *kvm = gmap->private;
346 	struct vsie_page *cur;
347 	unsigned long prefix;
348 	struct page *page;
349 	int i;
350 
351 	if (!gmap_is_shadow(gmap))
352 		return;
353 	if (start >= 1UL << 31)
354 		/* We are only interested in prefix pages */
355 		return;
356 
357 	/*
358 	 * Only new shadow blocks are added to the list during runtime,
359 	 * therefore we can safely reference them all the time.
360 	 */
361 	for (i = 0; i < kvm->arch.vsie.page_count; i++) {
362 		page = READ_ONCE(kvm->arch.vsie.pages[i]);
363 		if (!page)
364 			continue;
365 		cur = page_to_virt(page);
366 		if (READ_ONCE(cur->gmap) != gmap)
367 			continue;
368 		prefix = cur->scb_s.prefix << GUEST_PREFIX_SHIFT;
369 		/* with mso/msl, the prefix lies at an offset */
370 		prefix += cur->scb_s.mso;
371 		if (prefix <= end && start <= prefix + 2 * PAGE_SIZE - 1)
372 			prefix_unmapped_sync(cur);
373 	}
374 }
375 
376 /*
377  * Map the first prefix page and if tx is enabled also the second prefix page.
378  *
379  * The prefix will be protected, a gmap notifier will inform about unmaps.
380  * The shadow scb must not be executed until the prefix is remapped, this is
381  * guaranteed by properly handling PROG_REQUEST.
382  *
383  * Returns: - 0 on if successfully mapped or already mapped
384  *          - > 0 if control has to be given to guest 2
385  *          - -EAGAIN if the caller can retry immediately
386  *          - -ENOMEM if out of memory
387  */
388 static int map_prefix(struct kvm_vcpu *vcpu, struct vsie_page *vsie_page)
389 {
390 	struct kvm_s390_sie_block *scb_s = &vsie_page->scb_s;
391 	u64 prefix = scb_s->prefix << GUEST_PREFIX_SHIFT;
392 	int rc;
393 
394 	if (prefix_is_mapped(vsie_page))
395 		return 0;
396 
397 	/* mark it as mapped so we can catch any concurrent unmappers */
398 	prefix_mapped(vsie_page);
399 
400 	/* with mso/msl, the prefix lies at offset *mso* */
401 	prefix += scb_s->mso;
402 
403 	rc = kvm_s390_shadow_fault(vcpu, vsie_page->gmap, prefix);
404 	if (!rc && (scb_s->ecb & 0x10U))
405 		rc = kvm_s390_shadow_fault(vcpu, vsie_page->gmap,
406 					   prefix + PAGE_SIZE);
407 	/*
408 	 * We don't have to mprotect, we will be called for all unshadows.
409 	 * SIE will detect if protection applies and trigger a validity.
410 	 */
411 	if (rc)
412 		prefix_unmapped(vsie_page);
413 	if (rc > 0 || rc == -EFAULT)
414 		rc = set_validity_icpt(scb_s, 0x0037U);
415 	return rc;
416 }
417 
418 /*
419  * Pin the guest page given by gpa and set hpa to the pinned host address.
420  * Will always be pinned writable.
421  *
422  * Returns: - 0 on success
423  *          - -EINVAL if the gpa is not valid guest storage
424  *          - -ENOMEM if out of memory
425  */
426 static int pin_guest_page(struct kvm *kvm, gpa_t gpa, hpa_t *hpa)
427 {
428 	struct page *page;
429 	hva_t hva;
430 	int rc;
431 
432 	hva = gfn_to_hva(kvm, gpa_to_gfn(gpa));
433 	if (kvm_is_error_hva(hva))
434 		return -EINVAL;
435 	rc = get_user_pages_fast(hva, 1, 1, &page);
436 	if (rc < 0)
437 		return rc;
438 	else if (rc != 1)
439 		return -ENOMEM;
440 	*hpa = (hpa_t) page_to_virt(page) + (gpa & ~PAGE_MASK);
441 	return 0;
442 }
443 
444 /* Unpins a page previously pinned via pin_guest_page, marking it as dirty. */
445 static void unpin_guest_page(struct kvm *kvm, gpa_t gpa, hpa_t hpa)
446 {
447 	struct page *page;
448 
449 	page = virt_to_page(hpa);
450 	set_page_dirty_lock(page);
451 	put_page(page);
452 	/* mark the page always as dirty for migration */
453 	mark_page_dirty(kvm, gpa_to_gfn(gpa));
454 }
455 
456 /* unpin all blocks previously pinned by pin_blocks(), marking them dirty */
457 static void unpin_blocks(struct kvm_vcpu *vcpu, struct vsie_page *vsie_page)
458 {
459 	struct kvm_s390_sie_block *scb_o = vsie_page->scb_o;
460 	struct kvm_s390_sie_block *scb_s = &vsie_page->scb_s;
461 	hpa_t hpa;
462 	gpa_t gpa;
463 
464 	hpa = (u64) scb_s->scaoh << 32 | scb_s->scaol;
465 	if (hpa) {
466 		gpa = scb_o->scaol & ~0xfUL;
467 		if (test_kvm_cpu_feat(vcpu->kvm, KVM_S390_VM_CPU_FEAT_64BSCAO))
468 			gpa |= (u64) scb_o->scaoh << 32;
469 		unpin_guest_page(vcpu->kvm, gpa, hpa);
470 		scb_s->scaol = 0;
471 		scb_s->scaoh = 0;
472 	}
473 
474 	hpa = scb_s->itdba;
475 	if (hpa) {
476 		gpa = scb_o->itdba & ~0xffUL;
477 		unpin_guest_page(vcpu->kvm, gpa, hpa);
478 		scb_s->itdba = 0;
479 	}
480 
481 	hpa = scb_s->gvrd;
482 	if (hpa) {
483 		gpa = scb_o->gvrd & ~0x1ffUL;
484 		unpin_guest_page(vcpu->kvm, gpa, hpa);
485 		scb_s->gvrd = 0;
486 	}
487 
488 	hpa = scb_s->riccbd;
489 	if (hpa) {
490 		gpa = scb_o->riccbd & ~0x3fUL;
491 		unpin_guest_page(vcpu->kvm, gpa, hpa);
492 		scb_s->riccbd = 0;
493 	}
494 }
495 
496 /*
497  * Instead of shadowing some blocks, we can simply forward them because the
498  * addresses in the scb are 64 bit long.
499  *
500  * This works as long as the data lies in one page. If blocks ever exceed one
501  * page, we have to fall back to shadowing.
502  *
503  * As we reuse the sca, the vcpu pointers contained in it are invalid. We must
504  * therefore not enable any facilities that access these pointers (e.g. SIGPIF).
505  *
506  * Returns: - 0 if all blocks were pinned.
507  *          - > 0 if control has to be given to guest 2
508  *          - -ENOMEM if out of memory
509  */
510 static int pin_blocks(struct kvm_vcpu *vcpu, struct vsie_page *vsie_page)
511 {
512 	struct kvm_s390_sie_block *scb_o = vsie_page->scb_o;
513 	struct kvm_s390_sie_block *scb_s = &vsie_page->scb_s;
514 	hpa_t hpa;
515 	gpa_t gpa;
516 	int rc = 0;
517 
518 	gpa = scb_o->scaol & ~0xfUL;
519 	if (test_kvm_cpu_feat(vcpu->kvm, KVM_S390_VM_CPU_FEAT_64BSCAO))
520 		gpa |= (u64) scb_o->scaoh << 32;
521 	if (gpa) {
522 		if (!(gpa & ~0x1fffUL))
523 			rc = set_validity_icpt(scb_s, 0x0038U);
524 		else if ((gpa & ~0x1fffUL) == kvm_s390_get_prefix(vcpu))
525 			rc = set_validity_icpt(scb_s, 0x0011U);
526 		else if ((gpa & PAGE_MASK) !=
527 			 ((gpa + sizeof(struct bsca_block) - 1) & PAGE_MASK))
528 			rc = set_validity_icpt(scb_s, 0x003bU);
529 		if (!rc) {
530 			rc = pin_guest_page(vcpu->kvm, gpa, &hpa);
531 			if (rc == -EINVAL)
532 				rc = set_validity_icpt(scb_s, 0x0034U);
533 		}
534 		if (rc)
535 			goto unpin;
536 		scb_s->scaoh = (u32)((u64)hpa >> 32);
537 		scb_s->scaol = (u32)(u64)hpa;
538 	}
539 
540 	gpa = scb_o->itdba & ~0xffUL;
541 	if (gpa && (scb_s->ecb & 0x10U)) {
542 		if (!(gpa & ~0x1fffU)) {
543 			rc = set_validity_icpt(scb_s, 0x0080U);
544 			goto unpin;
545 		}
546 		/* 256 bytes cannot cross page boundaries */
547 		rc = pin_guest_page(vcpu->kvm, gpa, &hpa);
548 		if (rc == -EINVAL)
549 			rc = set_validity_icpt(scb_s, 0x0080U);
550 		if (rc)
551 			goto unpin;
552 		scb_s->itdba = hpa;
553 	}
554 
555 	gpa = scb_o->gvrd & ~0x1ffUL;
556 	if (gpa && (scb_s->eca & 0x00020000U) &&
557 	    !(scb_s->ecd & 0x20000000U)) {
558 		if (!(gpa & ~0x1fffUL)) {
559 			rc = set_validity_icpt(scb_s, 0x1310U);
560 			goto unpin;
561 		}
562 		/*
563 		 * 512 bytes vector registers cannot cross page boundaries
564 		 * if this block gets bigger, we have to shadow it.
565 		 */
566 		rc = pin_guest_page(vcpu->kvm, gpa, &hpa);
567 		if (rc == -EINVAL)
568 			rc = set_validity_icpt(scb_s, 0x1310U);
569 		if (rc)
570 			goto unpin;
571 		scb_s->gvrd = hpa;
572 	}
573 
574 	gpa = scb_o->riccbd & ~0x3fUL;
575 	if (gpa && (scb_s->ecb3 & 0x01U)) {
576 		if (!(gpa & ~0x1fffUL)) {
577 			rc = set_validity_icpt(scb_s, 0x0043U);
578 			goto unpin;
579 		}
580 		/* 64 bytes cannot cross page boundaries */
581 		rc = pin_guest_page(vcpu->kvm, gpa, &hpa);
582 		if (rc == -EINVAL)
583 			rc = set_validity_icpt(scb_s, 0x0043U);
584 		/* Validity 0x0044 will be checked by SIE */
585 		if (rc)
586 			goto unpin;
587 		scb_s->riccbd = hpa;
588 	}
589 	return 0;
590 unpin:
591 	unpin_blocks(vcpu, vsie_page);
592 	return rc;
593 }
594 
595 /* unpin the scb provided by guest 2, marking it as dirty */
596 static void unpin_scb(struct kvm_vcpu *vcpu, struct vsie_page *vsie_page,
597 		      gpa_t gpa)
598 {
599 	hpa_t hpa = (hpa_t) vsie_page->scb_o;
600 
601 	if (hpa)
602 		unpin_guest_page(vcpu->kvm, gpa, hpa);
603 	vsie_page->scb_o = NULL;
604 }
605 
606 /*
607  * Pin the scb at gpa provided by guest 2 at vsie_page->scb_o.
608  *
609  * Returns: - 0 if the scb was pinned.
610  *          - > 0 if control has to be given to guest 2
611  *          - -ENOMEM if out of memory
612  */
613 static int pin_scb(struct kvm_vcpu *vcpu, struct vsie_page *vsie_page,
614 		   gpa_t gpa)
615 {
616 	hpa_t hpa;
617 	int rc;
618 
619 	rc = pin_guest_page(vcpu->kvm, gpa, &hpa);
620 	if (rc == -EINVAL) {
621 		rc = kvm_s390_inject_program_int(vcpu, PGM_ADDRESSING);
622 		if (!rc)
623 			rc = 1;
624 	}
625 	if (!rc)
626 		vsie_page->scb_o = (struct kvm_s390_sie_block *) hpa;
627 	return rc;
628 }
629 
630 /*
631  * Inject a fault into guest 2.
632  *
633  * Returns: - > 0 if control has to be given to guest 2
634  *            < 0 if an error occurred during injection.
635  */
636 static int inject_fault(struct kvm_vcpu *vcpu, __u16 code, __u64 vaddr,
637 			bool write_flag)
638 {
639 	struct kvm_s390_pgm_info pgm = {
640 		.code = code,
641 		.trans_exc_code =
642 			/* 0-51: virtual address */
643 			(vaddr & 0xfffffffffffff000UL) |
644 			/* 52-53: store / fetch */
645 			(((unsigned int) !write_flag) + 1) << 10,
646 			/* 62-63: asce id (alway primary == 0) */
647 		.exc_access_id = 0, /* always primary */
648 		.op_access_id = 0, /* not MVPG */
649 	};
650 	int rc;
651 
652 	if (code == PGM_PROTECTION)
653 		pgm.trans_exc_code |= 0x4UL;
654 
655 	rc = kvm_s390_inject_prog_irq(vcpu, &pgm);
656 	return rc ? rc : 1;
657 }
658 
659 /*
660  * Handle a fault during vsie execution on a gmap shadow.
661  *
662  * Returns: - 0 if the fault was resolved
663  *          - > 0 if control has to be given to guest 2
664  *          - < 0 if an error occurred
665  */
666 static int handle_fault(struct kvm_vcpu *vcpu, struct vsie_page *vsie_page)
667 {
668 	int rc;
669 
670 	if (current->thread.gmap_int_code == PGM_PROTECTION)
671 		/* we can directly forward all protection exceptions */
672 		return inject_fault(vcpu, PGM_PROTECTION,
673 				    current->thread.gmap_addr, 1);
674 
675 	rc = kvm_s390_shadow_fault(vcpu, vsie_page->gmap,
676 				   current->thread.gmap_addr);
677 	if (rc > 0) {
678 		rc = inject_fault(vcpu, rc,
679 				  current->thread.gmap_addr,
680 				  current->thread.gmap_write_flag);
681 		if (rc >= 0)
682 			vsie_page->fault_addr = current->thread.gmap_addr;
683 	}
684 	return rc;
685 }
686 
687 /*
688  * Retry the previous fault that required guest 2 intervention. This avoids
689  * one superfluous SIE re-entry and direct exit.
690  *
691  * Will ignore any errors. The next SIE fault will do proper fault handling.
692  */
693 static void handle_last_fault(struct kvm_vcpu *vcpu,
694 			      struct vsie_page *vsie_page)
695 {
696 	if (vsie_page->fault_addr)
697 		kvm_s390_shadow_fault(vcpu, vsie_page->gmap,
698 				      vsie_page->fault_addr);
699 	vsie_page->fault_addr = 0;
700 }
701 
702 static inline void clear_vsie_icpt(struct vsie_page *vsie_page)
703 {
704 	vsie_page->scb_s.icptcode = 0;
705 }
706 
707 /* rewind the psw and clear the vsie icpt, so we can retry execution */
708 static void retry_vsie_icpt(struct vsie_page *vsie_page)
709 {
710 	struct kvm_s390_sie_block *scb_s = &vsie_page->scb_s;
711 	int ilen = insn_length(scb_s->ipa >> 8);
712 
713 	/* take care of EXECUTE instructions */
714 	if (scb_s->icptstatus & 1) {
715 		ilen = (scb_s->icptstatus >> 4) & 0x6;
716 		if (!ilen)
717 			ilen = 4;
718 	}
719 	scb_s->gpsw.addr = __rewind_psw(scb_s->gpsw, ilen);
720 	clear_vsie_icpt(vsie_page);
721 }
722 
723 /*
724  * Try to shadow + enable the guest 2 provided facility list.
725  * Retry instruction execution if enabled for and provided by guest 2.
726  *
727  * Returns: - 0 if handled (retry or guest 2 icpt)
728  *          - > 0 if control has to be given to guest 2
729  */
730 static int handle_stfle(struct kvm_vcpu *vcpu, struct vsie_page *vsie_page)
731 {
732 	struct kvm_s390_sie_block *scb_s = &vsie_page->scb_s;
733 	__u32 fac = vsie_page->scb_o->fac & 0x7ffffff8U;
734 
735 	if (fac && test_kvm_facility(vcpu->kvm, 7)) {
736 		retry_vsie_icpt(vsie_page);
737 		if (read_guest_real(vcpu, fac, &vsie_page->fac,
738 				    sizeof(vsie_page->fac)))
739 			return set_validity_icpt(scb_s, 0x1090U);
740 		scb_s->fac = (__u32)(__u64) &vsie_page->fac;
741 	}
742 	return 0;
743 }
744 
745 /*
746  * Run the vsie on a shadow scb and a shadow gmap, without any further
747  * sanity checks, handling SIE faults.
748  *
749  * Returns: - 0 everything went fine
750  *          - > 0 if control has to be given to guest 2
751  *          - < 0 if an error occurred
752  */
753 static int do_vsie_run(struct kvm_vcpu *vcpu, struct vsie_page *vsie_page)
754 {
755 	struct kvm_s390_sie_block *scb_s = &vsie_page->scb_s;
756 	struct kvm_s390_sie_block *scb_o = vsie_page->scb_o;
757 	int rc;
758 
759 	handle_last_fault(vcpu, vsie_page);
760 
761 	if (need_resched())
762 		schedule();
763 	if (test_cpu_flag(CIF_MCCK_PENDING))
764 		s390_handle_mcck();
765 
766 	srcu_read_unlock(&vcpu->kvm->srcu, vcpu->srcu_idx);
767 	local_irq_disable();
768 	guest_enter_irqoff();
769 	local_irq_enable();
770 
771 	rc = sie64a(scb_s, vcpu->run->s.regs.gprs);
772 
773 	local_irq_disable();
774 	guest_exit_irqoff();
775 	local_irq_enable();
776 	vcpu->srcu_idx = srcu_read_lock(&vcpu->kvm->srcu);
777 
778 	if (rc > 0)
779 		rc = 0; /* we could still have an icpt */
780 	else if (rc == -EFAULT)
781 		return handle_fault(vcpu, vsie_page);
782 
783 	switch (scb_s->icptcode) {
784 	case ICPT_INST:
785 		if (scb_s->ipa == 0xb2b0)
786 			rc = handle_stfle(vcpu, vsie_page);
787 		break;
788 	case ICPT_STOP:
789 		/* stop not requested by g2 - must have been a kick */
790 		if (!(atomic_read(&scb_o->cpuflags) & CPUSTAT_STOP_INT))
791 			clear_vsie_icpt(vsie_page);
792 		break;
793 	case ICPT_VALIDITY:
794 		if ((scb_s->ipa & 0xf000) != 0xf000)
795 			scb_s->ipa += 0x1000;
796 		break;
797 	}
798 	return rc;
799 }
800 
801 static void release_gmap_shadow(struct vsie_page *vsie_page)
802 {
803 	if (vsie_page->gmap)
804 		gmap_put(vsie_page->gmap);
805 	WRITE_ONCE(vsie_page->gmap, NULL);
806 	prefix_unmapped(vsie_page);
807 }
808 
809 static int acquire_gmap_shadow(struct kvm_vcpu *vcpu,
810 			       struct vsie_page *vsie_page)
811 {
812 	unsigned long asce;
813 	union ctlreg0 cr0;
814 	struct gmap *gmap;
815 	int edat;
816 
817 	asce = vcpu->arch.sie_block->gcr[1];
818 	cr0.val = vcpu->arch.sie_block->gcr[0];
819 	edat = cr0.edat && test_kvm_facility(vcpu->kvm, 8);
820 	edat += edat && test_kvm_facility(vcpu->kvm, 78);
821 
822 	/*
823 	 * ASCE or EDAT could have changed since last icpt, or the gmap
824 	 * we're holding has been unshadowed. If the gmap is still valid,
825 	 * we can safely reuse it.
826 	 */
827 	if (vsie_page->gmap && gmap_shadow_valid(vsie_page->gmap, asce, edat))
828 		return 0;
829 
830 	/* release the old shadow - if any, and mark the prefix as unmapped */
831 	release_gmap_shadow(vsie_page);
832 	gmap = gmap_shadow(vcpu->arch.gmap, asce, edat);
833 	if (IS_ERR(gmap))
834 		return PTR_ERR(gmap);
835 	gmap->private = vcpu->kvm;
836 	WRITE_ONCE(vsie_page->gmap, gmap);
837 	return 0;
838 }
839 
840 /*
841  * Register the shadow scb at the VCPU, e.g. for kicking out of vsie.
842  */
843 static void register_shadow_scb(struct kvm_vcpu *vcpu,
844 				struct vsie_page *vsie_page)
845 {
846 	struct kvm_s390_sie_block *scb_s = &vsie_page->scb_s;
847 
848 	WRITE_ONCE(vcpu->arch.vsie_block, &vsie_page->scb_s);
849 	/*
850 	 * External calls have to lead to a kick of the vcpu and
851 	 * therefore the vsie -> Simulate Wait state.
852 	 */
853 	atomic_or(CPUSTAT_WAIT, &vcpu->arch.sie_block->cpuflags);
854 	/*
855 	 * We have to adjust the g3 epoch by the g2 epoch. The epoch will
856 	 * automatically be adjusted on tod clock changes via kvm_sync_clock.
857 	 */
858 	preempt_disable();
859 	scb_s->epoch += vcpu->kvm->arch.epoch;
860 	preempt_enable();
861 }
862 
863 /*
864  * Unregister a shadow scb from a VCPU.
865  */
866 static void unregister_shadow_scb(struct kvm_vcpu *vcpu)
867 {
868 	atomic_andnot(CPUSTAT_WAIT, &vcpu->arch.sie_block->cpuflags);
869 	WRITE_ONCE(vcpu->arch.vsie_block, NULL);
870 }
871 
872 /*
873  * Run the vsie on a shadowed scb, managing the gmap shadow, handling
874  * prefix pages and faults.
875  *
876  * Returns: - 0 if no errors occurred
877  *          - > 0 if control has to be given to guest 2
878  *          - -ENOMEM if out of memory
879  */
880 static int vsie_run(struct kvm_vcpu *vcpu, struct vsie_page *vsie_page)
881 {
882 	struct kvm_s390_sie_block *scb_s = &vsie_page->scb_s;
883 	int rc = 0;
884 
885 	while (1) {
886 		rc = acquire_gmap_shadow(vcpu, vsie_page);
887 		if (!rc)
888 			rc = map_prefix(vcpu, vsie_page);
889 		if (!rc) {
890 			gmap_enable(vsie_page->gmap);
891 			update_intervention_requests(vsie_page);
892 			rc = do_vsie_run(vcpu, vsie_page);
893 			gmap_enable(vcpu->arch.gmap);
894 		}
895 		atomic_andnot(PROG_BLOCK_SIE, &scb_s->prog20);
896 
897 		if (rc == -EAGAIN)
898 			rc = 0;
899 		if (rc || scb_s->icptcode || signal_pending(current) ||
900 		    kvm_s390_vcpu_has_irq(vcpu, 0))
901 			break;
902 	}
903 
904 	if (rc == -EFAULT) {
905 		/*
906 		 * Addressing exceptions are always presentes as intercepts.
907 		 * As addressing exceptions are suppressing and our guest 3 PSW
908 		 * points at the responsible instruction, we have to
909 		 * forward the PSW and set the ilc. If we can't read guest 3
910 		 * instruction, we can use an arbitrary ilc. Let's always use
911 		 * ilen = 4 for now, so we can avoid reading in guest 3 virtual
912 		 * memory. (we could also fake the shadow so the hardware
913 		 * handles it).
914 		 */
915 		scb_s->icptcode = ICPT_PROGI;
916 		scb_s->iprcc = PGM_ADDRESSING;
917 		scb_s->pgmilc = 4;
918 		scb_s->gpsw.addr = __rewind_psw(scb_s->gpsw, 4);
919 	}
920 	return rc;
921 }
922 
923 /*
924  * Get or create a vsie page for a scb address.
925  *
926  * Returns: - address of a vsie page (cached or new one)
927  *          - NULL if the same scb address is already used by another VCPU
928  *          - ERR_PTR(-ENOMEM) if out of memory
929  */
930 static struct vsie_page *get_vsie_page(struct kvm *kvm, unsigned long addr)
931 {
932 	struct vsie_page *vsie_page;
933 	struct page *page;
934 	int nr_vcpus;
935 
936 	rcu_read_lock();
937 	page = radix_tree_lookup(&kvm->arch.vsie.addr_to_page, addr >> 9);
938 	rcu_read_unlock();
939 	if (page) {
940 		if (page_ref_inc_return(page) == 2)
941 			return page_to_virt(page);
942 		page_ref_dec(page);
943 	}
944 
945 	/*
946 	 * We want at least #online_vcpus shadows, so every VCPU can execute
947 	 * the VSIE in parallel.
948 	 */
949 	nr_vcpus = atomic_read(&kvm->online_vcpus);
950 
951 	mutex_lock(&kvm->arch.vsie.mutex);
952 	if (kvm->arch.vsie.page_count < nr_vcpus) {
953 		page = alloc_page(GFP_KERNEL | __GFP_ZERO | GFP_DMA);
954 		if (!page) {
955 			mutex_unlock(&kvm->arch.vsie.mutex);
956 			return ERR_PTR(-ENOMEM);
957 		}
958 		page_ref_inc(page);
959 		kvm->arch.vsie.pages[kvm->arch.vsie.page_count] = page;
960 		kvm->arch.vsie.page_count++;
961 	} else {
962 		/* reuse an existing entry that belongs to nobody */
963 		while (true) {
964 			page = kvm->arch.vsie.pages[kvm->arch.vsie.next];
965 			if (page_ref_inc_return(page) == 2)
966 				break;
967 			page_ref_dec(page);
968 			kvm->arch.vsie.next++;
969 			kvm->arch.vsie.next %= nr_vcpus;
970 		}
971 		radix_tree_delete(&kvm->arch.vsie.addr_to_page, page->index >> 9);
972 	}
973 	page->index = addr;
974 	/* double use of the same address */
975 	if (radix_tree_insert(&kvm->arch.vsie.addr_to_page, addr >> 9, page)) {
976 		page_ref_dec(page);
977 		mutex_unlock(&kvm->arch.vsie.mutex);
978 		return NULL;
979 	}
980 	mutex_unlock(&kvm->arch.vsie.mutex);
981 
982 	vsie_page = page_to_virt(page);
983 	memset(&vsie_page->scb_s, 0, sizeof(struct kvm_s390_sie_block));
984 	release_gmap_shadow(vsie_page);
985 	vsie_page->fault_addr = 0;
986 	vsie_page->scb_s.ihcpu = 0xffffU;
987 	return vsie_page;
988 }
989 
990 /* put a vsie page acquired via get_vsie_page */
991 static void put_vsie_page(struct kvm *kvm, struct vsie_page *vsie_page)
992 {
993 	struct page *page = pfn_to_page(__pa(vsie_page) >> PAGE_SHIFT);
994 
995 	page_ref_dec(page);
996 }
997 
998 int kvm_s390_handle_vsie(struct kvm_vcpu *vcpu)
999 {
1000 	struct vsie_page *vsie_page;
1001 	unsigned long scb_addr;
1002 	int rc;
1003 
1004 	vcpu->stat.instruction_sie++;
1005 	if (!test_kvm_cpu_feat(vcpu->kvm, KVM_S390_VM_CPU_FEAT_SIEF2))
1006 		return -EOPNOTSUPP;
1007 	if (vcpu->arch.sie_block->gpsw.mask & PSW_MASK_PSTATE)
1008 		return kvm_s390_inject_program_int(vcpu, PGM_PRIVILEGED_OP);
1009 
1010 	BUILD_BUG_ON(sizeof(struct vsie_page) != 4096);
1011 	scb_addr = kvm_s390_get_base_disp_s(vcpu, NULL);
1012 
1013 	/* 512 byte alignment */
1014 	if (unlikely(scb_addr & 0x1ffUL))
1015 		return kvm_s390_inject_program_int(vcpu, PGM_SPECIFICATION);
1016 
1017 	if (signal_pending(current) || kvm_s390_vcpu_has_irq(vcpu, 0))
1018 		return 0;
1019 
1020 	vsie_page = get_vsie_page(vcpu->kvm, scb_addr);
1021 	if (IS_ERR(vsie_page))
1022 		return PTR_ERR(vsie_page);
1023 	else if (!vsie_page)
1024 		/* double use of sie control block - simply do nothing */
1025 		return 0;
1026 
1027 	rc = pin_scb(vcpu, vsie_page, scb_addr);
1028 	if (rc)
1029 		goto out_put;
1030 	rc = shadow_scb(vcpu, vsie_page);
1031 	if (rc)
1032 		goto out_unpin_scb;
1033 	rc = pin_blocks(vcpu, vsie_page);
1034 	if (rc)
1035 		goto out_unshadow;
1036 	register_shadow_scb(vcpu, vsie_page);
1037 	rc = vsie_run(vcpu, vsie_page);
1038 	unregister_shadow_scb(vcpu);
1039 	unpin_blocks(vcpu, vsie_page);
1040 out_unshadow:
1041 	unshadow_scb(vcpu, vsie_page);
1042 out_unpin_scb:
1043 	unpin_scb(vcpu, vsie_page, scb_addr);
1044 out_put:
1045 	put_vsie_page(vcpu->kvm, vsie_page);
1046 
1047 	return rc < 0 ? rc : 0;
1048 }
1049 
1050 /* Init the vsie data structures. To be called when a vm is initialized. */
1051 void kvm_s390_vsie_init(struct kvm *kvm)
1052 {
1053 	mutex_init(&kvm->arch.vsie.mutex);
1054 	INIT_RADIX_TREE(&kvm->arch.vsie.addr_to_page, GFP_KERNEL);
1055 }
1056 
1057 /* Destroy the vsie data structures. To be called when a vm is destroyed. */
1058 void kvm_s390_vsie_destroy(struct kvm *kvm)
1059 {
1060 	struct vsie_page *vsie_page;
1061 	struct page *page;
1062 	int i;
1063 
1064 	mutex_lock(&kvm->arch.vsie.mutex);
1065 	for (i = 0; i < kvm->arch.vsie.page_count; i++) {
1066 		page = kvm->arch.vsie.pages[i];
1067 		kvm->arch.vsie.pages[i] = NULL;
1068 		vsie_page = page_to_virt(page);
1069 		release_gmap_shadow(vsie_page);
1070 		/* free the radix tree entry */
1071 		radix_tree_delete(&kvm->arch.vsie.addr_to_page, page->index >> 9);
1072 		__free_page(page);
1073 	}
1074 	kvm->arch.vsie.page_count = 0;
1075 	mutex_unlock(&kvm->arch.vsie.mutex);
1076 }
1077 
1078 void kvm_s390_vsie_kick(struct kvm_vcpu *vcpu)
1079 {
1080 	struct kvm_s390_sie_block *scb = READ_ONCE(vcpu->arch.vsie_block);
1081 
1082 	/*
1083 	 * Even if the VCPU lets go of the shadow sie block reference, it is
1084 	 * still valid in the cache. So we can safely kick it.
1085 	 */
1086 	if (scb) {
1087 		atomic_or(PROG_BLOCK_SIE, &scb->prog20);
1088 		if (scb->prog0c & PROG_IN_SIE)
1089 			atomic_or(CPUSTAT_STOP_INT, &scb->cpuflags);
1090 	}
1091 }
1092