1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * guest access functions
4 *
5 * Copyright IBM Corp. 2014
6 *
7 */
8
9 #include <linux/vmalloc.h>
10 #include <linux/mm_types.h>
11 #include <linux/err.h>
12
13 #include <asm/pgtable.h>
14 #include <asm/gmap.h>
15 #include "kvm-s390.h"
16 #include "gaccess.h"
17 #include <asm/switch_to.h>
18
19 union asce {
20 unsigned long val;
21 struct {
22 unsigned long origin : 52; /* Region- or Segment-Table Origin */
23 unsigned long : 2;
24 unsigned long g : 1; /* Subspace Group Control */
25 unsigned long p : 1; /* Private Space Control */
26 unsigned long s : 1; /* Storage-Alteration-Event Control */
27 unsigned long x : 1; /* Space-Switch-Event Control */
28 unsigned long r : 1; /* Real-Space Control */
29 unsigned long : 1;
30 unsigned long dt : 2; /* Designation-Type Control */
31 unsigned long tl : 2; /* Region- or Segment-Table Length */
32 };
33 };
34
35 enum {
36 ASCE_TYPE_SEGMENT = 0,
37 ASCE_TYPE_REGION3 = 1,
38 ASCE_TYPE_REGION2 = 2,
39 ASCE_TYPE_REGION1 = 3
40 };
41
42 union region1_table_entry {
43 unsigned long val;
44 struct {
45 unsigned long rto: 52;/* Region-Table Origin */
46 unsigned long : 2;
47 unsigned long p : 1; /* DAT-Protection Bit */
48 unsigned long : 1;
49 unsigned long tf : 2; /* Region-Second-Table Offset */
50 unsigned long i : 1; /* Region-Invalid Bit */
51 unsigned long : 1;
52 unsigned long tt : 2; /* Table-Type Bits */
53 unsigned long tl : 2; /* Region-Second-Table Length */
54 };
55 };
56
57 union region2_table_entry {
58 unsigned long val;
59 struct {
60 unsigned long rto: 52;/* Region-Table Origin */
61 unsigned long : 2;
62 unsigned long p : 1; /* DAT-Protection Bit */
63 unsigned long : 1;
64 unsigned long tf : 2; /* Region-Third-Table Offset */
65 unsigned long i : 1; /* Region-Invalid Bit */
66 unsigned long : 1;
67 unsigned long tt : 2; /* Table-Type Bits */
68 unsigned long tl : 2; /* Region-Third-Table Length */
69 };
70 };
71
72 struct region3_table_entry_fc0 {
73 unsigned long sto: 52;/* Segment-Table Origin */
74 unsigned long : 1;
75 unsigned long fc : 1; /* Format-Control */
76 unsigned long p : 1; /* DAT-Protection Bit */
77 unsigned long : 1;
78 unsigned long tf : 2; /* Segment-Table Offset */
79 unsigned long i : 1; /* Region-Invalid Bit */
80 unsigned long cr : 1; /* Common-Region Bit */
81 unsigned long tt : 2; /* Table-Type Bits */
82 unsigned long tl : 2; /* Segment-Table Length */
83 };
84
85 struct region3_table_entry_fc1 {
86 unsigned long rfaa : 33; /* Region-Frame Absolute Address */
87 unsigned long : 14;
88 unsigned long av : 1; /* ACCF-Validity Control */
89 unsigned long acc: 4; /* Access-Control Bits */
90 unsigned long f : 1; /* Fetch-Protection Bit */
91 unsigned long fc : 1; /* Format-Control */
92 unsigned long p : 1; /* DAT-Protection Bit */
93 unsigned long iep: 1; /* Instruction-Execution-Protection */
94 unsigned long : 2;
95 unsigned long i : 1; /* Region-Invalid Bit */
96 unsigned long cr : 1; /* Common-Region Bit */
97 unsigned long tt : 2; /* Table-Type Bits */
98 unsigned long : 2;
99 };
100
101 union region3_table_entry {
102 unsigned long val;
103 struct region3_table_entry_fc0 fc0;
104 struct region3_table_entry_fc1 fc1;
105 struct {
106 unsigned long : 53;
107 unsigned long fc : 1; /* Format-Control */
108 unsigned long : 4;
109 unsigned long i : 1; /* Region-Invalid Bit */
110 unsigned long cr : 1; /* Common-Region Bit */
111 unsigned long tt : 2; /* Table-Type Bits */
112 unsigned long : 2;
113 };
114 };
115
116 struct segment_entry_fc0 {
117 unsigned long pto: 53;/* Page-Table Origin */
118 unsigned long fc : 1; /* Format-Control */
119 unsigned long p : 1; /* DAT-Protection Bit */
120 unsigned long : 3;
121 unsigned long i : 1; /* Segment-Invalid Bit */
122 unsigned long cs : 1; /* Common-Segment Bit */
123 unsigned long tt : 2; /* Table-Type Bits */
124 unsigned long : 2;
125 };
126
127 struct segment_entry_fc1 {
128 unsigned long sfaa : 44; /* Segment-Frame Absolute Address */
129 unsigned long : 3;
130 unsigned long av : 1; /* ACCF-Validity Control */
131 unsigned long acc: 4; /* Access-Control Bits */
132 unsigned long f : 1; /* Fetch-Protection Bit */
133 unsigned long fc : 1; /* Format-Control */
134 unsigned long p : 1; /* DAT-Protection Bit */
135 unsigned long iep: 1; /* Instruction-Execution-Protection */
136 unsigned long : 2;
137 unsigned long i : 1; /* Segment-Invalid Bit */
138 unsigned long cs : 1; /* Common-Segment Bit */
139 unsigned long tt : 2; /* Table-Type Bits */
140 unsigned long : 2;
141 };
142
143 union segment_table_entry {
144 unsigned long val;
145 struct segment_entry_fc0 fc0;
146 struct segment_entry_fc1 fc1;
147 struct {
148 unsigned long : 53;
149 unsigned long fc : 1; /* Format-Control */
150 unsigned long : 4;
151 unsigned long i : 1; /* Segment-Invalid Bit */
152 unsigned long cs : 1; /* Common-Segment Bit */
153 unsigned long tt : 2; /* Table-Type Bits */
154 unsigned long : 2;
155 };
156 };
157
158 enum {
159 TABLE_TYPE_SEGMENT = 0,
160 TABLE_TYPE_REGION3 = 1,
161 TABLE_TYPE_REGION2 = 2,
162 TABLE_TYPE_REGION1 = 3
163 };
164
165 union page_table_entry {
166 unsigned long val;
167 struct {
168 unsigned long pfra : 52; /* Page-Frame Real Address */
169 unsigned long z : 1; /* Zero Bit */
170 unsigned long i : 1; /* Page-Invalid Bit */
171 unsigned long p : 1; /* DAT-Protection Bit */
172 unsigned long iep: 1; /* Instruction-Execution-Protection */
173 unsigned long : 8;
174 };
175 };
176
177 /*
178 * vaddress union in order to easily decode a virtual address into its
179 * region first index, region second index etc. parts.
180 */
181 union vaddress {
182 unsigned long addr;
183 struct {
184 unsigned long rfx : 11;
185 unsigned long rsx : 11;
186 unsigned long rtx : 11;
187 unsigned long sx : 11;
188 unsigned long px : 8;
189 unsigned long bx : 12;
190 };
191 struct {
192 unsigned long rfx01 : 2;
193 unsigned long : 9;
194 unsigned long rsx01 : 2;
195 unsigned long : 9;
196 unsigned long rtx01 : 2;
197 unsigned long : 9;
198 unsigned long sx01 : 2;
199 unsigned long : 29;
200 };
201 };
202
203 /*
204 * raddress union which will contain the result (real or absolute address)
205 * after a page table walk. The rfaa, sfaa and pfra members are used to
206 * simply assign them the value of a region, segment or page table entry.
207 */
208 union raddress {
209 unsigned long addr;
210 unsigned long rfaa : 33; /* Region-Frame Absolute Address */
211 unsigned long sfaa : 44; /* Segment-Frame Absolute Address */
212 unsigned long pfra : 52; /* Page-Frame Real Address */
213 };
214
215 union alet {
216 u32 val;
217 struct {
218 u32 reserved : 7;
219 u32 p : 1;
220 u32 alesn : 8;
221 u32 alen : 16;
222 };
223 };
224
225 union ald {
226 u32 val;
227 struct {
228 u32 : 1;
229 u32 alo : 24;
230 u32 all : 7;
231 };
232 };
233
234 struct ale {
235 unsigned long i : 1; /* ALEN-Invalid Bit */
236 unsigned long : 5;
237 unsigned long fo : 1; /* Fetch-Only Bit */
238 unsigned long p : 1; /* Private Bit */
239 unsigned long alesn : 8; /* Access-List-Entry Sequence Number */
240 unsigned long aleax : 16; /* Access-List-Entry Authorization Index */
241 unsigned long : 32;
242 unsigned long : 1;
243 unsigned long asteo : 25; /* ASN-Second-Table-Entry Origin */
244 unsigned long : 6;
245 unsigned long astesn : 32; /* ASTE Sequence Number */
246 };
247
248 struct aste {
249 unsigned long i : 1; /* ASX-Invalid Bit */
250 unsigned long ato : 29; /* Authority-Table Origin */
251 unsigned long : 1;
252 unsigned long b : 1; /* Base-Space Bit */
253 unsigned long ax : 16; /* Authorization Index */
254 unsigned long atl : 12; /* Authority-Table Length */
255 unsigned long : 2;
256 unsigned long ca : 1; /* Controlled-ASN Bit */
257 unsigned long ra : 1; /* Reusable-ASN Bit */
258 unsigned long asce : 64; /* Address-Space-Control Element */
259 unsigned long ald : 32;
260 unsigned long astesn : 32;
261 /* .. more fields there */
262 };
263
ipte_lock_held(struct kvm_vcpu * vcpu)264 int ipte_lock_held(struct kvm_vcpu *vcpu)
265 {
266 if (vcpu->arch.sie_block->eca & ECA_SII) {
267 int rc;
268
269 read_lock(&vcpu->kvm->arch.sca_lock);
270 rc = kvm_s390_get_ipte_control(vcpu->kvm)->kh != 0;
271 read_unlock(&vcpu->kvm->arch.sca_lock);
272 return rc;
273 }
274 return vcpu->kvm->arch.ipte_lock_count != 0;
275 }
276
ipte_lock_simple(struct kvm_vcpu * vcpu)277 static void ipte_lock_simple(struct kvm_vcpu *vcpu)
278 {
279 union ipte_control old, new, *ic;
280
281 mutex_lock(&vcpu->kvm->arch.ipte_mutex);
282 vcpu->kvm->arch.ipte_lock_count++;
283 if (vcpu->kvm->arch.ipte_lock_count > 1)
284 goto out;
285 retry:
286 read_lock(&vcpu->kvm->arch.sca_lock);
287 ic = kvm_s390_get_ipte_control(vcpu->kvm);
288 do {
289 old = READ_ONCE(*ic);
290 if (old.k) {
291 read_unlock(&vcpu->kvm->arch.sca_lock);
292 cond_resched();
293 goto retry;
294 }
295 new = old;
296 new.k = 1;
297 } while (cmpxchg(&ic->val, old.val, new.val) != old.val);
298 read_unlock(&vcpu->kvm->arch.sca_lock);
299 out:
300 mutex_unlock(&vcpu->kvm->arch.ipte_mutex);
301 }
302
ipte_unlock_simple(struct kvm_vcpu * vcpu)303 static void ipte_unlock_simple(struct kvm_vcpu *vcpu)
304 {
305 union ipte_control old, new, *ic;
306
307 mutex_lock(&vcpu->kvm->arch.ipte_mutex);
308 vcpu->kvm->arch.ipte_lock_count--;
309 if (vcpu->kvm->arch.ipte_lock_count)
310 goto out;
311 read_lock(&vcpu->kvm->arch.sca_lock);
312 ic = kvm_s390_get_ipte_control(vcpu->kvm);
313 do {
314 old = READ_ONCE(*ic);
315 new = old;
316 new.k = 0;
317 } while (cmpxchg(&ic->val, old.val, new.val) != old.val);
318 read_unlock(&vcpu->kvm->arch.sca_lock);
319 wake_up(&vcpu->kvm->arch.ipte_wq);
320 out:
321 mutex_unlock(&vcpu->kvm->arch.ipte_mutex);
322 }
323
ipte_lock_siif(struct kvm_vcpu * vcpu)324 static void ipte_lock_siif(struct kvm_vcpu *vcpu)
325 {
326 union ipte_control old, new, *ic;
327
328 retry:
329 read_lock(&vcpu->kvm->arch.sca_lock);
330 ic = kvm_s390_get_ipte_control(vcpu->kvm);
331 do {
332 old = READ_ONCE(*ic);
333 if (old.kg) {
334 read_unlock(&vcpu->kvm->arch.sca_lock);
335 cond_resched();
336 goto retry;
337 }
338 new = old;
339 new.k = 1;
340 new.kh++;
341 } while (cmpxchg(&ic->val, old.val, new.val) != old.val);
342 read_unlock(&vcpu->kvm->arch.sca_lock);
343 }
344
ipte_unlock_siif(struct kvm_vcpu * vcpu)345 static void ipte_unlock_siif(struct kvm_vcpu *vcpu)
346 {
347 union ipte_control old, new, *ic;
348
349 read_lock(&vcpu->kvm->arch.sca_lock);
350 ic = kvm_s390_get_ipte_control(vcpu->kvm);
351 do {
352 old = READ_ONCE(*ic);
353 new = old;
354 new.kh--;
355 if (!new.kh)
356 new.k = 0;
357 } while (cmpxchg(&ic->val, old.val, new.val) != old.val);
358 read_unlock(&vcpu->kvm->arch.sca_lock);
359 if (!new.kh)
360 wake_up(&vcpu->kvm->arch.ipte_wq);
361 }
362
ipte_lock(struct kvm_vcpu * vcpu)363 void ipte_lock(struct kvm_vcpu *vcpu)
364 {
365 if (vcpu->arch.sie_block->eca & ECA_SII)
366 ipte_lock_siif(vcpu);
367 else
368 ipte_lock_simple(vcpu);
369 }
370
ipte_unlock(struct kvm_vcpu * vcpu)371 void ipte_unlock(struct kvm_vcpu *vcpu)
372 {
373 if (vcpu->arch.sie_block->eca & ECA_SII)
374 ipte_unlock_siif(vcpu);
375 else
376 ipte_unlock_simple(vcpu);
377 }
378
ar_translation(struct kvm_vcpu * vcpu,union asce * asce,u8 ar,enum gacc_mode mode)379 static int ar_translation(struct kvm_vcpu *vcpu, union asce *asce, u8 ar,
380 enum gacc_mode mode)
381 {
382 union alet alet;
383 struct ale ale;
384 struct aste aste;
385 unsigned long ald_addr, authority_table_addr;
386 union ald ald;
387 int eax, rc;
388 u8 authority_table;
389
390 if (ar >= NUM_ACRS)
391 return -EINVAL;
392
393 save_access_regs(vcpu->run->s.regs.acrs);
394 alet.val = vcpu->run->s.regs.acrs[ar];
395
396 if (ar == 0 || alet.val == 0) {
397 asce->val = vcpu->arch.sie_block->gcr[1];
398 return 0;
399 } else if (alet.val == 1) {
400 asce->val = vcpu->arch.sie_block->gcr[7];
401 return 0;
402 }
403
404 if (alet.reserved)
405 return PGM_ALET_SPECIFICATION;
406
407 if (alet.p)
408 ald_addr = vcpu->arch.sie_block->gcr[5];
409 else
410 ald_addr = vcpu->arch.sie_block->gcr[2];
411 ald_addr &= 0x7fffffc0;
412
413 rc = read_guest_real(vcpu, ald_addr + 16, &ald.val, sizeof(union ald));
414 if (rc)
415 return rc;
416
417 if (alet.alen / 8 > ald.all)
418 return PGM_ALEN_TRANSLATION;
419
420 if (0x7fffffff - ald.alo * 128 < alet.alen * 16)
421 return PGM_ADDRESSING;
422
423 rc = read_guest_real(vcpu, ald.alo * 128 + alet.alen * 16, &ale,
424 sizeof(struct ale));
425 if (rc)
426 return rc;
427
428 if (ale.i == 1)
429 return PGM_ALEN_TRANSLATION;
430 if (ale.alesn != alet.alesn)
431 return PGM_ALE_SEQUENCE;
432
433 rc = read_guest_real(vcpu, ale.asteo * 64, &aste, sizeof(struct aste));
434 if (rc)
435 return rc;
436
437 if (aste.i)
438 return PGM_ASTE_VALIDITY;
439 if (aste.astesn != ale.astesn)
440 return PGM_ASTE_SEQUENCE;
441
442 if (ale.p == 1) {
443 eax = (vcpu->arch.sie_block->gcr[8] >> 16) & 0xffff;
444 if (ale.aleax != eax) {
445 if (eax / 16 > aste.atl)
446 return PGM_EXTENDED_AUTHORITY;
447
448 authority_table_addr = aste.ato * 4 + eax / 4;
449
450 rc = read_guest_real(vcpu, authority_table_addr,
451 &authority_table,
452 sizeof(u8));
453 if (rc)
454 return rc;
455
456 if ((authority_table & (0x40 >> ((eax & 3) * 2))) == 0)
457 return PGM_EXTENDED_AUTHORITY;
458 }
459 }
460
461 if (ale.fo == 1 && mode == GACC_STORE)
462 return PGM_PROTECTION;
463
464 asce->val = aste.asce;
465 return 0;
466 }
467
468 struct trans_exc_code_bits {
469 unsigned long addr : 52; /* Translation-exception Address */
470 unsigned long fsi : 2; /* Access Exception Fetch/Store Indication */
471 unsigned long : 2;
472 unsigned long b56 : 1;
473 unsigned long : 3;
474 unsigned long b60 : 1;
475 unsigned long b61 : 1;
476 unsigned long as : 2; /* ASCE Identifier */
477 };
478
479 enum {
480 FSI_UNKNOWN = 0, /* Unknown wether fetch or store */
481 FSI_STORE = 1, /* Exception was due to store operation */
482 FSI_FETCH = 2 /* Exception was due to fetch operation */
483 };
484
485 enum prot_type {
486 PROT_TYPE_LA = 0,
487 PROT_TYPE_KEYC = 1,
488 PROT_TYPE_ALC = 2,
489 PROT_TYPE_DAT = 3,
490 PROT_TYPE_IEP = 4,
491 };
492
trans_exc(struct kvm_vcpu * vcpu,int code,unsigned long gva,u8 ar,enum gacc_mode mode,enum prot_type prot)493 static int trans_exc(struct kvm_vcpu *vcpu, int code, unsigned long gva,
494 u8 ar, enum gacc_mode mode, enum prot_type prot)
495 {
496 struct kvm_s390_pgm_info *pgm = &vcpu->arch.pgm;
497 struct trans_exc_code_bits *tec;
498
499 memset(pgm, 0, sizeof(*pgm));
500 pgm->code = code;
501 tec = (struct trans_exc_code_bits *)&pgm->trans_exc_code;
502
503 switch (code) {
504 case PGM_PROTECTION:
505 switch (prot) {
506 case PROT_TYPE_IEP:
507 tec->b61 = 1;
508 /* FALL THROUGH */
509 case PROT_TYPE_LA:
510 tec->b56 = 1;
511 break;
512 case PROT_TYPE_KEYC:
513 tec->b60 = 1;
514 break;
515 case PROT_TYPE_ALC:
516 tec->b60 = 1;
517 /* FALL THROUGH */
518 case PROT_TYPE_DAT:
519 tec->b61 = 1;
520 break;
521 }
522 /* FALL THROUGH */
523 case PGM_ASCE_TYPE:
524 case PGM_PAGE_TRANSLATION:
525 case PGM_REGION_FIRST_TRANS:
526 case PGM_REGION_SECOND_TRANS:
527 case PGM_REGION_THIRD_TRANS:
528 case PGM_SEGMENT_TRANSLATION:
529 /*
530 * op_access_id only applies to MOVE_PAGE -> set bit 61
531 * exc_access_id has to be set to 0 for some instructions. Both
532 * cases have to be handled by the caller.
533 */
534 tec->addr = gva >> PAGE_SHIFT;
535 tec->fsi = mode == GACC_STORE ? FSI_STORE : FSI_FETCH;
536 tec->as = psw_bits(vcpu->arch.sie_block->gpsw).as;
537 /* FALL THROUGH */
538 case PGM_ALEN_TRANSLATION:
539 case PGM_ALE_SEQUENCE:
540 case PGM_ASTE_VALIDITY:
541 case PGM_ASTE_SEQUENCE:
542 case PGM_EXTENDED_AUTHORITY:
543 /*
544 * We can always store exc_access_id, as it is
545 * undefined for non-ar cases. It is undefined for
546 * most DAT protection exceptions.
547 */
548 pgm->exc_access_id = ar;
549 break;
550 }
551 return code;
552 }
553
get_vcpu_asce(struct kvm_vcpu * vcpu,union asce * asce,unsigned long ga,u8 ar,enum gacc_mode mode)554 static int get_vcpu_asce(struct kvm_vcpu *vcpu, union asce *asce,
555 unsigned long ga, u8 ar, enum gacc_mode mode)
556 {
557 int rc;
558 struct psw_bits psw = psw_bits(vcpu->arch.sie_block->gpsw);
559
560 if (!psw.dat) {
561 asce->val = 0;
562 asce->r = 1;
563 return 0;
564 }
565
566 if ((mode == GACC_IFETCH) && (psw.as != PSW_BITS_AS_HOME))
567 psw.as = PSW_BITS_AS_PRIMARY;
568
569 switch (psw.as) {
570 case PSW_BITS_AS_PRIMARY:
571 asce->val = vcpu->arch.sie_block->gcr[1];
572 return 0;
573 case PSW_BITS_AS_SECONDARY:
574 asce->val = vcpu->arch.sie_block->gcr[7];
575 return 0;
576 case PSW_BITS_AS_HOME:
577 asce->val = vcpu->arch.sie_block->gcr[13];
578 return 0;
579 case PSW_BITS_AS_ACCREG:
580 rc = ar_translation(vcpu, asce, ar, mode);
581 if (rc > 0)
582 return trans_exc(vcpu, rc, ga, ar, mode, PROT_TYPE_ALC);
583 return rc;
584 }
585 return 0;
586 }
587
deref_table(struct kvm * kvm,unsigned long gpa,unsigned long * val)588 static int deref_table(struct kvm *kvm, unsigned long gpa, unsigned long *val)
589 {
590 return kvm_read_guest(kvm, gpa, val, sizeof(*val));
591 }
592
593 /**
594 * guest_translate - translate a guest virtual into a guest absolute address
595 * @vcpu: virtual cpu
596 * @gva: guest virtual address
597 * @gpa: points to where guest physical (absolute) address should be stored
598 * @asce: effective asce
599 * @mode: indicates the access mode to be used
600 * @prot: returns the type for protection exceptions
601 *
602 * Translate a guest virtual address into a guest absolute address by means
603 * of dynamic address translation as specified by the architecture.
604 * If the resulting absolute address is not available in the configuration
605 * an addressing exception is indicated and @gpa will not be changed.
606 *
607 * Returns: - zero on success; @gpa contains the resulting absolute address
608 * - a negative value if guest access failed due to e.g. broken
609 * guest mapping
610 * - a positve value if an access exception happened. In this case
611 * the returned value is the program interruption code as defined
612 * by the architecture
613 */
guest_translate(struct kvm_vcpu * vcpu,unsigned long gva,unsigned long * gpa,const union asce asce,enum gacc_mode mode,enum prot_type * prot)614 static unsigned long guest_translate(struct kvm_vcpu *vcpu, unsigned long gva,
615 unsigned long *gpa, const union asce asce,
616 enum gacc_mode mode, enum prot_type *prot)
617 {
618 union vaddress vaddr = {.addr = gva};
619 union raddress raddr = {.addr = gva};
620 union page_table_entry pte;
621 int dat_protection = 0;
622 int iep_protection = 0;
623 union ctlreg0 ctlreg0;
624 unsigned long ptr;
625 int edat1, edat2, iep;
626
627 ctlreg0.val = vcpu->arch.sie_block->gcr[0];
628 edat1 = ctlreg0.edat && test_kvm_facility(vcpu->kvm, 8);
629 edat2 = edat1 && test_kvm_facility(vcpu->kvm, 78);
630 iep = ctlreg0.iep && test_kvm_facility(vcpu->kvm, 130);
631 if (asce.r)
632 goto real_address;
633 ptr = asce.origin * PAGE_SIZE;
634 switch (asce.dt) {
635 case ASCE_TYPE_REGION1:
636 if (vaddr.rfx01 > asce.tl)
637 return PGM_REGION_FIRST_TRANS;
638 ptr += vaddr.rfx * 8;
639 break;
640 case ASCE_TYPE_REGION2:
641 if (vaddr.rfx)
642 return PGM_ASCE_TYPE;
643 if (vaddr.rsx01 > asce.tl)
644 return PGM_REGION_SECOND_TRANS;
645 ptr += vaddr.rsx * 8;
646 break;
647 case ASCE_TYPE_REGION3:
648 if (vaddr.rfx || vaddr.rsx)
649 return PGM_ASCE_TYPE;
650 if (vaddr.rtx01 > asce.tl)
651 return PGM_REGION_THIRD_TRANS;
652 ptr += vaddr.rtx * 8;
653 break;
654 case ASCE_TYPE_SEGMENT:
655 if (vaddr.rfx || vaddr.rsx || vaddr.rtx)
656 return PGM_ASCE_TYPE;
657 if (vaddr.sx01 > asce.tl)
658 return PGM_SEGMENT_TRANSLATION;
659 ptr += vaddr.sx * 8;
660 break;
661 }
662 switch (asce.dt) {
663 case ASCE_TYPE_REGION1: {
664 union region1_table_entry rfte;
665
666 if (kvm_is_error_gpa(vcpu->kvm, ptr))
667 return PGM_ADDRESSING;
668 if (deref_table(vcpu->kvm, ptr, &rfte.val))
669 return -EFAULT;
670 if (rfte.i)
671 return PGM_REGION_FIRST_TRANS;
672 if (rfte.tt != TABLE_TYPE_REGION1)
673 return PGM_TRANSLATION_SPEC;
674 if (vaddr.rsx01 < rfte.tf || vaddr.rsx01 > rfte.tl)
675 return PGM_REGION_SECOND_TRANS;
676 if (edat1)
677 dat_protection |= rfte.p;
678 ptr = rfte.rto * PAGE_SIZE + vaddr.rsx * 8;
679 }
680 /* fallthrough */
681 case ASCE_TYPE_REGION2: {
682 union region2_table_entry rste;
683
684 if (kvm_is_error_gpa(vcpu->kvm, ptr))
685 return PGM_ADDRESSING;
686 if (deref_table(vcpu->kvm, ptr, &rste.val))
687 return -EFAULT;
688 if (rste.i)
689 return PGM_REGION_SECOND_TRANS;
690 if (rste.tt != TABLE_TYPE_REGION2)
691 return PGM_TRANSLATION_SPEC;
692 if (vaddr.rtx01 < rste.tf || vaddr.rtx01 > rste.tl)
693 return PGM_REGION_THIRD_TRANS;
694 if (edat1)
695 dat_protection |= rste.p;
696 ptr = rste.rto * PAGE_SIZE + vaddr.rtx * 8;
697 }
698 /* fallthrough */
699 case ASCE_TYPE_REGION3: {
700 union region3_table_entry rtte;
701
702 if (kvm_is_error_gpa(vcpu->kvm, ptr))
703 return PGM_ADDRESSING;
704 if (deref_table(vcpu->kvm, ptr, &rtte.val))
705 return -EFAULT;
706 if (rtte.i)
707 return PGM_REGION_THIRD_TRANS;
708 if (rtte.tt != TABLE_TYPE_REGION3)
709 return PGM_TRANSLATION_SPEC;
710 if (rtte.cr && asce.p && edat2)
711 return PGM_TRANSLATION_SPEC;
712 if (rtte.fc && edat2) {
713 dat_protection |= rtte.fc1.p;
714 iep_protection = rtte.fc1.iep;
715 raddr.rfaa = rtte.fc1.rfaa;
716 goto absolute_address;
717 }
718 if (vaddr.sx01 < rtte.fc0.tf)
719 return PGM_SEGMENT_TRANSLATION;
720 if (vaddr.sx01 > rtte.fc0.tl)
721 return PGM_SEGMENT_TRANSLATION;
722 if (edat1)
723 dat_protection |= rtte.fc0.p;
724 ptr = rtte.fc0.sto * PAGE_SIZE + vaddr.sx * 8;
725 }
726 /* fallthrough */
727 case ASCE_TYPE_SEGMENT: {
728 union segment_table_entry ste;
729
730 if (kvm_is_error_gpa(vcpu->kvm, ptr))
731 return PGM_ADDRESSING;
732 if (deref_table(vcpu->kvm, ptr, &ste.val))
733 return -EFAULT;
734 if (ste.i)
735 return PGM_SEGMENT_TRANSLATION;
736 if (ste.tt != TABLE_TYPE_SEGMENT)
737 return PGM_TRANSLATION_SPEC;
738 if (ste.cs && asce.p)
739 return PGM_TRANSLATION_SPEC;
740 if (ste.fc && edat1) {
741 dat_protection |= ste.fc1.p;
742 iep_protection = ste.fc1.iep;
743 raddr.sfaa = ste.fc1.sfaa;
744 goto absolute_address;
745 }
746 dat_protection |= ste.fc0.p;
747 ptr = ste.fc0.pto * (PAGE_SIZE / 2) + vaddr.px * 8;
748 }
749 }
750 if (kvm_is_error_gpa(vcpu->kvm, ptr))
751 return PGM_ADDRESSING;
752 if (deref_table(vcpu->kvm, ptr, &pte.val))
753 return -EFAULT;
754 if (pte.i)
755 return PGM_PAGE_TRANSLATION;
756 if (pte.z)
757 return PGM_TRANSLATION_SPEC;
758 dat_protection |= pte.p;
759 iep_protection = pte.iep;
760 raddr.pfra = pte.pfra;
761 real_address:
762 raddr.addr = kvm_s390_real_to_abs(vcpu, raddr.addr);
763 absolute_address:
764 if (mode == GACC_STORE && dat_protection) {
765 *prot = PROT_TYPE_DAT;
766 return PGM_PROTECTION;
767 }
768 if (mode == GACC_IFETCH && iep_protection && iep) {
769 *prot = PROT_TYPE_IEP;
770 return PGM_PROTECTION;
771 }
772 if (kvm_is_error_gpa(vcpu->kvm, raddr.addr))
773 return PGM_ADDRESSING;
774 *gpa = raddr.addr;
775 return 0;
776 }
777
is_low_address(unsigned long ga)778 static inline int is_low_address(unsigned long ga)
779 {
780 /* Check for address ranges 0..511 and 4096..4607 */
781 return (ga & ~0x11fful) == 0;
782 }
783
low_address_protection_enabled(struct kvm_vcpu * vcpu,const union asce asce)784 static int low_address_protection_enabled(struct kvm_vcpu *vcpu,
785 const union asce asce)
786 {
787 union ctlreg0 ctlreg0 = {.val = vcpu->arch.sie_block->gcr[0]};
788 psw_t *psw = &vcpu->arch.sie_block->gpsw;
789
790 if (!ctlreg0.lap)
791 return 0;
792 if (psw_bits(*psw).dat && asce.p)
793 return 0;
794 return 1;
795 }
796
guest_page_range(struct kvm_vcpu * vcpu,unsigned long ga,u8 ar,unsigned long * pages,unsigned long nr_pages,const union asce asce,enum gacc_mode mode)797 static int guest_page_range(struct kvm_vcpu *vcpu, unsigned long ga, u8 ar,
798 unsigned long *pages, unsigned long nr_pages,
799 const union asce asce, enum gacc_mode mode)
800 {
801 psw_t *psw = &vcpu->arch.sie_block->gpsw;
802 int lap_enabled, rc = 0;
803 enum prot_type prot;
804
805 lap_enabled = low_address_protection_enabled(vcpu, asce);
806 while (nr_pages) {
807 ga = kvm_s390_logical_to_effective(vcpu, ga);
808 if (mode == GACC_STORE && lap_enabled && is_low_address(ga))
809 return trans_exc(vcpu, PGM_PROTECTION, ga, ar, mode,
810 PROT_TYPE_LA);
811 ga &= PAGE_MASK;
812 if (psw_bits(*psw).dat) {
813 rc = guest_translate(vcpu, ga, pages, asce, mode, &prot);
814 if (rc < 0)
815 return rc;
816 } else {
817 *pages = kvm_s390_real_to_abs(vcpu, ga);
818 if (kvm_is_error_gpa(vcpu->kvm, *pages))
819 rc = PGM_ADDRESSING;
820 }
821 if (rc)
822 return trans_exc(vcpu, rc, ga, ar, mode, prot);
823 ga += PAGE_SIZE;
824 pages++;
825 nr_pages--;
826 }
827 return 0;
828 }
829
access_guest(struct kvm_vcpu * vcpu,unsigned long ga,u8 ar,void * data,unsigned long len,enum gacc_mode mode)830 int access_guest(struct kvm_vcpu *vcpu, unsigned long ga, u8 ar, void *data,
831 unsigned long len, enum gacc_mode mode)
832 {
833 psw_t *psw = &vcpu->arch.sie_block->gpsw;
834 unsigned long _len, nr_pages, gpa, idx;
835 unsigned long pages_array[2];
836 unsigned long *pages;
837 int need_ipte_lock;
838 union asce asce;
839 int rc;
840
841 if (!len)
842 return 0;
843 ga = kvm_s390_logical_to_effective(vcpu, ga);
844 rc = get_vcpu_asce(vcpu, &asce, ga, ar, mode);
845 if (rc)
846 return rc;
847 nr_pages = (((ga & ~PAGE_MASK) + len - 1) >> PAGE_SHIFT) + 1;
848 pages = pages_array;
849 if (nr_pages > ARRAY_SIZE(pages_array))
850 pages = vmalloc(array_size(nr_pages, sizeof(unsigned long)));
851 if (!pages)
852 return -ENOMEM;
853 need_ipte_lock = psw_bits(*psw).dat && !asce.r;
854 if (need_ipte_lock)
855 ipte_lock(vcpu);
856 rc = guest_page_range(vcpu, ga, ar, pages, nr_pages, asce, mode);
857 for (idx = 0; idx < nr_pages && !rc; idx++) {
858 gpa = *(pages + idx) + (ga & ~PAGE_MASK);
859 _len = min(PAGE_SIZE - (gpa & ~PAGE_MASK), len);
860 if (mode == GACC_STORE)
861 rc = kvm_write_guest(vcpu->kvm, gpa, data, _len);
862 else
863 rc = kvm_read_guest(vcpu->kvm, gpa, data, _len);
864 len -= _len;
865 ga += _len;
866 data += _len;
867 }
868 if (need_ipte_lock)
869 ipte_unlock(vcpu);
870 if (nr_pages > ARRAY_SIZE(pages_array))
871 vfree(pages);
872 return rc;
873 }
874
access_guest_real(struct kvm_vcpu * vcpu,unsigned long gra,void * data,unsigned long len,enum gacc_mode mode)875 int access_guest_real(struct kvm_vcpu *vcpu, unsigned long gra,
876 void *data, unsigned long len, enum gacc_mode mode)
877 {
878 unsigned long _len, gpa;
879 int rc = 0;
880
881 while (len && !rc) {
882 gpa = kvm_s390_real_to_abs(vcpu, gra);
883 _len = min(PAGE_SIZE - (gpa & ~PAGE_MASK), len);
884 if (mode)
885 rc = write_guest_abs(vcpu, gpa, data, _len);
886 else
887 rc = read_guest_abs(vcpu, gpa, data, _len);
888 len -= _len;
889 gra += _len;
890 data += _len;
891 }
892 return rc;
893 }
894
895 /**
896 * guest_translate_address - translate guest logical into guest absolute address
897 *
898 * Parameter semantics are the same as the ones from guest_translate.
899 * The memory contents at the guest address are not changed.
900 *
901 * Note: The IPTE lock is not taken during this function, so the caller
902 * has to take care of this.
903 */
guest_translate_address(struct kvm_vcpu * vcpu,unsigned long gva,u8 ar,unsigned long * gpa,enum gacc_mode mode)904 int guest_translate_address(struct kvm_vcpu *vcpu, unsigned long gva, u8 ar,
905 unsigned long *gpa, enum gacc_mode mode)
906 {
907 psw_t *psw = &vcpu->arch.sie_block->gpsw;
908 enum prot_type prot;
909 union asce asce;
910 int rc;
911
912 gva = kvm_s390_logical_to_effective(vcpu, gva);
913 rc = get_vcpu_asce(vcpu, &asce, gva, ar, mode);
914 if (rc)
915 return rc;
916 if (is_low_address(gva) && low_address_protection_enabled(vcpu, asce)) {
917 if (mode == GACC_STORE)
918 return trans_exc(vcpu, PGM_PROTECTION, gva, 0,
919 mode, PROT_TYPE_LA);
920 }
921
922 if (psw_bits(*psw).dat && !asce.r) { /* Use DAT? */
923 rc = guest_translate(vcpu, gva, gpa, asce, mode, &prot);
924 if (rc > 0)
925 return trans_exc(vcpu, rc, gva, 0, mode, prot);
926 } else {
927 *gpa = kvm_s390_real_to_abs(vcpu, gva);
928 if (kvm_is_error_gpa(vcpu->kvm, *gpa))
929 return trans_exc(vcpu, rc, gva, PGM_ADDRESSING, mode, 0);
930 }
931
932 return rc;
933 }
934
935 /**
936 * check_gva_range - test a range of guest virtual addresses for accessibility
937 */
check_gva_range(struct kvm_vcpu * vcpu,unsigned long gva,u8 ar,unsigned long length,enum gacc_mode mode)938 int check_gva_range(struct kvm_vcpu *vcpu, unsigned long gva, u8 ar,
939 unsigned long length, enum gacc_mode mode)
940 {
941 unsigned long gpa;
942 unsigned long currlen;
943 int rc = 0;
944
945 ipte_lock(vcpu);
946 while (length > 0 && !rc) {
947 currlen = min(length, PAGE_SIZE - (gva % PAGE_SIZE));
948 rc = guest_translate_address(vcpu, gva, ar, &gpa, mode);
949 gva += currlen;
950 length -= currlen;
951 }
952 ipte_unlock(vcpu);
953
954 return rc;
955 }
956
957 /**
958 * kvm_s390_check_low_addr_prot_real - check for low-address protection
959 * @gra: Guest real address
960 *
961 * Checks whether an address is subject to low-address protection and set
962 * up vcpu->arch.pgm accordingly if necessary.
963 *
964 * Return: 0 if no protection exception, or PGM_PROTECTION if protected.
965 */
kvm_s390_check_low_addr_prot_real(struct kvm_vcpu * vcpu,unsigned long gra)966 int kvm_s390_check_low_addr_prot_real(struct kvm_vcpu *vcpu, unsigned long gra)
967 {
968 union ctlreg0 ctlreg0 = {.val = vcpu->arch.sie_block->gcr[0]};
969
970 if (!ctlreg0.lap || !is_low_address(gra))
971 return 0;
972 return trans_exc(vcpu, PGM_PROTECTION, gra, 0, GACC_STORE, PROT_TYPE_LA);
973 }
974
975 /**
976 * kvm_s390_shadow_tables - walk the guest page table and create shadow tables
977 * @sg: pointer to the shadow guest address space structure
978 * @saddr: faulting address in the shadow gmap
979 * @pgt: pointer to the page table address result
980 * @fake: pgt references contiguous guest memory block, not a pgtable
981 */
kvm_s390_shadow_tables(struct gmap * sg,unsigned long saddr,unsigned long * pgt,int * dat_protection,int * fake)982 static int kvm_s390_shadow_tables(struct gmap *sg, unsigned long saddr,
983 unsigned long *pgt, int *dat_protection,
984 int *fake)
985 {
986 struct gmap *parent;
987 union asce asce;
988 union vaddress vaddr;
989 unsigned long ptr;
990 int rc;
991
992 *fake = 0;
993 *dat_protection = 0;
994 parent = sg->parent;
995 vaddr.addr = saddr;
996 asce.val = sg->orig_asce;
997 ptr = asce.origin * PAGE_SIZE;
998 if (asce.r) {
999 *fake = 1;
1000 ptr = 0;
1001 asce.dt = ASCE_TYPE_REGION1;
1002 }
1003 switch (asce.dt) {
1004 case ASCE_TYPE_REGION1:
1005 if (vaddr.rfx01 > asce.tl && !*fake)
1006 return PGM_REGION_FIRST_TRANS;
1007 break;
1008 case ASCE_TYPE_REGION2:
1009 if (vaddr.rfx)
1010 return PGM_ASCE_TYPE;
1011 if (vaddr.rsx01 > asce.tl)
1012 return PGM_REGION_SECOND_TRANS;
1013 break;
1014 case ASCE_TYPE_REGION3:
1015 if (vaddr.rfx || vaddr.rsx)
1016 return PGM_ASCE_TYPE;
1017 if (vaddr.rtx01 > asce.tl)
1018 return PGM_REGION_THIRD_TRANS;
1019 break;
1020 case ASCE_TYPE_SEGMENT:
1021 if (vaddr.rfx || vaddr.rsx || vaddr.rtx)
1022 return PGM_ASCE_TYPE;
1023 if (vaddr.sx01 > asce.tl)
1024 return PGM_SEGMENT_TRANSLATION;
1025 break;
1026 }
1027
1028 switch (asce.dt) {
1029 case ASCE_TYPE_REGION1: {
1030 union region1_table_entry rfte;
1031
1032 if (*fake) {
1033 ptr += vaddr.rfx * _REGION1_SIZE;
1034 rfte.val = ptr;
1035 goto shadow_r2t;
1036 }
1037 rc = gmap_read_table(parent, ptr + vaddr.rfx * 8, &rfte.val);
1038 if (rc)
1039 return rc;
1040 if (rfte.i)
1041 return PGM_REGION_FIRST_TRANS;
1042 if (rfte.tt != TABLE_TYPE_REGION1)
1043 return PGM_TRANSLATION_SPEC;
1044 if (vaddr.rsx01 < rfte.tf || vaddr.rsx01 > rfte.tl)
1045 return PGM_REGION_SECOND_TRANS;
1046 if (sg->edat_level >= 1)
1047 *dat_protection |= rfte.p;
1048 ptr = rfte.rto * PAGE_SIZE;
1049 shadow_r2t:
1050 rc = gmap_shadow_r2t(sg, saddr, rfte.val, *fake);
1051 if (rc)
1052 return rc;
1053 } /* fallthrough */
1054 case ASCE_TYPE_REGION2: {
1055 union region2_table_entry rste;
1056
1057 if (*fake) {
1058 ptr += vaddr.rsx * _REGION2_SIZE;
1059 rste.val = ptr;
1060 goto shadow_r3t;
1061 }
1062 rc = gmap_read_table(parent, ptr + vaddr.rsx * 8, &rste.val);
1063 if (rc)
1064 return rc;
1065 if (rste.i)
1066 return PGM_REGION_SECOND_TRANS;
1067 if (rste.tt != TABLE_TYPE_REGION2)
1068 return PGM_TRANSLATION_SPEC;
1069 if (vaddr.rtx01 < rste.tf || vaddr.rtx01 > rste.tl)
1070 return PGM_REGION_THIRD_TRANS;
1071 if (sg->edat_level >= 1)
1072 *dat_protection |= rste.p;
1073 ptr = rste.rto * PAGE_SIZE;
1074 shadow_r3t:
1075 rste.p |= *dat_protection;
1076 rc = gmap_shadow_r3t(sg, saddr, rste.val, *fake);
1077 if (rc)
1078 return rc;
1079 } /* fallthrough */
1080 case ASCE_TYPE_REGION3: {
1081 union region3_table_entry rtte;
1082
1083 if (*fake) {
1084 ptr += vaddr.rtx * _REGION3_SIZE;
1085 rtte.val = ptr;
1086 goto shadow_sgt;
1087 }
1088 rc = gmap_read_table(parent, ptr + vaddr.rtx * 8, &rtte.val);
1089 if (rc)
1090 return rc;
1091 if (rtte.i)
1092 return PGM_REGION_THIRD_TRANS;
1093 if (rtte.tt != TABLE_TYPE_REGION3)
1094 return PGM_TRANSLATION_SPEC;
1095 if (rtte.cr && asce.p && sg->edat_level >= 2)
1096 return PGM_TRANSLATION_SPEC;
1097 if (rtte.fc && sg->edat_level >= 2) {
1098 *dat_protection |= rtte.fc0.p;
1099 *fake = 1;
1100 ptr = rtte.fc1.rfaa * _REGION3_SIZE;
1101 rtte.val = ptr;
1102 goto shadow_sgt;
1103 }
1104 if (vaddr.sx01 < rtte.fc0.tf || vaddr.sx01 > rtte.fc0.tl)
1105 return PGM_SEGMENT_TRANSLATION;
1106 if (sg->edat_level >= 1)
1107 *dat_protection |= rtte.fc0.p;
1108 ptr = rtte.fc0.sto * PAGE_SIZE;
1109 shadow_sgt:
1110 rtte.fc0.p |= *dat_protection;
1111 rc = gmap_shadow_sgt(sg, saddr, rtte.val, *fake);
1112 if (rc)
1113 return rc;
1114 } /* fallthrough */
1115 case ASCE_TYPE_SEGMENT: {
1116 union segment_table_entry ste;
1117
1118 if (*fake) {
1119 ptr += vaddr.sx * _SEGMENT_SIZE;
1120 ste.val = ptr;
1121 goto shadow_pgt;
1122 }
1123 rc = gmap_read_table(parent, ptr + vaddr.sx * 8, &ste.val);
1124 if (rc)
1125 return rc;
1126 if (ste.i)
1127 return PGM_SEGMENT_TRANSLATION;
1128 if (ste.tt != TABLE_TYPE_SEGMENT)
1129 return PGM_TRANSLATION_SPEC;
1130 if (ste.cs && asce.p)
1131 return PGM_TRANSLATION_SPEC;
1132 *dat_protection |= ste.fc0.p;
1133 if (ste.fc && sg->edat_level >= 1) {
1134 *fake = 1;
1135 ptr = ste.fc1.sfaa * _SEGMENT_SIZE;
1136 ste.val = ptr;
1137 goto shadow_pgt;
1138 }
1139 ptr = ste.fc0.pto * (PAGE_SIZE / 2);
1140 shadow_pgt:
1141 ste.fc0.p |= *dat_protection;
1142 rc = gmap_shadow_pgt(sg, saddr, ste.val, *fake);
1143 if (rc)
1144 return rc;
1145 }
1146 }
1147 /* Return the parent address of the page table */
1148 *pgt = ptr;
1149 return 0;
1150 }
1151
1152 /**
1153 * kvm_s390_shadow_fault - handle fault on a shadow page table
1154 * @vcpu: virtual cpu
1155 * @sg: pointer to the shadow guest address space structure
1156 * @saddr: faulting address in the shadow gmap
1157 *
1158 * Returns: - 0 if the shadow fault was successfully resolved
1159 * - > 0 (pgm exception code) on exceptions while faulting
1160 * - -EAGAIN if the caller can retry immediately
1161 * - -EFAULT when accessing invalid guest addresses
1162 * - -ENOMEM if out of memory
1163 */
kvm_s390_shadow_fault(struct kvm_vcpu * vcpu,struct gmap * sg,unsigned long saddr)1164 int kvm_s390_shadow_fault(struct kvm_vcpu *vcpu, struct gmap *sg,
1165 unsigned long saddr)
1166 {
1167 union vaddress vaddr;
1168 union page_table_entry pte;
1169 unsigned long pgt;
1170 int dat_protection, fake;
1171 int rc;
1172
1173 down_read(&sg->mm->mmap_sem);
1174 /*
1175 * We don't want any guest-2 tables to change - so the parent
1176 * tables/pointers we read stay valid - unshadowing is however
1177 * always possible - only guest_table_lock protects us.
1178 */
1179 ipte_lock(vcpu);
1180
1181 rc = gmap_shadow_pgt_lookup(sg, saddr, &pgt, &dat_protection, &fake);
1182 if (rc)
1183 rc = kvm_s390_shadow_tables(sg, saddr, &pgt, &dat_protection,
1184 &fake);
1185
1186 vaddr.addr = saddr;
1187 if (fake) {
1188 pte.val = pgt + vaddr.px * PAGE_SIZE;
1189 goto shadow_page;
1190 }
1191 if (!rc)
1192 rc = gmap_read_table(sg->parent, pgt + vaddr.px * 8, &pte.val);
1193 if (!rc && pte.i)
1194 rc = PGM_PAGE_TRANSLATION;
1195 if (!rc && pte.z)
1196 rc = PGM_TRANSLATION_SPEC;
1197 shadow_page:
1198 pte.p |= dat_protection;
1199 if (!rc)
1200 rc = gmap_shadow_page(sg, saddr, __pte(pte.val));
1201 ipte_unlock(vcpu);
1202 up_read(&sg->mm->mmap_sem);
1203 return rc;
1204 }
1205
