xref: /linux/drivers/iommu/s390-iommu.c (revision 07fdad3a93756b872da7b53647715c48d0f4a2d0)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * IOMMU API for s390 PCI devices
4  *
5  * Copyright IBM Corp. 2015
6  * Author(s): Gerald Schaefer <gerald.schaefer@de.ibm.com>
7  */
8 
9 #include <linux/pci.h>
10 #include <linux/iommu.h>
11 #include <linux/iommu-helper.h>
12 #include <linux/sizes.h>
13 #include <linux/rculist.h>
14 #include <linux/rcupdate.h>
15 #include <asm/pci_dma.h>
16 
17 #include "dma-iommu.h"
18 
19 static const struct iommu_ops s390_iommu_ops, s390_iommu_rtr_ops;
20 
21 static struct kmem_cache *dma_region_table_cache;
22 static struct kmem_cache *dma_page_table_cache;
23 
24 static u64 s390_iommu_aperture;
25 static u32 s390_iommu_aperture_factor = 1;
26 
27 struct s390_domain {
28 	struct iommu_domain	domain;
29 	struct list_head	devices;
30 	struct zpci_iommu_ctrs	ctrs;
31 	unsigned long		*dma_table;
32 	spinlock_t		list_lock;
33 	struct rcu_head		rcu;
34 	u8			origin_type;
35 };
36 
37 static struct iommu_domain blocking_domain;
38 
39 static inline unsigned int calc_rfx(dma_addr_t ptr)
40 {
41 	return ((unsigned long)ptr >> ZPCI_RF_SHIFT) & ZPCI_INDEX_MASK;
42 }
43 
44 static inline unsigned int calc_rsx(dma_addr_t ptr)
45 {
46 	return ((unsigned long)ptr >> ZPCI_RS_SHIFT) & ZPCI_INDEX_MASK;
47 }
48 
49 static inline unsigned int calc_rtx(dma_addr_t ptr)
50 {
51 	return ((unsigned long)ptr >> ZPCI_RT_SHIFT) & ZPCI_INDEX_MASK;
52 }
53 
54 static inline unsigned int calc_sx(dma_addr_t ptr)
55 {
56 	return ((unsigned long)ptr >> ZPCI_ST_SHIFT) & ZPCI_INDEX_MASK;
57 }
58 
59 static inline unsigned int calc_px(dma_addr_t ptr)
60 {
61 	return ((unsigned long)ptr >> PAGE_SHIFT) & ZPCI_PT_MASK;
62 }
63 
64 static inline void set_pt_pfaa(unsigned long *entry, phys_addr_t pfaa)
65 {
66 	*entry &= ZPCI_PTE_FLAG_MASK;
67 	*entry |= (pfaa & ZPCI_PTE_ADDR_MASK);
68 }
69 
70 static inline void set_rf_rso(unsigned long *entry, phys_addr_t rso)
71 {
72 	*entry &= ZPCI_RTE_FLAG_MASK;
73 	*entry |= (rso & ZPCI_RTE_ADDR_MASK);
74 	*entry |= ZPCI_TABLE_TYPE_RFX;
75 }
76 
77 static inline void set_rs_rto(unsigned long *entry, phys_addr_t rto)
78 {
79 	*entry &= ZPCI_RTE_FLAG_MASK;
80 	*entry |= (rto & ZPCI_RTE_ADDR_MASK);
81 	*entry |= ZPCI_TABLE_TYPE_RSX;
82 }
83 
84 static inline void set_rt_sto(unsigned long *entry, phys_addr_t sto)
85 {
86 	*entry &= ZPCI_RTE_FLAG_MASK;
87 	*entry |= (sto & ZPCI_RTE_ADDR_MASK);
88 	*entry |= ZPCI_TABLE_TYPE_RTX;
89 }
90 
91 static inline void set_st_pto(unsigned long *entry, phys_addr_t pto)
92 {
93 	*entry &= ZPCI_STE_FLAG_MASK;
94 	*entry |= (pto & ZPCI_STE_ADDR_MASK);
95 	*entry |= ZPCI_TABLE_TYPE_SX;
96 }
97 
98 static inline void validate_rf_entry(unsigned long *entry)
99 {
100 	*entry &= ~ZPCI_TABLE_VALID_MASK;
101 	*entry &= ~ZPCI_TABLE_OFFSET_MASK;
102 	*entry |= ZPCI_TABLE_VALID;
103 	*entry |= ZPCI_TABLE_LEN_RFX;
104 }
105 
106 static inline void validate_rs_entry(unsigned long *entry)
107 {
108 	*entry &= ~ZPCI_TABLE_VALID_MASK;
109 	*entry &= ~ZPCI_TABLE_OFFSET_MASK;
110 	*entry |= ZPCI_TABLE_VALID;
111 	*entry |= ZPCI_TABLE_LEN_RSX;
112 }
113 
114 static inline void validate_rt_entry(unsigned long *entry)
115 {
116 	*entry &= ~ZPCI_TABLE_VALID_MASK;
117 	*entry &= ~ZPCI_TABLE_OFFSET_MASK;
118 	*entry |= ZPCI_TABLE_VALID;
119 	*entry |= ZPCI_TABLE_LEN_RTX;
120 }
121 
122 static inline void validate_st_entry(unsigned long *entry)
123 {
124 	*entry &= ~ZPCI_TABLE_VALID_MASK;
125 	*entry |= ZPCI_TABLE_VALID;
126 }
127 
128 static inline void invalidate_pt_entry(unsigned long *entry)
129 {
130 	WARN_ON_ONCE((*entry & ZPCI_PTE_VALID_MASK) == ZPCI_PTE_INVALID);
131 	*entry &= ~ZPCI_PTE_VALID_MASK;
132 	*entry |= ZPCI_PTE_INVALID;
133 }
134 
135 static inline void validate_pt_entry(unsigned long *entry)
136 {
137 	WARN_ON_ONCE((*entry & ZPCI_PTE_VALID_MASK) == ZPCI_PTE_VALID);
138 	*entry &= ~ZPCI_PTE_VALID_MASK;
139 	*entry |= ZPCI_PTE_VALID;
140 }
141 
142 static inline void entry_set_protected(unsigned long *entry)
143 {
144 	*entry &= ~ZPCI_TABLE_PROT_MASK;
145 	*entry |= ZPCI_TABLE_PROTECTED;
146 }
147 
148 static inline void entry_clr_protected(unsigned long *entry)
149 {
150 	*entry &= ~ZPCI_TABLE_PROT_MASK;
151 	*entry |= ZPCI_TABLE_UNPROTECTED;
152 }
153 
154 static inline int reg_entry_isvalid(unsigned long entry)
155 {
156 	return (entry & ZPCI_TABLE_VALID_MASK) == ZPCI_TABLE_VALID;
157 }
158 
159 static inline int pt_entry_isvalid(unsigned long entry)
160 {
161 	return (entry & ZPCI_PTE_VALID_MASK) == ZPCI_PTE_VALID;
162 }
163 
164 static inline unsigned long *get_rf_rso(unsigned long entry)
165 {
166 	if ((entry & ZPCI_TABLE_TYPE_MASK) == ZPCI_TABLE_TYPE_RFX)
167 		return phys_to_virt(entry & ZPCI_RTE_ADDR_MASK);
168 	else
169 		return NULL;
170 }
171 
172 static inline unsigned long *get_rs_rto(unsigned long entry)
173 {
174 	if ((entry & ZPCI_TABLE_TYPE_MASK) == ZPCI_TABLE_TYPE_RSX)
175 		return phys_to_virt(entry & ZPCI_RTE_ADDR_MASK);
176 	else
177 		return NULL;
178 }
179 
180 static inline unsigned long *get_rt_sto(unsigned long entry)
181 {
182 	if ((entry & ZPCI_TABLE_TYPE_MASK) == ZPCI_TABLE_TYPE_RTX)
183 		return phys_to_virt(entry & ZPCI_RTE_ADDR_MASK);
184 	else
185 		return NULL;
186 }
187 
188 static inline unsigned long *get_st_pto(unsigned long entry)
189 {
190 	if ((entry & ZPCI_TABLE_TYPE_MASK) == ZPCI_TABLE_TYPE_SX)
191 		return phys_to_virt(entry & ZPCI_STE_ADDR_MASK);
192 	else
193 		return NULL;
194 }
195 
196 static int __init dma_alloc_cpu_table_caches(void)
197 {
198 	dma_region_table_cache = kmem_cache_create("PCI_DMA_region_tables",
199 						   ZPCI_TABLE_SIZE,
200 						   ZPCI_TABLE_ALIGN,
201 						   0, NULL);
202 	if (!dma_region_table_cache)
203 		return -ENOMEM;
204 
205 	dma_page_table_cache = kmem_cache_create("PCI_DMA_page_tables",
206 						 ZPCI_PT_SIZE,
207 						 ZPCI_PT_ALIGN,
208 						 0, NULL);
209 	if (!dma_page_table_cache) {
210 		kmem_cache_destroy(dma_region_table_cache);
211 		return -ENOMEM;
212 	}
213 	return 0;
214 }
215 
216 static unsigned long *dma_alloc_cpu_table(gfp_t gfp)
217 {
218 	unsigned long *table, *entry;
219 
220 	table = kmem_cache_alloc(dma_region_table_cache, gfp);
221 	if (!table)
222 		return NULL;
223 
224 	for (entry = table; entry < table + ZPCI_TABLE_ENTRIES; entry++)
225 		*entry = ZPCI_TABLE_INVALID;
226 	return table;
227 }
228 
229 static void dma_free_cpu_table(void *table)
230 {
231 	kmem_cache_free(dma_region_table_cache, table);
232 }
233 
234 static void dma_free_page_table(void *table)
235 {
236 	kmem_cache_free(dma_page_table_cache, table);
237 }
238 
239 static void dma_free_seg_table(unsigned long entry)
240 {
241 	unsigned long *sto = get_rt_sto(entry);
242 	int sx;
243 
244 	for (sx = 0; sx < ZPCI_TABLE_ENTRIES; sx++)
245 		if (reg_entry_isvalid(sto[sx]))
246 			dma_free_page_table(get_st_pto(sto[sx]));
247 
248 	dma_free_cpu_table(sto);
249 }
250 
251 static void dma_free_rt_table(unsigned long entry)
252 {
253 	unsigned long *rto = get_rs_rto(entry);
254 	int rtx;
255 
256 	for (rtx = 0; rtx < ZPCI_TABLE_ENTRIES; rtx++)
257 		if (reg_entry_isvalid(rto[rtx]))
258 			dma_free_seg_table(rto[rtx]);
259 
260 	dma_free_cpu_table(rto);
261 }
262 
263 static void dma_free_rs_table(unsigned long entry)
264 {
265 	unsigned long *rso = get_rf_rso(entry);
266 	int rsx;
267 
268 	for (rsx = 0; rsx < ZPCI_TABLE_ENTRIES; rsx++)
269 		if (reg_entry_isvalid(rso[rsx]))
270 			dma_free_rt_table(rso[rsx]);
271 
272 	dma_free_cpu_table(rso);
273 }
274 
275 static void dma_cleanup_tables(struct s390_domain *domain)
276 {
277 	int rtx, rsx, rfx;
278 
279 	if (!domain->dma_table)
280 		return;
281 
282 	switch (domain->origin_type) {
283 	case ZPCI_TABLE_TYPE_RFX:
284 		for (rfx = 0; rfx < ZPCI_TABLE_ENTRIES; rfx++)
285 			if (reg_entry_isvalid(domain->dma_table[rfx]))
286 				dma_free_rs_table(domain->dma_table[rfx]);
287 		break;
288 	case ZPCI_TABLE_TYPE_RSX:
289 		for (rsx = 0; rsx < ZPCI_TABLE_ENTRIES; rsx++)
290 			if (reg_entry_isvalid(domain->dma_table[rsx]))
291 				dma_free_rt_table(domain->dma_table[rsx]);
292 		break;
293 	case ZPCI_TABLE_TYPE_RTX:
294 		for (rtx = 0; rtx < ZPCI_TABLE_ENTRIES; rtx++)
295 			if (reg_entry_isvalid(domain->dma_table[rtx]))
296 				dma_free_seg_table(domain->dma_table[rtx]);
297 		break;
298 	default:
299 		WARN_ONCE(1, "Invalid IOMMU table (%x)\n", domain->origin_type);
300 		return;
301 	}
302 
303 	dma_free_cpu_table(domain->dma_table);
304 }
305 
306 static unsigned long *dma_alloc_page_table(gfp_t gfp)
307 {
308 	unsigned long *table, *entry;
309 
310 	table = kmem_cache_alloc(dma_page_table_cache, gfp);
311 	if (!table)
312 		return NULL;
313 
314 	for (entry = table; entry < table + ZPCI_PT_ENTRIES; entry++)
315 		*entry = ZPCI_PTE_INVALID;
316 	return table;
317 }
318 
319 static unsigned long *dma_walk_rs_table(unsigned long *rso,
320 					dma_addr_t dma_addr, gfp_t gfp)
321 {
322 	unsigned int rsx = calc_rsx(dma_addr);
323 	unsigned long old_rse, rse;
324 	unsigned long *rsep, *rto;
325 
326 	rsep = &rso[rsx];
327 	rse = READ_ONCE(*rsep);
328 	if (reg_entry_isvalid(rse)) {
329 		rto = get_rs_rto(rse);
330 	} else {
331 		rto = dma_alloc_cpu_table(gfp);
332 		if (!rto)
333 			return NULL;
334 
335 		set_rs_rto(&rse, virt_to_phys(rto));
336 		validate_rs_entry(&rse);
337 		entry_clr_protected(&rse);
338 
339 		old_rse = cmpxchg(rsep, ZPCI_TABLE_INVALID, rse);
340 		if (old_rse != ZPCI_TABLE_INVALID) {
341 			/* Somone else was faster, use theirs */
342 			dma_free_cpu_table(rto);
343 			rto = get_rs_rto(old_rse);
344 		}
345 	}
346 	return rto;
347 }
348 
349 static unsigned long *dma_walk_rf_table(unsigned long *rfo,
350 					dma_addr_t dma_addr, gfp_t gfp)
351 {
352 	unsigned int rfx = calc_rfx(dma_addr);
353 	unsigned long old_rfe, rfe;
354 	unsigned long *rfep, *rso;
355 
356 	rfep = &rfo[rfx];
357 	rfe = READ_ONCE(*rfep);
358 	if (reg_entry_isvalid(rfe)) {
359 		rso = get_rf_rso(rfe);
360 	} else {
361 		rso = dma_alloc_cpu_table(gfp);
362 		if (!rso)
363 			return NULL;
364 
365 		set_rf_rso(&rfe, virt_to_phys(rso));
366 		validate_rf_entry(&rfe);
367 		entry_clr_protected(&rfe);
368 
369 		old_rfe = cmpxchg(rfep, ZPCI_TABLE_INVALID, rfe);
370 		if (old_rfe != ZPCI_TABLE_INVALID) {
371 			/* Somone else was faster, use theirs */
372 			dma_free_cpu_table(rso);
373 			rso = get_rf_rso(old_rfe);
374 		}
375 	}
376 
377 	if (!rso)
378 		return NULL;
379 
380 	return dma_walk_rs_table(rso, dma_addr, gfp);
381 }
382 
383 static unsigned long *dma_get_seg_table_origin(unsigned long *rtep, gfp_t gfp)
384 {
385 	unsigned long old_rte, rte;
386 	unsigned long *sto;
387 
388 	rte = READ_ONCE(*rtep);
389 	if (reg_entry_isvalid(rte)) {
390 		sto = get_rt_sto(rte);
391 	} else {
392 		sto = dma_alloc_cpu_table(gfp);
393 		if (!sto)
394 			return NULL;
395 
396 		set_rt_sto(&rte, virt_to_phys(sto));
397 		validate_rt_entry(&rte);
398 		entry_clr_protected(&rte);
399 
400 		old_rte = cmpxchg(rtep, ZPCI_TABLE_INVALID, rte);
401 		if (old_rte != ZPCI_TABLE_INVALID) {
402 			/* Somone else was faster, use theirs */
403 			dma_free_cpu_table(sto);
404 			sto = get_rt_sto(old_rte);
405 		}
406 	}
407 	return sto;
408 }
409 
410 static unsigned long *dma_get_page_table_origin(unsigned long *step, gfp_t gfp)
411 {
412 	unsigned long old_ste, ste;
413 	unsigned long *pto;
414 
415 	ste = READ_ONCE(*step);
416 	if (reg_entry_isvalid(ste)) {
417 		pto = get_st_pto(ste);
418 	} else {
419 		pto = dma_alloc_page_table(gfp);
420 		if (!pto)
421 			return NULL;
422 		set_st_pto(&ste, virt_to_phys(pto));
423 		validate_st_entry(&ste);
424 		entry_clr_protected(&ste);
425 
426 		old_ste = cmpxchg(step, ZPCI_TABLE_INVALID, ste);
427 		if (old_ste != ZPCI_TABLE_INVALID) {
428 			/* Somone else was faster, use theirs */
429 			dma_free_page_table(pto);
430 			pto = get_st_pto(old_ste);
431 		}
432 	}
433 	return pto;
434 }
435 
436 static unsigned long *dma_walk_region_tables(struct s390_domain *domain,
437 					     dma_addr_t dma_addr, gfp_t gfp)
438 {
439 	switch (domain->origin_type) {
440 	case ZPCI_TABLE_TYPE_RFX:
441 		return dma_walk_rf_table(domain->dma_table, dma_addr, gfp);
442 	case ZPCI_TABLE_TYPE_RSX:
443 		return dma_walk_rs_table(domain->dma_table, dma_addr, gfp);
444 	case ZPCI_TABLE_TYPE_RTX:
445 		return domain->dma_table;
446 	default:
447 		return NULL;
448 	}
449 }
450 
451 static unsigned long *dma_walk_cpu_trans(struct s390_domain *domain,
452 					 dma_addr_t dma_addr, gfp_t gfp)
453 {
454 	unsigned long *rto, *sto, *pto;
455 	unsigned int rtx, sx, px;
456 
457 	rto = dma_walk_region_tables(domain, dma_addr, gfp);
458 	if (!rto)
459 		return NULL;
460 
461 	rtx = calc_rtx(dma_addr);
462 	sto = dma_get_seg_table_origin(&rto[rtx], gfp);
463 	if (!sto)
464 		return NULL;
465 
466 	sx = calc_sx(dma_addr);
467 	pto = dma_get_page_table_origin(&sto[sx], gfp);
468 	if (!pto)
469 		return NULL;
470 
471 	px = calc_px(dma_addr);
472 	return &pto[px];
473 }
474 
475 static void dma_update_cpu_trans(unsigned long *ptep, phys_addr_t page_addr, int flags)
476 {
477 	unsigned long pte;
478 
479 	pte = READ_ONCE(*ptep);
480 	if (flags & ZPCI_PTE_INVALID) {
481 		invalidate_pt_entry(&pte);
482 	} else {
483 		set_pt_pfaa(&pte, page_addr);
484 		validate_pt_entry(&pte);
485 	}
486 
487 	if (flags & ZPCI_TABLE_PROTECTED)
488 		entry_set_protected(&pte);
489 	else
490 		entry_clr_protected(&pte);
491 
492 	xchg(ptep, pte);
493 }
494 
495 static struct s390_domain *to_s390_domain(struct iommu_domain *dom)
496 {
497 	return container_of(dom, struct s390_domain, domain);
498 }
499 
500 static bool s390_iommu_capable(struct device *dev, enum iommu_cap cap)
501 {
502 	struct zpci_dev *zdev = to_zpci_dev(dev);
503 
504 	switch (cap) {
505 	case IOMMU_CAP_CACHE_COHERENCY:
506 		return true;
507 	case IOMMU_CAP_DEFERRED_FLUSH:
508 		return zdev->pft != PCI_FUNC_TYPE_ISM;
509 	default:
510 		return false;
511 	}
512 }
513 
514 static inline u64 max_tbl_size(struct s390_domain *domain)
515 {
516 	switch (domain->origin_type) {
517 	case ZPCI_TABLE_TYPE_RTX:
518 		return ZPCI_TABLE_SIZE_RT - 1;
519 	case ZPCI_TABLE_TYPE_RSX:
520 		return ZPCI_TABLE_SIZE_RS - 1;
521 	case ZPCI_TABLE_TYPE_RFX:
522 		return U64_MAX;
523 	default:
524 		return 0;
525 	}
526 }
527 
528 static struct iommu_domain *s390_domain_alloc_paging(struct device *dev)
529 {
530 	struct zpci_dev *zdev = to_zpci_dev(dev);
531 	struct s390_domain *s390_domain;
532 	u64 aperture_size;
533 
534 	s390_domain = kzalloc(sizeof(*s390_domain), GFP_KERNEL);
535 	if (!s390_domain)
536 		return NULL;
537 
538 	s390_domain->dma_table = dma_alloc_cpu_table(GFP_KERNEL);
539 	if (!s390_domain->dma_table) {
540 		kfree(s390_domain);
541 		return NULL;
542 	}
543 
544 	aperture_size = min(s390_iommu_aperture,
545 			    zdev->end_dma - zdev->start_dma + 1);
546 	if (aperture_size <= (ZPCI_TABLE_SIZE_RT - zdev->start_dma)) {
547 		s390_domain->origin_type = ZPCI_TABLE_TYPE_RTX;
548 	} else if (aperture_size <= (ZPCI_TABLE_SIZE_RS - zdev->start_dma) &&
549 		  (zdev->dtsm & ZPCI_IOTA_DT_RS)) {
550 		s390_domain->origin_type = ZPCI_TABLE_TYPE_RSX;
551 	} else if (zdev->dtsm & ZPCI_IOTA_DT_RF) {
552 		s390_domain->origin_type = ZPCI_TABLE_TYPE_RFX;
553 	} else {
554 		/* Assume RTX available */
555 		s390_domain->origin_type = ZPCI_TABLE_TYPE_RTX;
556 		aperture_size = ZPCI_TABLE_SIZE_RT - zdev->start_dma;
557 	}
558 	zdev->end_dma = zdev->start_dma + aperture_size - 1;
559 
560 	s390_domain->domain.pgsize_bitmap = SZ_4K;
561 	s390_domain->domain.geometry.force_aperture = true;
562 	s390_domain->domain.geometry.aperture_start = 0;
563 	s390_domain->domain.geometry.aperture_end = max_tbl_size(s390_domain);
564 
565 	spin_lock_init(&s390_domain->list_lock);
566 	INIT_LIST_HEAD_RCU(&s390_domain->devices);
567 
568 	return &s390_domain->domain;
569 }
570 
571 static void s390_iommu_rcu_free_domain(struct rcu_head *head)
572 {
573 	struct s390_domain *s390_domain = container_of(head, struct s390_domain, rcu);
574 
575 	dma_cleanup_tables(s390_domain);
576 	kfree(s390_domain);
577 }
578 
579 static void s390_domain_free(struct iommu_domain *domain)
580 {
581 	struct s390_domain *s390_domain = to_s390_domain(domain);
582 
583 	rcu_read_lock();
584 	WARN_ON(!list_empty(&s390_domain->devices));
585 	rcu_read_unlock();
586 
587 	call_rcu(&s390_domain->rcu, s390_iommu_rcu_free_domain);
588 }
589 
590 static void zdev_s390_domain_update(struct zpci_dev *zdev,
591 				    struct iommu_domain *domain)
592 {
593 	unsigned long flags;
594 
595 	spin_lock_irqsave(&zdev->dom_lock, flags);
596 	zdev->s390_domain = domain;
597 	spin_unlock_irqrestore(&zdev->dom_lock, flags);
598 }
599 
600 static u64 get_iota_region_flag(struct s390_domain *domain)
601 {
602 	switch (domain->origin_type) {
603 	case ZPCI_TABLE_TYPE_RTX:
604 		return ZPCI_IOTA_RTTO_FLAG;
605 	case ZPCI_TABLE_TYPE_RSX:
606 		return ZPCI_IOTA_RSTO_FLAG;
607 	case ZPCI_TABLE_TYPE_RFX:
608 		return ZPCI_IOTA_RFTO_FLAG;
609 	default:
610 		WARN_ONCE(1, "Invalid IOMMU table (%x)\n", domain->origin_type);
611 		return 0;
612 	}
613 }
614 
615 static bool reg_ioat_propagate_error(int cc, u8 status)
616 {
617 	/*
618 	 * If the device is in the error state the reset routine
619 	 * will register the IOAT of the newly set domain on re-enable
620 	 */
621 	if (cc == ZPCI_CC_ERR && status == ZPCI_PCI_ST_FUNC_NOT_AVAIL)
622 		return false;
623 	/*
624 	 * If the device was removed treat registration as success
625 	 * and let the subsequent error event trigger tear down.
626 	 */
627 	if (cc == ZPCI_CC_INVAL_HANDLE)
628 		return false;
629 	return cc != ZPCI_CC_OK;
630 }
631 
632 static int s390_iommu_domain_reg_ioat(struct zpci_dev *zdev,
633 				      struct iommu_domain *domain, u8 *status)
634 {
635 	struct s390_domain *s390_domain;
636 	int rc = 0;
637 	u64 iota;
638 
639 	switch (domain->type) {
640 	case IOMMU_DOMAIN_IDENTITY:
641 		rc = zpci_register_ioat(zdev, 0, zdev->start_dma,
642 					zdev->end_dma, 0, status);
643 		break;
644 	case IOMMU_DOMAIN_BLOCKED:
645 		/* Nothing to do in this case */
646 		break;
647 	default:
648 		s390_domain = to_s390_domain(domain);
649 		iota = virt_to_phys(s390_domain->dma_table) |
650 		       get_iota_region_flag(s390_domain);
651 		rc = zpci_register_ioat(zdev, 0, zdev->start_dma,
652 					zdev->end_dma, iota, status);
653 	}
654 
655 	return rc;
656 }
657 
658 int zpci_iommu_register_ioat(struct zpci_dev *zdev, u8 *status)
659 {
660 	unsigned long flags;
661 	int rc;
662 
663 	spin_lock_irqsave(&zdev->dom_lock, flags);
664 
665 	rc = s390_iommu_domain_reg_ioat(zdev, zdev->s390_domain, status);
666 
667 	spin_unlock_irqrestore(&zdev->dom_lock, flags);
668 
669 	return rc;
670 }
671 
672 static int blocking_domain_attach_device(struct iommu_domain *domain,
673 					 struct device *dev)
674 {
675 	struct zpci_dev *zdev = to_zpci_dev(dev);
676 	struct s390_domain *s390_domain;
677 	unsigned long flags;
678 
679 	if (zdev->s390_domain->type == IOMMU_DOMAIN_BLOCKED)
680 		return 0;
681 
682 	s390_domain = to_s390_domain(zdev->s390_domain);
683 	if (zdev->dma_table) {
684 		spin_lock_irqsave(&s390_domain->list_lock, flags);
685 		list_del_rcu(&zdev->iommu_list);
686 		spin_unlock_irqrestore(&s390_domain->list_lock, flags);
687 	}
688 
689 	zpci_unregister_ioat(zdev, 0);
690 	zdev->dma_table = NULL;
691 	zdev_s390_domain_update(zdev, domain);
692 
693 	return 0;
694 }
695 
696 static int s390_iommu_attach_device(struct iommu_domain *domain,
697 				    struct device *dev)
698 {
699 	struct s390_domain *s390_domain = to_s390_domain(domain);
700 	struct zpci_dev *zdev = to_zpci_dev(dev);
701 	unsigned long flags;
702 	u8 status;
703 	int cc;
704 
705 	if (!zdev)
706 		return -ENODEV;
707 
708 	if (WARN_ON(domain->geometry.aperture_start > zdev->end_dma ||
709 		domain->geometry.aperture_end < zdev->start_dma))
710 		return -EINVAL;
711 
712 	blocking_domain_attach_device(&blocking_domain, dev);
713 
714 	/* If we fail now DMA remains blocked via blocking domain */
715 	cc = s390_iommu_domain_reg_ioat(zdev, domain, &status);
716 	if (reg_ioat_propagate_error(cc, status))
717 		return -EIO;
718 	zdev->dma_table = s390_domain->dma_table;
719 	zdev_s390_domain_update(zdev, domain);
720 
721 	spin_lock_irqsave(&s390_domain->list_lock, flags);
722 	list_add_rcu(&zdev->iommu_list, &s390_domain->devices);
723 	spin_unlock_irqrestore(&s390_domain->list_lock, flags);
724 
725 	return 0;
726 }
727 
728 static void s390_iommu_get_resv_regions(struct device *dev,
729 					struct list_head *list)
730 {
731 	struct zpci_dev *zdev = to_zpci_dev(dev);
732 	struct iommu_resv_region *region;
733 	u64 max_size, end_resv;
734 	unsigned long flags;
735 
736 	if (zdev->start_dma) {
737 		region = iommu_alloc_resv_region(0, zdev->start_dma, 0,
738 						 IOMMU_RESV_RESERVED, GFP_KERNEL);
739 		if (!region)
740 			return;
741 		list_add_tail(&region->list, list);
742 	}
743 
744 	spin_lock_irqsave(&zdev->dom_lock, flags);
745 	if (zdev->s390_domain->type == IOMMU_DOMAIN_BLOCKED ||
746 	    zdev->s390_domain->type == IOMMU_DOMAIN_IDENTITY) {
747 		spin_unlock_irqrestore(&zdev->dom_lock, flags);
748 		return;
749 	}
750 
751 	max_size = max_tbl_size(to_s390_domain(zdev->s390_domain));
752 	spin_unlock_irqrestore(&zdev->dom_lock, flags);
753 
754 	if (zdev->end_dma < max_size) {
755 		end_resv = max_size - zdev->end_dma;
756 		region = iommu_alloc_resv_region(zdev->end_dma + 1, end_resv,
757 						 0, IOMMU_RESV_RESERVED,
758 						 GFP_KERNEL);
759 		if (!region)
760 			return;
761 		list_add_tail(&region->list, list);
762 	}
763 }
764 
765 static struct iommu_device *s390_iommu_probe_device(struct device *dev)
766 {
767 	struct zpci_dev *zdev;
768 
769 	if (!dev_is_pci(dev))
770 		return ERR_PTR(-ENODEV);
771 
772 	zdev = to_zpci_dev(dev);
773 
774 	if (zdev->start_dma > zdev->end_dma)
775 		return ERR_PTR(-EINVAL);
776 
777 	if (zdev->tlb_refresh)
778 		dev->iommu->shadow_on_flush = 1;
779 
780 	/* Start with DMA blocked */
781 	spin_lock_init(&zdev->dom_lock);
782 	zdev_s390_domain_update(zdev, &blocking_domain);
783 
784 	return &zdev->iommu_dev;
785 }
786 
787 static int zpci_refresh_all(struct zpci_dev *zdev)
788 {
789 	return zpci_refresh_trans((u64)zdev->fh << 32, zdev->start_dma,
790 				  zdev->end_dma - zdev->start_dma + 1);
791 }
792 
793 static void s390_iommu_flush_iotlb_all(struct iommu_domain *domain)
794 {
795 	struct s390_domain *s390_domain = to_s390_domain(domain);
796 	struct zpci_dev *zdev;
797 
798 	rcu_read_lock();
799 	list_for_each_entry_rcu(zdev, &s390_domain->devices, iommu_list) {
800 		atomic64_inc(&s390_domain->ctrs.global_rpcits);
801 		zpci_refresh_all(zdev);
802 	}
803 	rcu_read_unlock();
804 }
805 
806 static void s390_iommu_iotlb_sync(struct iommu_domain *domain,
807 				  struct iommu_iotlb_gather *gather)
808 {
809 	struct s390_domain *s390_domain = to_s390_domain(domain);
810 	size_t size = gather->end - gather->start + 1;
811 	struct zpci_dev *zdev;
812 
813 	/* If gather was never added to there is nothing to flush */
814 	if (!gather->end)
815 		return;
816 
817 	rcu_read_lock();
818 	list_for_each_entry_rcu(zdev, &s390_domain->devices, iommu_list) {
819 		atomic64_inc(&s390_domain->ctrs.sync_rpcits);
820 		zpci_refresh_trans((u64)zdev->fh << 32, gather->start,
821 				   size);
822 	}
823 	rcu_read_unlock();
824 }
825 
826 static int s390_iommu_iotlb_sync_map(struct iommu_domain *domain,
827 				     unsigned long iova, size_t size)
828 {
829 	struct s390_domain *s390_domain = to_s390_domain(domain);
830 	struct zpci_dev *zdev;
831 	int ret = 0;
832 
833 	rcu_read_lock();
834 	list_for_each_entry_rcu(zdev, &s390_domain->devices, iommu_list) {
835 		if (!zdev->tlb_refresh)
836 			continue;
837 		atomic64_inc(&s390_domain->ctrs.sync_map_rpcits);
838 		ret = zpci_refresh_trans((u64)zdev->fh << 32,
839 					 iova, size);
840 		/*
841 		 * let the hypervisor discover invalidated entries
842 		 * allowing it to free IOVAs and unpin pages
843 		 */
844 		if (ret == -ENOMEM) {
845 			ret = zpci_refresh_all(zdev);
846 			if (ret)
847 				break;
848 		}
849 	}
850 	rcu_read_unlock();
851 
852 	return ret;
853 }
854 
855 static int s390_iommu_validate_trans(struct s390_domain *s390_domain,
856 				     phys_addr_t pa, dma_addr_t dma_addr,
857 				     unsigned long nr_pages, int flags,
858 				     gfp_t gfp)
859 {
860 	phys_addr_t page_addr = pa & PAGE_MASK;
861 	unsigned long *entry;
862 	unsigned long i;
863 	int rc;
864 
865 	for (i = 0; i < nr_pages; i++) {
866 		entry = dma_walk_cpu_trans(s390_domain, dma_addr, gfp);
867 		if (unlikely(!entry)) {
868 			rc = -ENOMEM;
869 			goto undo_cpu_trans;
870 		}
871 		dma_update_cpu_trans(entry, page_addr, flags);
872 		page_addr += PAGE_SIZE;
873 		dma_addr += PAGE_SIZE;
874 	}
875 
876 	return 0;
877 
878 undo_cpu_trans:
879 	while (i-- > 0) {
880 		dma_addr -= PAGE_SIZE;
881 		entry = dma_walk_cpu_trans(s390_domain, dma_addr, gfp);
882 		if (!entry)
883 			break;
884 		dma_update_cpu_trans(entry, 0, ZPCI_PTE_INVALID);
885 	}
886 
887 	return rc;
888 }
889 
890 static int s390_iommu_invalidate_trans(struct s390_domain *s390_domain,
891 				       dma_addr_t dma_addr, unsigned long nr_pages)
892 {
893 	unsigned long *entry;
894 	unsigned long i;
895 	int rc = 0;
896 
897 	for (i = 0; i < nr_pages; i++) {
898 		entry = dma_walk_cpu_trans(s390_domain, dma_addr, GFP_ATOMIC);
899 		if (unlikely(!entry)) {
900 			rc = -EINVAL;
901 			break;
902 		}
903 		dma_update_cpu_trans(entry, 0, ZPCI_PTE_INVALID);
904 		dma_addr += PAGE_SIZE;
905 	}
906 
907 	return rc;
908 }
909 
910 static int s390_iommu_map_pages(struct iommu_domain *domain,
911 				unsigned long iova, phys_addr_t paddr,
912 				size_t pgsize, size_t pgcount,
913 				int prot, gfp_t gfp, size_t *mapped)
914 {
915 	struct s390_domain *s390_domain = to_s390_domain(domain);
916 	size_t size = pgcount << __ffs(pgsize);
917 	int flags = ZPCI_PTE_VALID, rc = 0;
918 
919 	if (pgsize != SZ_4K)
920 		return -EINVAL;
921 
922 	if (iova < s390_domain->domain.geometry.aperture_start ||
923 	    (iova + size - 1) > s390_domain->domain.geometry.aperture_end)
924 		return -EINVAL;
925 
926 	if (!IS_ALIGNED(iova | paddr, pgsize))
927 		return -EINVAL;
928 
929 	if (!(prot & IOMMU_WRITE))
930 		flags |= ZPCI_TABLE_PROTECTED;
931 
932 	rc = s390_iommu_validate_trans(s390_domain, paddr, iova,
933 				     pgcount, flags, gfp);
934 	if (!rc) {
935 		*mapped = size;
936 		atomic64_add(pgcount, &s390_domain->ctrs.mapped_pages);
937 	}
938 
939 	return rc;
940 }
941 
942 static unsigned long *get_rso_from_iova(struct s390_domain *domain,
943 					dma_addr_t iova)
944 {
945 	unsigned long *rfo;
946 	unsigned long rfe;
947 	unsigned int rfx;
948 
949 	switch (domain->origin_type) {
950 	case ZPCI_TABLE_TYPE_RFX:
951 		rfo = domain->dma_table;
952 		rfx = calc_rfx(iova);
953 		rfe = READ_ONCE(rfo[rfx]);
954 		if (!reg_entry_isvalid(rfe))
955 			return NULL;
956 		return get_rf_rso(rfe);
957 	case ZPCI_TABLE_TYPE_RSX:
958 		return domain->dma_table;
959 	default:
960 		return NULL;
961 	}
962 }
963 
964 static unsigned long *get_rto_from_iova(struct s390_domain *domain,
965 					dma_addr_t iova)
966 {
967 	unsigned long *rso;
968 	unsigned long rse;
969 	unsigned int rsx;
970 
971 	switch (domain->origin_type) {
972 	case ZPCI_TABLE_TYPE_RFX:
973 	case ZPCI_TABLE_TYPE_RSX:
974 		rso = get_rso_from_iova(domain, iova);
975 		rsx = calc_rsx(iova);
976 		rse = READ_ONCE(rso[rsx]);
977 		if (!reg_entry_isvalid(rse))
978 			return NULL;
979 		return get_rs_rto(rse);
980 	case ZPCI_TABLE_TYPE_RTX:
981 		return domain->dma_table;
982 	default:
983 		return NULL;
984 	}
985 }
986 
987 static phys_addr_t s390_iommu_iova_to_phys(struct iommu_domain *domain,
988 					   dma_addr_t iova)
989 {
990 	struct s390_domain *s390_domain = to_s390_domain(domain);
991 	unsigned long *rto, *sto, *pto;
992 	unsigned long ste, pte, rte;
993 	unsigned int rtx, sx, px;
994 	phys_addr_t phys = 0;
995 
996 	if (iova < domain->geometry.aperture_start ||
997 	    iova > domain->geometry.aperture_end)
998 		return 0;
999 
1000 	rto = get_rto_from_iova(s390_domain, iova);
1001 	if (!rto)
1002 		return 0;
1003 
1004 	rtx = calc_rtx(iova);
1005 	sx = calc_sx(iova);
1006 	px = calc_px(iova);
1007 
1008 	rte = READ_ONCE(rto[rtx]);
1009 	if (reg_entry_isvalid(rte)) {
1010 		sto = get_rt_sto(rte);
1011 		ste = READ_ONCE(sto[sx]);
1012 		if (reg_entry_isvalid(ste)) {
1013 			pto = get_st_pto(ste);
1014 			pte = READ_ONCE(pto[px]);
1015 			if (pt_entry_isvalid(pte))
1016 				phys = pte & ZPCI_PTE_ADDR_MASK;
1017 		}
1018 	}
1019 
1020 	return phys;
1021 }
1022 
1023 static size_t s390_iommu_unmap_pages(struct iommu_domain *domain,
1024 				     unsigned long iova,
1025 				     size_t pgsize, size_t pgcount,
1026 				     struct iommu_iotlb_gather *gather)
1027 {
1028 	struct s390_domain *s390_domain = to_s390_domain(domain);
1029 	size_t size = pgcount << __ffs(pgsize);
1030 	int rc;
1031 
1032 	if (WARN_ON(iova < s390_domain->domain.geometry.aperture_start ||
1033 	    (iova + size - 1) > s390_domain->domain.geometry.aperture_end))
1034 		return 0;
1035 
1036 	rc = s390_iommu_invalidate_trans(s390_domain, iova, pgcount);
1037 	if (rc)
1038 		return 0;
1039 
1040 	iommu_iotlb_gather_add_range(gather, iova, size);
1041 	atomic64_add(pgcount, &s390_domain->ctrs.unmapped_pages);
1042 
1043 	return size;
1044 }
1045 
1046 struct zpci_iommu_ctrs *zpci_get_iommu_ctrs(struct zpci_dev *zdev)
1047 {
1048 	struct s390_domain *s390_domain;
1049 
1050 	lockdep_assert_held(&zdev->dom_lock);
1051 
1052 	if (zdev->s390_domain->type == IOMMU_DOMAIN_BLOCKED ||
1053 	    zdev->s390_domain->type == IOMMU_DOMAIN_IDENTITY)
1054 		return NULL;
1055 
1056 	s390_domain = to_s390_domain(zdev->s390_domain);
1057 	return &s390_domain->ctrs;
1058 }
1059 
1060 int zpci_init_iommu(struct zpci_dev *zdev)
1061 {
1062 	int rc = 0;
1063 
1064 	rc = iommu_device_sysfs_add(&zdev->iommu_dev, NULL, NULL,
1065 				    "s390-iommu.%08x", zdev->fid);
1066 	if (rc)
1067 		goto out_err;
1068 
1069 	if (zdev->rtr_avail) {
1070 		rc = iommu_device_register(&zdev->iommu_dev,
1071 					   &s390_iommu_rtr_ops, NULL);
1072 	} else {
1073 		rc = iommu_device_register(&zdev->iommu_dev, &s390_iommu_ops,
1074 					   NULL);
1075 	}
1076 	if (rc)
1077 		goto out_sysfs;
1078 
1079 	return 0;
1080 
1081 out_sysfs:
1082 	iommu_device_sysfs_remove(&zdev->iommu_dev);
1083 
1084 out_err:
1085 	return rc;
1086 }
1087 
1088 void zpci_destroy_iommu(struct zpci_dev *zdev)
1089 {
1090 	iommu_device_unregister(&zdev->iommu_dev);
1091 	iommu_device_sysfs_remove(&zdev->iommu_dev);
1092 }
1093 
1094 static int __init s390_iommu_setup(char *str)
1095 {
1096 	if (!strcmp(str, "strict")) {
1097 		pr_warn("s390_iommu=strict deprecated; use iommu.strict=1 instead\n");
1098 		iommu_set_dma_strict();
1099 	}
1100 	return 1;
1101 }
1102 
1103 __setup("s390_iommu=", s390_iommu_setup);
1104 
1105 static int __init s390_iommu_aperture_setup(char *str)
1106 {
1107 	if (kstrtou32(str, 10, &s390_iommu_aperture_factor))
1108 		s390_iommu_aperture_factor = 1;
1109 	return 1;
1110 }
1111 
1112 __setup("s390_iommu_aperture=", s390_iommu_aperture_setup);
1113 
1114 static int __init s390_iommu_init(void)
1115 {
1116 	int rc;
1117 
1118 	iommu_dma_forcedac = true;
1119 	s390_iommu_aperture = (u64)virt_to_phys(high_memory);
1120 	if (!s390_iommu_aperture_factor)
1121 		s390_iommu_aperture = ULONG_MAX;
1122 	else
1123 		s390_iommu_aperture *= s390_iommu_aperture_factor;
1124 
1125 	rc = dma_alloc_cpu_table_caches();
1126 	if (rc)
1127 		return rc;
1128 
1129 	return rc;
1130 }
1131 subsys_initcall(s390_iommu_init);
1132 
1133 static int s390_attach_dev_identity(struct iommu_domain *domain,
1134 				    struct device *dev)
1135 {
1136 	struct zpci_dev *zdev = to_zpci_dev(dev);
1137 	u8 status;
1138 	int cc;
1139 
1140 	blocking_domain_attach_device(&blocking_domain, dev);
1141 
1142 	/* If we fail now DMA remains blocked via blocking domain */
1143 	cc = s390_iommu_domain_reg_ioat(zdev, domain, &status);
1144 	if (reg_ioat_propagate_error(cc, status))
1145 		return -EIO;
1146 
1147 	zdev_s390_domain_update(zdev, domain);
1148 
1149 	return 0;
1150 }
1151 
1152 static const struct iommu_domain_ops s390_identity_ops = {
1153 	.attach_dev = s390_attach_dev_identity,
1154 };
1155 
1156 static struct iommu_domain s390_identity_domain = {
1157 	.type = IOMMU_DOMAIN_IDENTITY,
1158 	.ops = &s390_identity_ops,
1159 };
1160 
1161 static struct iommu_domain blocking_domain = {
1162 	.type = IOMMU_DOMAIN_BLOCKED,
1163 	.ops = &(const struct iommu_domain_ops) {
1164 		.attach_dev	= blocking_domain_attach_device,
1165 	}
1166 };
1167 
1168 #define S390_IOMMU_COMMON_OPS() \
1169 	.blocked_domain		= &blocking_domain, \
1170 	.release_domain		= &blocking_domain, \
1171 	.capable = s390_iommu_capable, \
1172 	.domain_alloc_paging = s390_domain_alloc_paging, \
1173 	.probe_device = s390_iommu_probe_device, \
1174 	.device_group = generic_device_group, \
1175 	.get_resv_regions = s390_iommu_get_resv_regions, \
1176 	.default_domain_ops = &(const struct iommu_domain_ops) { \
1177 		.attach_dev	= s390_iommu_attach_device, \
1178 		.map_pages	= s390_iommu_map_pages, \
1179 		.unmap_pages	= s390_iommu_unmap_pages, \
1180 		.flush_iotlb_all = s390_iommu_flush_iotlb_all, \
1181 		.iotlb_sync      = s390_iommu_iotlb_sync, \
1182 		.iotlb_sync_map  = s390_iommu_iotlb_sync_map, \
1183 		.iova_to_phys	= s390_iommu_iova_to_phys, \
1184 		.free		= s390_domain_free, \
1185 	}
1186 
1187 static const struct iommu_ops s390_iommu_ops = {
1188 	S390_IOMMU_COMMON_OPS()
1189 };
1190 
1191 static const struct iommu_ops s390_iommu_rtr_ops = {
1192 	.identity_domain	= &s390_identity_domain,
1193 	S390_IOMMU_COMMON_OPS()
1194 };
1195