xref: /openbmc/linux/arch/s390/kvm/vsie.c (revision 2891f2d5)
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 <linux/sched/signal.h>
18 
19 #include <asm/gmap.h>
20 #include <asm/mmu_context.h>
21 #include <asm/sclp.h>
22 #include <asm/nmi.h>
23 #include <asm/dis.h>
24 #include "kvm-s390.h"
25 #include "gaccess.h"
26 
27 struct vsie_page {
28 	struct kvm_s390_sie_block scb_s;	/* 0x0000 */
29 	/* the pinned originial scb */
30 	struct kvm_s390_sie_block *scb_o;	/* 0x0200 */
31 	/* the shadow gmap in use by the vsie_page */
32 	struct gmap *gmap;			/* 0x0208 */
33 	/* address of the last reported fault to guest2 */
34 	unsigned long fault_addr;		/* 0x0210 */
35 	__u8 reserved[0x0700 - 0x0218];		/* 0x0218 */
36 	struct kvm_s390_crypto_cb crycb;	/* 0x0700 */
37 	__u8 fac[S390_ARCH_FAC_LIST_SIZE_BYTE];	/* 0x0800 */
38 } __packed;
39 
40 /* trigger a validity icpt for the given scb */
41 static int set_validity_icpt(struct kvm_s390_sie_block *scb,
42 			     __u16 reason_code)
43 {
44 	scb->ipa = 0x1000;
45 	scb->ipb = ((__u32) reason_code) << 16;
46 	scb->icptcode = ICPT_VALIDITY;
47 	return 1;
48 }
49 
50 /* mark the prefix as unmapped, this will block the VSIE */
51 static void prefix_unmapped(struct vsie_page *vsie_page)
52 {
53 	atomic_or(PROG_REQUEST, &vsie_page->scb_s.prog20);
54 }
55 
56 /* mark the prefix as unmapped and wait until the VSIE has been left */
57 static void prefix_unmapped_sync(struct vsie_page *vsie_page)
58 {
59 	prefix_unmapped(vsie_page);
60 	if (vsie_page->scb_s.prog0c & PROG_IN_SIE)
61 		atomic_or(CPUSTAT_STOP_INT, &vsie_page->scb_s.cpuflags);
62 	while (vsie_page->scb_s.prog0c & PROG_IN_SIE)
63 		cpu_relax();
64 }
65 
66 /* mark the prefix as mapped, this will allow the VSIE to run */
67 static void prefix_mapped(struct vsie_page *vsie_page)
68 {
69 	atomic_andnot(PROG_REQUEST, &vsie_page->scb_s.prog20);
70 }
71 
72 /* test if the prefix is mapped into the gmap shadow */
73 static int prefix_is_mapped(struct vsie_page *vsie_page)
74 {
75 	return !(atomic_read(&vsie_page->scb_s.prog20) & PROG_REQUEST);
76 }
77 
78 /* copy the updated intervention request bits into the shadow scb */
79 static void update_intervention_requests(struct vsie_page *vsie_page)
80 {
81 	const int bits = CPUSTAT_STOP_INT | CPUSTAT_IO_INT | CPUSTAT_EXT_INT;
82 	int cpuflags;
83 
84 	cpuflags = atomic_read(&vsie_page->scb_o->cpuflags);
85 	atomic_andnot(bits, &vsie_page->scb_s.cpuflags);
86 	atomic_or(cpuflags & bits, &vsie_page->scb_s.cpuflags);
87 }
88 
89 /* shadow (filter and validate) the cpuflags  */
90 static int prepare_cpuflags(struct kvm_vcpu *vcpu, struct vsie_page *vsie_page)
91 {
92 	struct kvm_s390_sie_block *scb_s = &vsie_page->scb_s;
93 	struct kvm_s390_sie_block *scb_o = vsie_page->scb_o;
94 	int newflags, cpuflags = atomic_read(&scb_o->cpuflags);
95 
96 	/* we don't allow ESA/390 guests */
97 	if (!(cpuflags & CPUSTAT_ZARCH))
98 		return set_validity_icpt(scb_s, 0x0001U);
99 
100 	if (cpuflags & (CPUSTAT_RRF | CPUSTAT_MCDS))
101 		return set_validity_icpt(scb_s, 0x0001U);
102 	else if (cpuflags & (CPUSTAT_SLSV | CPUSTAT_SLSR))
103 		return set_validity_icpt(scb_s, 0x0007U);
104 
105 	/* intervention requests will be set later */
106 	newflags = CPUSTAT_ZARCH;
107 	if (cpuflags & CPUSTAT_GED && test_kvm_facility(vcpu->kvm, 8))
108 		newflags |= CPUSTAT_GED;
109 	if (cpuflags & CPUSTAT_GED2 && test_kvm_facility(vcpu->kvm, 78)) {
110 		if (cpuflags & CPUSTAT_GED)
111 			return set_validity_icpt(scb_s, 0x0001U);
112 		newflags |= CPUSTAT_GED2;
113 	}
114 	if (test_kvm_cpu_feat(vcpu->kvm, KVM_S390_VM_CPU_FEAT_GPERE))
115 		newflags |= cpuflags & CPUSTAT_P;
116 	if (test_kvm_cpu_feat(vcpu->kvm, KVM_S390_VM_CPU_FEAT_GSLS))
117 		newflags |= cpuflags & CPUSTAT_SM;
118 	if (test_kvm_cpu_feat(vcpu->kvm, KVM_S390_VM_CPU_FEAT_IBS))
119 		newflags |= cpuflags & CPUSTAT_IBS;
120 
121 	atomic_set(&scb_s->cpuflags, newflags);
122 	return 0;
123 }
124 
125 /*
126  * Create a shadow copy of the crycb block and setup key wrapping, if
127  * requested for guest 3 and enabled for guest 2.
128  *
129  * We only accept format-1 (no AP in g2), but convert it into format-2
130  * There is nothing to do for format-0.
131  *
132  * Returns: - 0 if shadowed or nothing to do
133  *          - > 0 if control has to be given to guest 2
134  */
135 static int shadow_crycb(struct kvm_vcpu *vcpu, struct vsie_page *vsie_page)
136 {
137 	struct kvm_s390_sie_block *scb_s = &vsie_page->scb_s;
138 	struct kvm_s390_sie_block *scb_o = vsie_page->scb_o;
139 	u32 crycb_addr = scb_o->crycbd & 0x7ffffff8U;
140 	unsigned long *b1, *b2;
141 	u8 ecb3_flags;
142 
143 	scb_s->crycbd = 0;
144 	if (!(scb_o->crycbd & vcpu->arch.sie_block->crycbd & CRYCB_FORMAT1))
145 		return 0;
146 	/* format-1 is supported with message-security-assist extension 3 */
147 	if (!test_kvm_facility(vcpu->kvm, 76))
148 		return 0;
149 	/* we may only allow it if enabled for guest 2 */
150 	ecb3_flags = scb_o->ecb3 & vcpu->arch.sie_block->ecb3 &
151 		     (ECB3_AES | ECB3_DEA);
152 	if (!ecb3_flags)
153 		return 0;
154 
155 	if ((crycb_addr & PAGE_MASK) != ((crycb_addr + 128) & PAGE_MASK))
156 		return set_validity_icpt(scb_s, 0x003CU);
157 	else if (!crycb_addr)
158 		return set_validity_icpt(scb_s, 0x0039U);
159 
160 	/* copy only the wrapping keys */
161 	if (read_guest_real(vcpu, crycb_addr + 72, &vsie_page->crycb, 56))
162 		return set_validity_icpt(scb_s, 0x0035U);
163 
164 	scb_s->ecb3 |= ecb3_flags;
165 	scb_s->crycbd = ((__u32)(__u64) &vsie_page->crycb) | CRYCB_FORMAT1 |
166 			CRYCB_FORMAT2;
167 
168 	/* xor both blocks in one run */
169 	b1 = (unsigned long *) vsie_page->crycb.dea_wrapping_key_mask;
170 	b2 = (unsigned long *)
171 			    vcpu->kvm->arch.crypto.crycb->dea_wrapping_key_mask;
172 	/* as 56%8 == 0, bitmap_xor won't overwrite any data */
173 	bitmap_xor(b1, b1, b2, BITS_PER_BYTE * 56);
174 	return 0;
175 }
176 
177 /* shadow (round up/down) the ibc to avoid validity icpt */
178 static void prepare_ibc(struct kvm_vcpu *vcpu, struct vsie_page *vsie_page)
179 {
180 	struct kvm_s390_sie_block *scb_s = &vsie_page->scb_s;
181 	struct kvm_s390_sie_block *scb_o = vsie_page->scb_o;
182 	__u64 min_ibc = (sclp.ibc >> 16) & 0x0fffU;
183 
184 	scb_s->ibc = 0;
185 	/* ibc installed in g2 and requested for g3 */
186 	if (vcpu->kvm->arch.model.ibc && (scb_o->ibc & 0x0fffU)) {
187 		scb_s->ibc = scb_o->ibc & 0x0fffU;
188 		/* takte care of the minimum ibc level of the machine */
189 		if (scb_s->ibc < min_ibc)
190 			scb_s->ibc = min_ibc;
191 		/* take care of the maximum ibc level set for the guest */
192 		if (scb_s->ibc > vcpu->kvm->arch.model.ibc)
193 			scb_s->ibc = vcpu->kvm->arch.model.ibc;
194 	}
195 }
196 
197 /* unshadow the scb, copying parameters back to the real scb */
198 static void unshadow_scb(struct kvm_vcpu *vcpu, struct vsie_page *vsie_page)
199 {
200 	struct kvm_s390_sie_block *scb_s = &vsie_page->scb_s;
201 	struct kvm_s390_sie_block *scb_o = vsie_page->scb_o;
202 
203 	/* interception */
204 	scb_o->icptcode = scb_s->icptcode;
205 	scb_o->icptstatus = scb_s->icptstatus;
206 	scb_o->ipa = scb_s->ipa;
207 	scb_o->ipb = scb_s->ipb;
208 	scb_o->gbea = scb_s->gbea;
209 
210 	/* timer */
211 	scb_o->cputm = scb_s->cputm;
212 	scb_o->ckc = scb_s->ckc;
213 	scb_o->todpr = scb_s->todpr;
214 
215 	/* guest state */
216 	scb_o->gpsw = scb_s->gpsw;
217 	scb_o->gg14 = scb_s->gg14;
218 	scb_o->gg15 = scb_s->gg15;
219 	memcpy(scb_o->gcr, scb_s->gcr, 128);
220 	scb_o->pp = scb_s->pp;
221 
222 	/* interrupt intercept */
223 	switch (scb_s->icptcode) {
224 	case ICPT_PROGI:
225 	case ICPT_INSTPROGI:
226 	case ICPT_EXTINT:
227 		memcpy((void *)((u64)scb_o + 0xc0),
228 		       (void *)((u64)scb_s + 0xc0), 0xf0 - 0xc0);
229 		break;
230 	case ICPT_PARTEXEC:
231 		/* MVPG only */
232 		memcpy((void *)((u64)scb_o + 0xc0),
233 		       (void *)((u64)scb_s + 0xc0), 0xd0 - 0xc0);
234 		break;
235 	}
236 
237 	if (scb_s->ihcpu != 0xffffU)
238 		scb_o->ihcpu = scb_s->ihcpu;
239 }
240 
241 /*
242  * Setup the shadow scb by copying and checking the relevant parts of the g2
243  * provided scb.
244  *
245  * Returns: - 0 if the scb has been shadowed
246  *          - > 0 if control has to be given to guest 2
247  */
248 static int shadow_scb(struct kvm_vcpu *vcpu, struct vsie_page *vsie_page)
249 {
250 	struct kvm_s390_sie_block *scb_o = vsie_page->scb_o;
251 	struct kvm_s390_sie_block *scb_s = &vsie_page->scb_s;
252 	bool had_tx = scb_s->ecb & 0x10U;
253 	unsigned long new_mso = 0;
254 	int rc;
255 
256 	/* make sure we don't have any leftovers when reusing the scb */
257 	scb_s->icptcode = 0;
258 	scb_s->eca = 0;
259 	scb_s->ecb = 0;
260 	scb_s->ecb2 = 0;
261 	scb_s->ecb3 = 0;
262 	scb_s->ecd = 0;
263 	scb_s->fac = 0;
264 
265 	rc = prepare_cpuflags(vcpu, vsie_page);
266 	if (rc)
267 		goto out;
268 
269 	/* timer */
270 	scb_s->cputm = scb_o->cputm;
271 	scb_s->ckc = scb_o->ckc;
272 	scb_s->todpr = scb_o->todpr;
273 	scb_s->epoch = scb_o->epoch;
274 
275 	/* guest state */
276 	scb_s->gpsw = scb_o->gpsw;
277 	scb_s->gg14 = scb_o->gg14;
278 	scb_s->gg15 = scb_o->gg15;
279 	memcpy(scb_s->gcr, scb_o->gcr, 128);
280 	scb_s->pp = scb_o->pp;
281 
282 	/* interception / execution handling */
283 	scb_s->gbea = scb_o->gbea;
284 	scb_s->lctl = scb_o->lctl;
285 	scb_s->svcc = scb_o->svcc;
286 	scb_s->ictl = scb_o->ictl;
287 	/*
288 	 * SKEY handling functions can't deal with false setting of PTE invalid
289 	 * bits. Therefore we cannot provide interpretation and would later
290 	 * have to provide own emulation handlers.
291 	 */
292 	scb_s->ictl |= ICTL_ISKE | ICTL_SSKE | ICTL_RRBE;
293 	scb_s->icpua = scb_o->icpua;
294 
295 	if (!(atomic_read(&scb_s->cpuflags) & CPUSTAT_SM))
296 		new_mso = scb_o->mso & 0xfffffffffff00000UL;
297 	/* if the hva of the prefix changes, we have to remap the prefix */
298 	if (scb_s->mso != new_mso || scb_s->prefix != scb_o->prefix)
299 		prefix_unmapped(vsie_page);
300 	 /* SIE will do mso/msl validity and exception checks for us */
301 	scb_s->msl = scb_o->msl & 0xfffffffffff00000UL;
302 	scb_s->mso = new_mso;
303 	scb_s->prefix = scb_o->prefix;
304 
305 	/* We have to definetly flush the tlb if this scb never ran */
306 	if (scb_s->ihcpu != 0xffffU)
307 		scb_s->ihcpu = scb_o->ihcpu;
308 
309 	/* MVPG and Protection Exception Interpretation are always available */
310 	scb_s->eca |= scb_o->eca & 0x01002000U;
311 	/* Host-protection-interruption introduced with ESOP */
312 	if (test_kvm_cpu_feat(vcpu->kvm, KVM_S390_VM_CPU_FEAT_ESOP))
313 		scb_s->ecb |= scb_o->ecb & 0x02U;
314 	/* transactional execution */
315 	if (test_kvm_facility(vcpu->kvm, 73)) {
316 		/* remap the prefix is tx is toggled on */
317 		if ((scb_o->ecb & 0x10U) && !had_tx)
318 			prefix_unmapped(vsie_page);
319 		scb_s->ecb |= scb_o->ecb & 0x10U;
320 	}
321 	/* SIMD */
322 	if (test_kvm_facility(vcpu->kvm, 129)) {
323 		scb_s->eca |= scb_o->eca & 0x00020000U;
324 		scb_s->ecd |= scb_o->ecd & 0x20000000U;
325 	}
326 	/* Run-time-Instrumentation */
327 	if (test_kvm_facility(vcpu->kvm, 64))
328 		scb_s->ecb3 |= scb_o->ecb3 & 0x01U;
329 	/* Instruction Execution Prevention */
330 	if (test_kvm_facility(vcpu->kvm, 130))
331 		scb_s->ecb2 |= scb_o->ecb2 & 0x20U;
332 	if (test_kvm_cpu_feat(vcpu->kvm, KVM_S390_VM_CPU_FEAT_SIIF))
333 		scb_s->eca |= scb_o->eca & 0x00000001U;
334 	if (test_kvm_cpu_feat(vcpu->kvm, KVM_S390_VM_CPU_FEAT_IB))
335 		scb_s->eca |= scb_o->eca & 0x40000000U;
336 	if (test_kvm_cpu_feat(vcpu->kvm, KVM_S390_VM_CPU_FEAT_CEI))
337 		scb_s->eca |= scb_o->eca & 0x80000000U;
338 
339 	prepare_ibc(vcpu, vsie_page);
340 	rc = shadow_crycb(vcpu, vsie_page);
341 out:
342 	if (rc)
343 		unshadow_scb(vcpu, vsie_page);
344 	return rc;
345 }
346 
347 void kvm_s390_vsie_gmap_notifier(struct gmap *gmap, unsigned long start,
348 				 unsigned long end)
349 {
350 	struct kvm *kvm = gmap->private;
351 	struct vsie_page *cur;
352 	unsigned long prefix;
353 	struct page *page;
354 	int i;
355 
356 	if (!gmap_is_shadow(gmap))
357 		return;
358 	if (start >= 1UL << 31)
359 		/* We are only interested in prefix pages */
360 		return;
361 
362 	/*
363 	 * Only new shadow blocks are added to the list during runtime,
364 	 * therefore we can safely reference them all the time.
365 	 */
366 	for (i = 0; i < kvm->arch.vsie.page_count; i++) {
367 		page = READ_ONCE(kvm->arch.vsie.pages[i]);
368 		if (!page)
369 			continue;
370 		cur = page_to_virt(page);
371 		if (READ_ONCE(cur->gmap) != gmap)
372 			continue;
373 		prefix = cur->scb_s.prefix << GUEST_PREFIX_SHIFT;
374 		/* with mso/msl, the prefix lies at an offset */
375 		prefix += cur->scb_s.mso;
376 		if (prefix <= end && start <= prefix + 2 * PAGE_SIZE - 1)
377 			prefix_unmapped_sync(cur);
378 	}
379 }
380 
381 /*
382  * Map the first prefix page and if tx is enabled also the second prefix page.
383  *
384  * The prefix will be protected, a gmap notifier will inform about unmaps.
385  * The shadow scb must not be executed until the prefix is remapped, this is
386  * guaranteed by properly handling PROG_REQUEST.
387  *
388  * Returns: - 0 on if successfully mapped or already mapped
389  *          - > 0 if control has to be given to guest 2
390  *          - -EAGAIN if the caller can retry immediately
391  *          - -ENOMEM if out of memory
392  */
393 static int map_prefix(struct kvm_vcpu *vcpu, struct vsie_page *vsie_page)
394 {
395 	struct kvm_s390_sie_block *scb_s = &vsie_page->scb_s;
396 	u64 prefix = scb_s->prefix << GUEST_PREFIX_SHIFT;
397 	int rc;
398 
399 	if (prefix_is_mapped(vsie_page))
400 		return 0;
401 
402 	/* mark it as mapped so we can catch any concurrent unmappers */
403 	prefix_mapped(vsie_page);
404 
405 	/* with mso/msl, the prefix lies at offset *mso* */
406 	prefix += scb_s->mso;
407 
408 	rc = kvm_s390_shadow_fault(vcpu, vsie_page->gmap, prefix);
409 	if (!rc && (scb_s->ecb & 0x10U))
410 		rc = kvm_s390_shadow_fault(vcpu, vsie_page->gmap,
411 					   prefix + PAGE_SIZE);
412 	/*
413 	 * We don't have to mprotect, we will be called for all unshadows.
414 	 * SIE will detect if protection applies and trigger a validity.
415 	 */
416 	if (rc)
417 		prefix_unmapped(vsie_page);
418 	if (rc > 0 || rc == -EFAULT)
419 		rc = set_validity_icpt(scb_s, 0x0037U);
420 	return rc;
421 }
422 
423 /*
424  * Pin the guest page given by gpa and set hpa to the pinned host address.
425  * Will always be pinned writable.
426  *
427  * Returns: - 0 on success
428  *          - -EINVAL if the gpa is not valid guest storage
429  *          - -ENOMEM if out of memory
430  */
431 static int pin_guest_page(struct kvm *kvm, gpa_t gpa, hpa_t *hpa)
432 {
433 	struct page *page;
434 	hva_t hva;
435 	int rc;
436 
437 	hva = gfn_to_hva(kvm, gpa_to_gfn(gpa));
438 	if (kvm_is_error_hva(hva))
439 		return -EINVAL;
440 	rc = get_user_pages_fast(hva, 1, 1, &page);
441 	if (rc < 0)
442 		return rc;
443 	else if (rc != 1)
444 		return -ENOMEM;
445 	*hpa = (hpa_t) page_to_virt(page) + (gpa & ~PAGE_MASK);
446 	return 0;
447 }
448 
449 /* Unpins a page previously pinned via pin_guest_page, marking it as dirty. */
450 static void unpin_guest_page(struct kvm *kvm, gpa_t gpa, hpa_t hpa)
451 {
452 	struct page *page;
453 
454 	page = virt_to_page(hpa);
455 	set_page_dirty_lock(page);
456 	put_page(page);
457 	/* mark the page always as dirty for migration */
458 	mark_page_dirty(kvm, gpa_to_gfn(gpa));
459 }
460 
461 /* unpin all blocks previously pinned by pin_blocks(), marking them dirty */
462 static void unpin_blocks(struct kvm_vcpu *vcpu, struct vsie_page *vsie_page)
463 {
464 	struct kvm_s390_sie_block *scb_o = vsie_page->scb_o;
465 	struct kvm_s390_sie_block *scb_s = &vsie_page->scb_s;
466 	hpa_t hpa;
467 	gpa_t gpa;
468 
469 	hpa = (u64) scb_s->scaoh << 32 | scb_s->scaol;
470 	if (hpa) {
471 		gpa = scb_o->scaol & ~0xfUL;
472 		if (test_kvm_cpu_feat(vcpu->kvm, KVM_S390_VM_CPU_FEAT_64BSCAO))
473 			gpa |= (u64) scb_o->scaoh << 32;
474 		unpin_guest_page(vcpu->kvm, gpa, hpa);
475 		scb_s->scaol = 0;
476 		scb_s->scaoh = 0;
477 	}
478 
479 	hpa = scb_s->itdba;
480 	if (hpa) {
481 		gpa = scb_o->itdba & ~0xffUL;
482 		unpin_guest_page(vcpu->kvm, gpa, hpa);
483 		scb_s->itdba = 0;
484 	}
485 
486 	hpa = scb_s->gvrd;
487 	if (hpa) {
488 		gpa = scb_o->gvrd & ~0x1ffUL;
489 		unpin_guest_page(vcpu->kvm, gpa, hpa);
490 		scb_s->gvrd = 0;
491 	}
492 
493 	hpa = scb_s->riccbd;
494 	if (hpa) {
495 		gpa = scb_o->riccbd & ~0x3fUL;
496 		unpin_guest_page(vcpu->kvm, gpa, hpa);
497 		scb_s->riccbd = 0;
498 	}
499 }
500 
501 /*
502  * Instead of shadowing some blocks, we can simply forward them because the
503  * addresses in the scb are 64 bit long.
504  *
505  * This works as long as the data lies in one page. If blocks ever exceed one
506  * page, we have to fall back to shadowing.
507  *
508  * As we reuse the sca, the vcpu pointers contained in it are invalid. We must
509  * therefore not enable any facilities that access these pointers (e.g. SIGPIF).
510  *
511  * Returns: - 0 if all blocks were pinned.
512  *          - > 0 if control has to be given to guest 2
513  *          - -ENOMEM if out of memory
514  */
515 static int pin_blocks(struct kvm_vcpu *vcpu, struct vsie_page *vsie_page)
516 {
517 	struct kvm_s390_sie_block *scb_o = vsie_page->scb_o;
518 	struct kvm_s390_sie_block *scb_s = &vsie_page->scb_s;
519 	hpa_t hpa;
520 	gpa_t gpa;
521 	int rc = 0;
522 
523 	gpa = scb_o->scaol & ~0xfUL;
524 	if (test_kvm_cpu_feat(vcpu->kvm, KVM_S390_VM_CPU_FEAT_64BSCAO))
525 		gpa |= (u64) scb_o->scaoh << 32;
526 	if (gpa) {
527 		if (!(gpa & ~0x1fffUL))
528 			rc = set_validity_icpt(scb_s, 0x0038U);
529 		else if ((gpa & ~0x1fffUL) == kvm_s390_get_prefix(vcpu))
530 			rc = set_validity_icpt(scb_s, 0x0011U);
531 		else if ((gpa & PAGE_MASK) !=
532 			 ((gpa + sizeof(struct bsca_block) - 1) & PAGE_MASK))
533 			rc = set_validity_icpt(scb_s, 0x003bU);
534 		if (!rc) {
535 			rc = pin_guest_page(vcpu->kvm, gpa, &hpa);
536 			if (rc == -EINVAL)
537 				rc = set_validity_icpt(scb_s, 0x0034U);
538 		}
539 		if (rc)
540 			goto unpin;
541 		scb_s->scaoh = (u32)((u64)hpa >> 32);
542 		scb_s->scaol = (u32)(u64)hpa;
543 	}
544 
545 	gpa = scb_o->itdba & ~0xffUL;
546 	if (gpa && (scb_s->ecb & 0x10U)) {
547 		if (!(gpa & ~0x1fffU)) {
548 			rc = set_validity_icpt(scb_s, 0x0080U);
549 			goto unpin;
550 		}
551 		/* 256 bytes cannot cross page boundaries */
552 		rc = pin_guest_page(vcpu->kvm, gpa, &hpa);
553 		if (rc == -EINVAL)
554 			rc = set_validity_icpt(scb_s, 0x0080U);
555 		if (rc)
556 			goto unpin;
557 		scb_s->itdba = hpa;
558 	}
559 
560 	gpa = scb_o->gvrd & ~0x1ffUL;
561 	if (gpa && (scb_s->eca & 0x00020000U) &&
562 	    !(scb_s->ecd & 0x20000000U)) {
563 		if (!(gpa & ~0x1fffUL)) {
564 			rc = set_validity_icpt(scb_s, 0x1310U);
565 			goto unpin;
566 		}
567 		/*
568 		 * 512 bytes vector registers cannot cross page boundaries
569 		 * if this block gets bigger, we have to shadow it.
570 		 */
571 		rc = pin_guest_page(vcpu->kvm, gpa, &hpa);
572 		if (rc == -EINVAL)
573 			rc = set_validity_icpt(scb_s, 0x1310U);
574 		if (rc)
575 			goto unpin;
576 		scb_s->gvrd = hpa;
577 	}
578 
579 	gpa = scb_o->riccbd & ~0x3fUL;
580 	if (gpa && (scb_s->ecb3 & 0x01U)) {
581 		if (!(gpa & ~0x1fffUL)) {
582 			rc = set_validity_icpt(scb_s, 0x0043U);
583 			goto unpin;
584 		}
585 		/* 64 bytes cannot cross page boundaries */
586 		rc = pin_guest_page(vcpu->kvm, gpa, &hpa);
587 		if (rc == -EINVAL)
588 			rc = set_validity_icpt(scb_s, 0x0043U);
589 		/* Validity 0x0044 will be checked by SIE */
590 		if (rc)
591 			goto unpin;
592 		scb_s->riccbd = hpa;
593 	}
594 	return 0;
595 unpin:
596 	unpin_blocks(vcpu, vsie_page);
597 	return rc;
598 }
599 
600 /* unpin the scb provided by guest 2, marking it as dirty */
601 static void unpin_scb(struct kvm_vcpu *vcpu, struct vsie_page *vsie_page,
602 		      gpa_t gpa)
603 {
604 	hpa_t hpa = (hpa_t) vsie_page->scb_o;
605 
606 	if (hpa)
607 		unpin_guest_page(vcpu->kvm, gpa, hpa);
608 	vsie_page->scb_o = NULL;
609 }
610 
611 /*
612  * Pin the scb at gpa provided by guest 2 at vsie_page->scb_o.
613  *
614  * Returns: - 0 if the scb was pinned.
615  *          - > 0 if control has to be given to guest 2
616  *          - -ENOMEM if out of memory
617  */
618 static int pin_scb(struct kvm_vcpu *vcpu, struct vsie_page *vsie_page,
619 		   gpa_t gpa)
620 {
621 	hpa_t hpa;
622 	int rc;
623 
624 	rc = pin_guest_page(vcpu->kvm, gpa, &hpa);
625 	if (rc == -EINVAL) {
626 		rc = kvm_s390_inject_program_int(vcpu, PGM_ADDRESSING);
627 		if (!rc)
628 			rc = 1;
629 	}
630 	if (!rc)
631 		vsie_page->scb_o = (struct kvm_s390_sie_block *) hpa;
632 	return rc;
633 }
634 
635 /*
636  * Inject a fault into guest 2.
637  *
638  * Returns: - > 0 if control has to be given to guest 2
639  *            < 0 if an error occurred during injection.
640  */
641 static int inject_fault(struct kvm_vcpu *vcpu, __u16 code, __u64 vaddr,
642 			bool write_flag)
643 {
644 	struct kvm_s390_pgm_info pgm = {
645 		.code = code,
646 		.trans_exc_code =
647 			/* 0-51: virtual address */
648 			(vaddr & 0xfffffffffffff000UL) |
649 			/* 52-53: store / fetch */
650 			(((unsigned int) !write_flag) + 1) << 10,
651 			/* 62-63: asce id (alway primary == 0) */
652 		.exc_access_id = 0, /* always primary */
653 		.op_access_id = 0, /* not MVPG */
654 	};
655 	int rc;
656 
657 	if (code == PGM_PROTECTION)
658 		pgm.trans_exc_code |= 0x4UL;
659 
660 	rc = kvm_s390_inject_prog_irq(vcpu, &pgm);
661 	return rc ? rc : 1;
662 }
663 
664 /*
665  * Handle a fault during vsie execution on a gmap shadow.
666  *
667  * Returns: - 0 if the fault was resolved
668  *          - > 0 if control has to be given to guest 2
669  *          - < 0 if an error occurred
670  */
671 static int handle_fault(struct kvm_vcpu *vcpu, struct vsie_page *vsie_page)
672 {
673 	int rc;
674 
675 	if (current->thread.gmap_int_code == PGM_PROTECTION)
676 		/* we can directly forward all protection exceptions */
677 		return inject_fault(vcpu, PGM_PROTECTION,
678 				    current->thread.gmap_addr, 1);
679 
680 	rc = kvm_s390_shadow_fault(vcpu, vsie_page->gmap,
681 				   current->thread.gmap_addr);
682 	if (rc > 0) {
683 		rc = inject_fault(vcpu, rc,
684 				  current->thread.gmap_addr,
685 				  current->thread.gmap_write_flag);
686 		if (rc >= 0)
687 			vsie_page->fault_addr = current->thread.gmap_addr;
688 	}
689 	return rc;
690 }
691 
692 /*
693  * Retry the previous fault that required guest 2 intervention. This avoids
694  * one superfluous SIE re-entry and direct exit.
695  *
696  * Will ignore any errors. The next SIE fault will do proper fault handling.
697  */
698 static void handle_last_fault(struct kvm_vcpu *vcpu,
699 			      struct vsie_page *vsie_page)
700 {
701 	if (vsie_page->fault_addr)
702 		kvm_s390_shadow_fault(vcpu, vsie_page->gmap,
703 				      vsie_page->fault_addr);
704 	vsie_page->fault_addr = 0;
705 }
706 
707 static inline void clear_vsie_icpt(struct vsie_page *vsie_page)
708 {
709 	vsie_page->scb_s.icptcode = 0;
710 }
711 
712 /* rewind the psw and clear the vsie icpt, so we can retry execution */
713 static void retry_vsie_icpt(struct vsie_page *vsie_page)
714 {
715 	struct kvm_s390_sie_block *scb_s = &vsie_page->scb_s;
716 	int ilen = insn_length(scb_s->ipa >> 8);
717 
718 	/* take care of EXECUTE instructions */
719 	if (scb_s->icptstatus & 1) {
720 		ilen = (scb_s->icptstatus >> 4) & 0x6;
721 		if (!ilen)
722 			ilen = 4;
723 	}
724 	scb_s->gpsw.addr = __rewind_psw(scb_s->gpsw, ilen);
725 	clear_vsie_icpt(vsie_page);
726 }
727 
728 /*
729  * Try to shadow + enable the guest 2 provided facility list.
730  * Retry instruction execution if enabled for and provided by guest 2.
731  *
732  * Returns: - 0 if handled (retry or guest 2 icpt)
733  *          - > 0 if control has to be given to guest 2
734  */
735 static int handle_stfle(struct kvm_vcpu *vcpu, struct vsie_page *vsie_page)
736 {
737 	struct kvm_s390_sie_block *scb_s = &vsie_page->scb_s;
738 	__u32 fac = vsie_page->scb_o->fac & 0x7ffffff8U;
739 
740 	if (fac && test_kvm_facility(vcpu->kvm, 7)) {
741 		retry_vsie_icpt(vsie_page);
742 		if (read_guest_real(vcpu, fac, &vsie_page->fac,
743 				    sizeof(vsie_page->fac)))
744 			return set_validity_icpt(scb_s, 0x1090U);
745 		scb_s->fac = (__u32)(__u64) &vsie_page->fac;
746 	}
747 	return 0;
748 }
749 
750 /*
751  * Run the vsie on a shadow scb and a shadow gmap, without any further
752  * sanity checks, handling SIE faults.
753  *
754  * Returns: - 0 everything went fine
755  *          - > 0 if control has to be given to guest 2
756  *          - < 0 if an error occurred
757  */
758 static int do_vsie_run(struct kvm_vcpu *vcpu, struct vsie_page *vsie_page)
759 {
760 	struct kvm_s390_sie_block *scb_s = &vsie_page->scb_s;
761 	struct kvm_s390_sie_block *scb_o = vsie_page->scb_o;
762 	int rc;
763 
764 	handle_last_fault(vcpu, vsie_page);
765 
766 	if (need_resched())
767 		schedule();
768 	if (test_cpu_flag(CIF_MCCK_PENDING))
769 		s390_handle_mcck();
770 
771 	srcu_read_unlock(&vcpu->kvm->srcu, vcpu->srcu_idx);
772 	local_irq_disable();
773 	guest_enter_irqoff();
774 	local_irq_enable();
775 
776 	rc = sie64a(scb_s, vcpu->run->s.regs.gprs);
777 
778 	local_irq_disable();
779 	guest_exit_irqoff();
780 	local_irq_enable();
781 	vcpu->srcu_idx = srcu_read_lock(&vcpu->kvm->srcu);
782 
783 	if (rc > 0)
784 		rc = 0; /* we could still have an icpt */
785 	else if (rc == -EFAULT)
786 		return handle_fault(vcpu, vsie_page);
787 
788 	switch (scb_s->icptcode) {
789 	case ICPT_INST:
790 		if (scb_s->ipa == 0xb2b0)
791 			rc = handle_stfle(vcpu, vsie_page);
792 		break;
793 	case ICPT_STOP:
794 		/* stop not requested by g2 - must have been a kick */
795 		if (!(atomic_read(&scb_o->cpuflags) & CPUSTAT_STOP_INT))
796 			clear_vsie_icpt(vsie_page);
797 		break;
798 	case ICPT_VALIDITY:
799 		if ((scb_s->ipa & 0xf000) != 0xf000)
800 			scb_s->ipa += 0x1000;
801 		break;
802 	}
803 	return rc;
804 }
805 
806 static void release_gmap_shadow(struct vsie_page *vsie_page)
807 {
808 	if (vsie_page->gmap)
809 		gmap_put(vsie_page->gmap);
810 	WRITE_ONCE(vsie_page->gmap, NULL);
811 	prefix_unmapped(vsie_page);
812 }
813 
814 static int acquire_gmap_shadow(struct kvm_vcpu *vcpu,
815 			       struct vsie_page *vsie_page)
816 {
817 	unsigned long asce;
818 	union ctlreg0 cr0;
819 	struct gmap *gmap;
820 	int edat;
821 
822 	asce = vcpu->arch.sie_block->gcr[1];
823 	cr0.val = vcpu->arch.sie_block->gcr[0];
824 	edat = cr0.edat && test_kvm_facility(vcpu->kvm, 8);
825 	edat += edat && test_kvm_facility(vcpu->kvm, 78);
826 
827 	/*
828 	 * ASCE or EDAT could have changed since last icpt, or the gmap
829 	 * we're holding has been unshadowed. If the gmap is still valid,
830 	 * we can safely reuse it.
831 	 */
832 	if (vsie_page->gmap && gmap_shadow_valid(vsie_page->gmap, asce, edat))
833 		return 0;
834 
835 	/* release the old shadow - if any, and mark the prefix as unmapped */
836 	release_gmap_shadow(vsie_page);
837 	gmap = gmap_shadow(vcpu->arch.gmap, asce, edat);
838 	if (IS_ERR(gmap))
839 		return PTR_ERR(gmap);
840 	gmap->private = vcpu->kvm;
841 	WRITE_ONCE(vsie_page->gmap, gmap);
842 	return 0;
843 }
844 
845 /*
846  * Register the shadow scb at the VCPU, e.g. for kicking out of vsie.
847  */
848 static void register_shadow_scb(struct kvm_vcpu *vcpu,
849 				struct vsie_page *vsie_page)
850 {
851 	struct kvm_s390_sie_block *scb_s = &vsie_page->scb_s;
852 
853 	WRITE_ONCE(vcpu->arch.vsie_block, &vsie_page->scb_s);
854 	/*
855 	 * External calls have to lead to a kick of the vcpu and
856 	 * therefore the vsie -> Simulate Wait state.
857 	 */
858 	atomic_or(CPUSTAT_WAIT, &vcpu->arch.sie_block->cpuflags);
859 	/*
860 	 * We have to adjust the g3 epoch by the g2 epoch. The epoch will
861 	 * automatically be adjusted on tod clock changes via kvm_sync_clock.
862 	 */
863 	preempt_disable();
864 	scb_s->epoch += vcpu->kvm->arch.epoch;
865 	preempt_enable();
866 }
867 
868 /*
869  * Unregister a shadow scb from a VCPU.
870  */
871 static void unregister_shadow_scb(struct kvm_vcpu *vcpu)
872 {
873 	atomic_andnot(CPUSTAT_WAIT, &vcpu->arch.sie_block->cpuflags);
874 	WRITE_ONCE(vcpu->arch.vsie_block, NULL);
875 }
876 
877 /*
878  * Run the vsie on a shadowed scb, managing the gmap shadow, handling
879  * prefix pages and faults.
880  *
881  * Returns: - 0 if no errors occurred
882  *          - > 0 if control has to be given to guest 2
883  *          - -ENOMEM if out of memory
884  */
885 static int vsie_run(struct kvm_vcpu *vcpu, struct vsie_page *vsie_page)
886 {
887 	struct kvm_s390_sie_block *scb_s = &vsie_page->scb_s;
888 	int rc = 0;
889 
890 	while (1) {
891 		rc = acquire_gmap_shadow(vcpu, vsie_page);
892 		if (!rc)
893 			rc = map_prefix(vcpu, vsie_page);
894 		if (!rc) {
895 			gmap_enable(vsie_page->gmap);
896 			update_intervention_requests(vsie_page);
897 			rc = do_vsie_run(vcpu, vsie_page);
898 			gmap_enable(vcpu->arch.gmap);
899 		}
900 		atomic_andnot(PROG_BLOCK_SIE, &scb_s->prog20);
901 
902 		if (rc == -EAGAIN)
903 			rc = 0;
904 		if (rc || scb_s->icptcode || signal_pending(current) ||
905 		    kvm_s390_vcpu_has_irq(vcpu, 0))
906 			break;
907 	}
908 
909 	if (rc == -EFAULT) {
910 		/*
911 		 * Addressing exceptions are always presentes as intercepts.
912 		 * As addressing exceptions are suppressing and our guest 3 PSW
913 		 * points at the responsible instruction, we have to
914 		 * forward the PSW and set the ilc. If we can't read guest 3
915 		 * instruction, we can use an arbitrary ilc. Let's always use
916 		 * ilen = 4 for now, so we can avoid reading in guest 3 virtual
917 		 * memory. (we could also fake the shadow so the hardware
918 		 * handles it).
919 		 */
920 		scb_s->icptcode = ICPT_PROGI;
921 		scb_s->iprcc = PGM_ADDRESSING;
922 		scb_s->pgmilc = 4;
923 		scb_s->gpsw.addr = __rewind_psw(scb_s->gpsw, 4);
924 	}
925 	return rc;
926 }
927 
928 /*
929  * Get or create a vsie page for a scb address.
930  *
931  * Returns: - address of a vsie page (cached or new one)
932  *          - NULL if the same scb address is already used by another VCPU
933  *          - ERR_PTR(-ENOMEM) if out of memory
934  */
935 static struct vsie_page *get_vsie_page(struct kvm *kvm, unsigned long addr)
936 {
937 	struct vsie_page *vsie_page;
938 	struct page *page;
939 	int nr_vcpus;
940 
941 	rcu_read_lock();
942 	page = radix_tree_lookup(&kvm->arch.vsie.addr_to_page, addr >> 9);
943 	rcu_read_unlock();
944 	if (page) {
945 		if (page_ref_inc_return(page) == 2)
946 			return page_to_virt(page);
947 		page_ref_dec(page);
948 	}
949 
950 	/*
951 	 * We want at least #online_vcpus shadows, so every VCPU can execute
952 	 * the VSIE in parallel.
953 	 */
954 	nr_vcpus = atomic_read(&kvm->online_vcpus);
955 
956 	mutex_lock(&kvm->arch.vsie.mutex);
957 	if (kvm->arch.vsie.page_count < nr_vcpus) {
958 		page = alloc_page(GFP_KERNEL | __GFP_ZERO | GFP_DMA);
959 		if (!page) {
960 			mutex_unlock(&kvm->arch.vsie.mutex);
961 			return ERR_PTR(-ENOMEM);
962 		}
963 		page_ref_inc(page);
964 		kvm->arch.vsie.pages[kvm->arch.vsie.page_count] = page;
965 		kvm->arch.vsie.page_count++;
966 	} else {
967 		/* reuse an existing entry that belongs to nobody */
968 		while (true) {
969 			page = kvm->arch.vsie.pages[kvm->arch.vsie.next];
970 			if (page_ref_inc_return(page) == 2)
971 				break;
972 			page_ref_dec(page);
973 			kvm->arch.vsie.next++;
974 			kvm->arch.vsie.next %= nr_vcpus;
975 		}
976 		radix_tree_delete(&kvm->arch.vsie.addr_to_page, page->index >> 9);
977 	}
978 	page->index = addr;
979 	/* double use of the same address */
980 	if (radix_tree_insert(&kvm->arch.vsie.addr_to_page, addr >> 9, page)) {
981 		page_ref_dec(page);
982 		mutex_unlock(&kvm->arch.vsie.mutex);
983 		return NULL;
984 	}
985 	mutex_unlock(&kvm->arch.vsie.mutex);
986 
987 	vsie_page = page_to_virt(page);
988 	memset(&vsie_page->scb_s, 0, sizeof(struct kvm_s390_sie_block));
989 	release_gmap_shadow(vsie_page);
990 	vsie_page->fault_addr = 0;
991 	vsie_page->scb_s.ihcpu = 0xffffU;
992 	return vsie_page;
993 }
994 
995 /* put a vsie page acquired via get_vsie_page */
996 static void put_vsie_page(struct kvm *kvm, struct vsie_page *vsie_page)
997 {
998 	struct page *page = pfn_to_page(__pa(vsie_page) >> PAGE_SHIFT);
999 
1000 	page_ref_dec(page);
1001 }
1002 
1003 int kvm_s390_handle_vsie(struct kvm_vcpu *vcpu)
1004 {
1005 	struct vsie_page *vsie_page;
1006 	unsigned long scb_addr;
1007 	int rc;
1008 
1009 	vcpu->stat.instruction_sie++;
1010 	if (!test_kvm_cpu_feat(vcpu->kvm, KVM_S390_VM_CPU_FEAT_SIEF2))
1011 		return -EOPNOTSUPP;
1012 	if (vcpu->arch.sie_block->gpsw.mask & PSW_MASK_PSTATE)
1013 		return kvm_s390_inject_program_int(vcpu, PGM_PRIVILEGED_OP);
1014 
1015 	BUILD_BUG_ON(sizeof(struct vsie_page) != 4096);
1016 	scb_addr = kvm_s390_get_base_disp_s(vcpu, NULL);
1017 
1018 	/* 512 byte alignment */
1019 	if (unlikely(scb_addr & 0x1ffUL))
1020 		return kvm_s390_inject_program_int(vcpu, PGM_SPECIFICATION);
1021 
1022 	if (signal_pending(current) || kvm_s390_vcpu_has_irq(vcpu, 0))
1023 		return 0;
1024 
1025 	vsie_page = get_vsie_page(vcpu->kvm, scb_addr);
1026 	if (IS_ERR(vsie_page))
1027 		return PTR_ERR(vsie_page);
1028 	else if (!vsie_page)
1029 		/* double use of sie control block - simply do nothing */
1030 		return 0;
1031 
1032 	rc = pin_scb(vcpu, vsie_page, scb_addr);
1033 	if (rc)
1034 		goto out_put;
1035 	rc = shadow_scb(vcpu, vsie_page);
1036 	if (rc)
1037 		goto out_unpin_scb;
1038 	rc = pin_blocks(vcpu, vsie_page);
1039 	if (rc)
1040 		goto out_unshadow;
1041 	register_shadow_scb(vcpu, vsie_page);
1042 	rc = vsie_run(vcpu, vsie_page);
1043 	unregister_shadow_scb(vcpu);
1044 	unpin_blocks(vcpu, vsie_page);
1045 out_unshadow:
1046 	unshadow_scb(vcpu, vsie_page);
1047 out_unpin_scb:
1048 	unpin_scb(vcpu, vsie_page, scb_addr);
1049 out_put:
1050 	put_vsie_page(vcpu->kvm, vsie_page);
1051 
1052 	return rc < 0 ? rc : 0;
1053 }
1054 
1055 /* Init the vsie data structures. To be called when a vm is initialized. */
1056 void kvm_s390_vsie_init(struct kvm *kvm)
1057 {
1058 	mutex_init(&kvm->arch.vsie.mutex);
1059 	INIT_RADIX_TREE(&kvm->arch.vsie.addr_to_page, GFP_KERNEL);
1060 }
1061 
1062 /* Destroy the vsie data structures. To be called when a vm is destroyed. */
1063 void kvm_s390_vsie_destroy(struct kvm *kvm)
1064 {
1065 	struct vsie_page *vsie_page;
1066 	struct page *page;
1067 	int i;
1068 
1069 	mutex_lock(&kvm->arch.vsie.mutex);
1070 	for (i = 0; i < kvm->arch.vsie.page_count; i++) {
1071 		page = kvm->arch.vsie.pages[i];
1072 		kvm->arch.vsie.pages[i] = NULL;
1073 		vsie_page = page_to_virt(page);
1074 		release_gmap_shadow(vsie_page);
1075 		/* free the radix tree entry */
1076 		radix_tree_delete(&kvm->arch.vsie.addr_to_page, page->index >> 9);
1077 		__free_page(page);
1078 	}
1079 	kvm->arch.vsie.page_count = 0;
1080 	mutex_unlock(&kvm->arch.vsie.mutex);
1081 }
1082 
1083 void kvm_s390_vsie_kick(struct kvm_vcpu *vcpu)
1084 {
1085 	struct kvm_s390_sie_block *scb = READ_ONCE(vcpu->arch.vsie_block);
1086 
1087 	/*
1088 	 * Even if the VCPU lets go of the shadow sie block reference, it is
1089 	 * still valid in the cache. So we can safely kick it.
1090 	 */
1091 	if (scb) {
1092 		atomic_or(PROG_BLOCK_SIE, &scb->prog20);
1093 		if (scb->prog0c & PROG_IN_SIE)
1094 			atomic_or(CPUSTAT_STOP_INT, &scb->cpuflags);
1095 	}
1096 }
1097