xref: /linux/drivers/iommu/intel/pasid.c (revision 42b16d3ac371a2fac9b6f08fd75f23f34ba3955a)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * intel-pasid.c - PASID idr, table and entry manipulation
4  *
5  * Copyright (C) 2018 Intel Corporation
6  *
7  * Author: Lu Baolu <baolu.lu@linux.intel.com>
8  */
9 
10 #define pr_fmt(fmt)	"DMAR: " fmt
11 
12 #include <linux/bitops.h>
13 #include <linux/cpufeature.h>
14 #include <linux/dmar.h>
15 #include <linux/iommu.h>
16 #include <linux/memory.h>
17 #include <linux/pci.h>
18 #include <linux/pci-ats.h>
19 #include <linux/spinlock.h>
20 
21 #include "iommu.h"
22 #include "pasid.h"
23 #include "../iommu-pages.h"
24 
25 /*
26  * Intel IOMMU system wide PASID name space:
27  */
28 u32 intel_pasid_max_id = PASID_MAX;
29 
30 /*
31  * Per device pasid table management:
32  */
33 
34 /*
35  * Allocate a pasid table for @dev. It should be called in a
36  * single-thread context.
37  */
intel_pasid_alloc_table(struct device * dev)38 int intel_pasid_alloc_table(struct device *dev)
39 {
40 	struct device_domain_info *info;
41 	struct pasid_table *pasid_table;
42 	struct pasid_dir_entry *dir;
43 	u32 max_pasid = 0;
44 	int order, size;
45 
46 	might_sleep();
47 	info = dev_iommu_priv_get(dev);
48 	if (WARN_ON(!info || !dev_is_pci(dev)))
49 		return -ENODEV;
50 	if (WARN_ON(info->pasid_table))
51 		return -EEXIST;
52 
53 	pasid_table = kzalloc(sizeof(*pasid_table), GFP_KERNEL);
54 	if (!pasid_table)
55 		return -ENOMEM;
56 
57 	if (info->pasid_supported)
58 		max_pasid = min_t(u32, pci_max_pasids(to_pci_dev(dev)),
59 				  intel_pasid_max_id);
60 
61 	size = max_pasid >> (PASID_PDE_SHIFT - 3);
62 	order = size ? get_order(size) : 0;
63 	dir = iommu_alloc_pages_node(info->iommu->node, GFP_KERNEL, order);
64 	if (!dir) {
65 		kfree(pasid_table);
66 		return -ENOMEM;
67 	}
68 
69 	pasid_table->table = dir;
70 	pasid_table->order = order;
71 	pasid_table->max_pasid = 1 << (order + PAGE_SHIFT + 3);
72 	info->pasid_table = pasid_table;
73 
74 	if (!ecap_coherent(info->iommu->ecap))
75 		clflush_cache_range(pasid_table->table, (1 << order) * PAGE_SIZE);
76 
77 	return 0;
78 }
79 
intel_pasid_free_table(struct device * dev)80 void intel_pasid_free_table(struct device *dev)
81 {
82 	struct device_domain_info *info;
83 	struct pasid_table *pasid_table;
84 	struct pasid_dir_entry *dir;
85 	struct pasid_entry *table;
86 	int i, max_pde;
87 
88 	info = dev_iommu_priv_get(dev);
89 	if (!info || !dev_is_pci(dev) || !info->pasid_table)
90 		return;
91 
92 	pasid_table = info->pasid_table;
93 	info->pasid_table = NULL;
94 
95 	/* Free scalable mode PASID directory tables: */
96 	dir = pasid_table->table;
97 	max_pde = pasid_table->max_pasid >> PASID_PDE_SHIFT;
98 	for (i = 0; i < max_pde; i++) {
99 		table = get_pasid_table_from_pde(&dir[i]);
100 		iommu_free_page(table);
101 	}
102 
103 	iommu_free_pages(pasid_table->table, pasid_table->order);
104 	kfree(pasid_table);
105 }
106 
intel_pasid_get_table(struct device * dev)107 struct pasid_table *intel_pasid_get_table(struct device *dev)
108 {
109 	struct device_domain_info *info;
110 
111 	info = dev_iommu_priv_get(dev);
112 	if (!info)
113 		return NULL;
114 
115 	return info->pasid_table;
116 }
117 
intel_pasid_get_dev_max_id(struct device * dev)118 static int intel_pasid_get_dev_max_id(struct device *dev)
119 {
120 	struct device_domain_info *info;
121 
122 	info = dev_iommu_priv_get(dev);
123 	if (!info || !info->pasid_table)
124 		return 0;
125 
126 	return info->pasid_table->max_pasid;
127 }
128 
intel_pasid_get_entry(struct device * dev,u32 pasid)129 static struct pasid_entry *intel_pasid_get_entry(struct device *dev, u32 pasid)
130 {
131 	struct device_domain_info *info;
132 	struct pasid_table *pasid_table;
133 	struct pasid_dir_entry *dir;
134 	struct pasid_entry *entries;
135 	int dir_index, index;
136 
137 	pasid_table = intel_pasid_get_table(dev);
138 	if (WARN_ON(!pasid_table || pasid >= intel_pasid_get_dev_max_id(dev)))
139 		return NULL;
140 
141 	dir = pasid_table->table;
142 	info = dev_iommu_priv_get(dev);
143 	dir_index = pasid >> PASID_PDE_SHIFT;
144 	index = pasid & PASID_PTE_MASK;
145 
146 retry:
147 	entries = get_pasid_table_from_pde(&dir[dir_index]);
148 	if (!entries) {
149 		u64 tmp;
150 
151 		entries = iommu_alloc_page_node(info->iommu->node, GFP_ATOMIC);
152 		if (!entries)
153 			return NULL;
154 
155 		/*
156 		 * The pasid directory table entry won't be freed after
157 		 * allocation. No worry about the race with free and
158 		 * clear. However, this entry might be populated by others
159 		 * while we are preparing it. Use theirs with a retry.
160 		 */
161 		tmp = 0ULL;
162 		if (!try_cmpxchg64(&dir[dir_index].val, &tmp,
163 				   (u64)virt_to_phys(entries) | PASID_PTE_PRESENT)) {
164 			iommu_free_page(entries);
165 			goto retry;
166 		}
167 		if (!ecap_coherent(info->iommu->ecap)) {
168 			clflush_cache_range(entries, VTD_PAGE_SIZE);
169 			clflush_cache_range(&dir[dir_index].val, sizeof(*dir));
170 		}
171 	}
172 
173 	return &entries[index];
174 }
175 
176 /*
177  * Interfaces for PASID table entry manipulation:
178  */
179 static void
intel_pasid_clear_entry(struct device * dev,u32 pasid,bool fault_ignore)180 intel_pasid_clear_entry(struct device *dev, u32 pasid, bool fault_ignore)
181 {
182 	struct pasid_entry *pe;
183 
184 	pe = intel_pasid_get_entry(dev, pasid);
185 	if (WARN_ON(!pe))
186 		return;
187 
188 	if (fault_ignore && pasid_pte_is_present(pe))
189 		pasid_clear_entry_with_fpd(pe);
190 	else
191 		pasid_clear_entry(pe);
192 }
193 
194 static void
pasid_cache_invalidation_with_pasid(struct intel_iommu * iommu,u16 did,u32 pasid)195 pasid_cache_invalidation_with_pasid(struct intel_iommu *iommu,
196 				    u16 did, u32 pasid)
197 {
198 	struct qi_desc desc;
199 
200 	desc.qw0 = QI_PC_DID(did) | QI_PC_GRAN(QI_PC_PASID_SEL) |
201 		QI_PC_PASID(pasid) | QI_PC_TYPE;
202 	desc.qw1 = 0;
203 	desc.qw2 = 0;
204 	desc.qw3 = 0;
205 
206 	qi_submit_sync(iommu, &desc, 1, 0);
207 }
208 
209 static void
devtlb_invalidation_with_pasid(struct intel_iommu * iommu,struct device * dev,u32 pasid)210 devtlb_invalidation_with_pasid(struct intel_iommu *iommu,
211 			       struct device *dev, u32 pasid)
212 {
213 	struct device_domain_info *info;
214 	u16 sid, qdep, pfsid;
215 
216 	info = dev_iommu_priv_get(dev);
217 	if (!info || !info->ats_enabled)
218 		return;
219 
220 	if (pci_dev_is_disconnected(to_pci_dev(dev)))
221 		return;
222 
223 	sid = info->bus << 8 | info->devfn;
224 	qdep = info->ats_qdep;
225 	pfsid = info->pfsid;
226 
227 	/*
228 	 * When PASID 0 is used, it indicates RID2PASID(DMA request w/o PASID),
229 	 * devTLB flush w/o PASID should be used. For non-zero PASID under
230 	 * SVA usage, device could do DMA with multiple PASIDs. It is more
231 	 * efficient to flush devTLB specific to the PASID.
232 	 */
233 	if (pasid == IOMMU_NO_PASID)
234 		qi_flush_dev_iotlb(iommu, sid, pfsid, qdep, 0, 64 - VTD_PAGE_SHIFT);
235 	else
236 		qi_flush_dev_iotlb_pasid(iommu, sid, pfsid, pasid, qdep, 0, 64 - VTD_PAGE_SHIFT);
237 }
238 
intel_pasid_tear_down_entry(struct intel_iommu * iommu,struct device * dev,u32 pasid,bool fault_ignore)239 void intel_pasid_tear_down_entry(struct intel_iommu *iommu, struct device *dev,
240 				 u32 pasid, bool fault_ignore)
241 {
242 	struct pasid_entry *pte;
243 	u16 did, pgtt;
244 
245 	spin_lock(&iommu->lock);
246 	pte = intel_pasid_get_entry(dev, pasid);
247 	if (WARN_ON(!pte) || !pasid_pte_is_present(pte)) {
248 		spin_unlock(&iommu->lock);
249 		return;
250 	}
251 
252 	did = pasid_get_domain_id(pte);
253 	pgtt = pasid_pte_get_pgtt(pte);
254 	intel_pasid_clear_entry(dev, pasid, fault_ignore);
255 	spin_unlock(&iommu->lock);
256 
257 	if (!ecap_coherent(iommu->ecap))
258 		clflush_cache_range(pte, sizeof(*pte));
259 
260 	pasid_cache_invalidation_with_pasid(iommu, did, pasid);
261 
262 	if (pgtt == PASID_ENTRY_PGTT_PT || pgtt == PASID_ENTRY_PGTT_FL_ONLY)
263 		qi_flush_piotlb(iommu, did, pasid, 0, -1, 0);
264 	else
265 		iommu->flush.flush_iotlb(iommu, did, 0, 0, DMA_TLB_DSI_FLUSH);
266 
267 	devtlb_invalidation_with_pasid(iommu, dev, pasid);
268 }
269 
270 /*
271  * This function flushes cache for a newly setup pasid table entry.
272  * Caller of it should not modify the in-use pasid table entries.
273  */
pasid_flush_caches(struct intel_iommu * iommu,struct pasid_entry * pte,u32 pasid,u16 did)274 static void pasid_flush_caches(struct intel_iommu *iommu,
275 				struct pasid_entry *pte,
276 			       u32 pasid, u16 did)
277 {
278 	if (!ecap_coherent(iommu->ecap))
279 		clflush_cache_range(pte, sizeof(*pte));
280 
281 	if (cap_caching_mode(iommu->cap)) {
282 		pasid_cache_invalidation_with_pasid(iommu, did, pasid);
283 		qi_flush_piotlb(iommu, did, pasid, 0, -1, 0);
284 	} else {
285 		iommu_flush_write_buffer(iommu);
286 	}
287 }
288 
289 /*
290  * Set up the scalable mode pasid table entry for first only
291  * translation type.
292  */
intel_pasid_setup_first_level(struct intel_iommu * iommu,struct device * dev,pgd_t * pgd,u32 pasid,u16 did,int flags)293 int intel_pasid_setup_first_level(struct intel_iommu *iommu,
294 				  struct device *dev, pgd_t *pgd,
295 				  u32 pasid, u16 did, int flags)
296 {
297 	struct pasid_entry *pte;
298 
299 	if (!ecap_flts(iommu->ecap)) {
300 		pr_err("No first level translation support on %s\n",
301 		       iommu->name);
302 		return -EINVAL;
303 	}
304 
305 	if ((flags & PASID_FLAG_FL5LP) && !cap_fl5lp_support(iommu->cap)) {
306 		pr_err("No 5-level paging support for first-level on %s\n",
307 		       iommu->name);
308 		return -EINVAL;
309 	}
310 
311 	spin_lock(&iommu->lock);
312 	pte = intel_pasid_get_entry(dev, pasid);
313 	if (!pte) {
314 		spin_unlock(&iommu->lock);
315 		return -ENODEV;
316 	}
317 
318 	if (pasid_pte_is_present(pte)) {
319 		spin_unlock(&iommu->lock);
320 		return -EBUSY;
321 	}
322 
323 	pasid_clear_entry(pte);
324 
325 	/* Setup the first level page table pointer: */
326 	pasid_set_flptr(pte, (u64)__pa(pgd));
327 
328 	if (flags & PASID_FLAG_FL5LP)
329 		pasid_set_flpm(pte, 1);
330 
331 	if (flags & PASID_FLAG_PAGE_SNOOP)
332 		pasid_set_pgsnp(pte);
333 
334 	pasid_set_domain_id(pte, did);
335 	pasid_set_address_width(pte, iommu->agaw);
336 	pasid_set_page_snoop(pte, !!ecap_smpwc(iommu->ecap));
337 
338 	/* Setup Present and PASID Granular Transfer Type: */
339 	pasid_set_translation_type(pte, PASID_ENTRY_PGTT_FL_ONLY);
340 	pasid_set_present(pte);
341 	spin_unlock(&iommu->lock);
342 
343 	pasid_flush_caches(iommu, pte, pasid, did);
344 
345 	return 0;
346 }
347 
348 /*
349  * Skip top levels of page tables for iommu which has less agaw
350  * than default. Unnecessary for PT mode.
351  */
iommu_skip_agaw(struct dmar_domain * domain,struct intel_iommu * iommu,struct dma_pte ** pgd)352 static int iommu_skip_agaw(struct dmar_domain *domain,
353 			   struct intel_iommu *iommu,
354 			   struct dma_pte **pgd)
355 {
356 	int agaw;
357 
358 	for (agaw = domain->agaw; agaw > iommu->agaw; agaw--) {
359 		*pgd = phys_to_virt(dma_pte_addr(*pgd));
360 		if (!dma_pte_present(*pgd))
361 			return -EINVAL;
362 	}
363 
364 	return agaw;
365 }
366 
367 /*
368  * Set up the scalable mode pasid entry for second only translation type.
369  */
intel_pasid_setup_second_level(struct intel_iommu * iommu,struct dmar_domain * domain,struct device * dev,u32 pasid)370 int intel_pasid_setup_second_level(struct intel_iommu *iommu,
371 				   struct dmar_domain *domain,
372 				   struct device *dev, u32 pasid)
373 {
374 	struct pasid_entry *pte;
375 	struct dma_pte *pgd;
376 	u64 pgd_val;
377 	int agaw;
378 	u16 did;
379 
380 	/*
381 	 * If hardware advertises no support for second level
382 	 * translation, return directly.
383 	 */
384 	if (!ecap_slts(iommu->ecap)) {
385 		pr_err("No second level translation support on %s\n",
386 		       iommu->name);
387 		return -EINVAL;
388 	}
389 
390 	pgd = domain->pgd;
391 	agaw = iommu_skip_agaw(domain, iommu, &pgd);
392 	if (agaw < 0) {
393 		dev_err(dev, "Invalid domain page table\n");
394 		return -EINVAL;
395 	}
396 
397 	pgd_val = virt_to_phys(pgd);
398 	did = domain_id_iommu(domain, iommu);
399 
400 	spin_lock(&iommu->lock);
401 	pte = intel_pasid_get_entry(dev, pasid);
402 	if (!pte) {
403 		spin_unlock(&iommu->lock);
404 		return -ENODEV;
405 	}
406 
407 	if (pasid_pte_is_present(pte)) {
408 		spin_unlock(&iommu->lock);
409 		return -EBUSY;
410 	}
411 
412 	pasid_clear_entry(pte);
413 	pasid_set_domain_id(pte, did);
414 	pasid_set_slptr(pte, pgd_val);
415 	pasid_set_address_width(pte, agaw);
416 	pasid_set_translation_type(pte, PASID_ENTRY_PGTT_SL_ONLY);
417 	pasid_set_fault_enable(pte);
418 	pasid_set_page_snoop(pte, !!ecap_smpwc(iommu->ecap));
419 	if (domain->dirty_tracking)
420 		pasid_set_ssade(pte);
421 
422 	pasid_set_present(pte);
423 	spin_unlock(&iommu->lock);
424 
425 	pasid_flush_caches(iommu, pte, pasid, did);
426 
427 	return 0;
428 }
429 
430 /*
431  * Set up dirty tracking on a second only or nested translation type.
432  */
intel_pasid_setup_dirty_tracking(struct intel_iommu * iommu,struct device * dev,u32 pasid,bool enabled)433 int intel_pasid_setup_dirty_tracking(struct intel_iommu *iommu,
434 				     struct device *dev, u32 pasid,
435 				     bool enabled)
436 {
437 	struct pasid_entry *pte;
438 	u16 did, pgtt;
439 
440 	spin_lock(&iommu->lock);
441 
442 	pte = intel_pasid_get_entry(dev, pasid);
443 	if (!pte) {
444 		spin_unlock(&iommu->lock);
445 		dev_err_ratelimited(
446 			dev, "Failed to get pasid entry of PASID %d\n", pasid);
447 		return -ENODEV;
448 	}
449 
450 	did = pasid_get_domain_id(pte);
451 	pgtt = pasid_pte_get_pgtt(pte);
452 	if (pgtt != PASID_ENTRY_PGTT_SL_ONLY &&
453 	    pgtt != PASID_ENTRY_PGTT_NESTED) {
454 		spin_unlock(&iommu->lock);
455 		dev_err_ratelimited(
456 			dev,
457 			"Dirty tracking not supported on translation type %d\n",
458 			pgtt);
459 		return -EOPNOTSUPP;
460 	}
461 
462 	if (pasid_get_ssade(pte) == enabled) {
463 		spin_unlock(&iommu->lock);
464 		return 0;
465 	}
466 
467 	if (enabled)
468 		pasid_set_ssade(pte);
469 	else
470 		pasid_clear_ssade(pte);
471 	spin_unlock(&iommu->lock);
472 
473 	if (!ecap_coherent(iommu->ecap))
474 		clflush_cache_range(pte, sizeof(*pte));
475 
476 	/*
477 	 * From VT-d spec table 25 "Guidance to Software for Invalidations":
478 	 *
479 	 * - PASID-selective-within-Domain PASID-cache invalidation
480 	 *   If (PGTT=SS or Nested)
481 	 *    - Domain-selective IOTLB invalidation
482 	 *   Else
483 	 *    - PASID-selective PASID-based IOTLB invalidation
484 	 * - If (pasid is RID_PASID)
485 	 *    - Global Device-TLB invalidation to affected functions
486 	 *   Else
487 	 *    - PASID-based Device-TLB invalidation (with S=1 and
488 	 *      Addr[63:12]=0x7FFFFFFF_FFFFF) to affected functions
489 	 */
490 	pasid_cache_invalidation_with_pasid(iommu, did, pasid);
491 
492 	iommu->flush.flush_iotlb(iommu, did, 0, 0, DMA_TLB_DSI_FLUSH);
493 
494 	devtlb_invalidation_with_pasid(iommu, dev, pasid);
495 
496 	return 0;
497 }
498 
499 /*
500  * Set up the scalable mode pasid entry for passthrough translation type.
501  */
intel_pasid_setup_pass_through(struct intel_iommu * iommu,struct device * dev,u32 pasid)502 int intel_pasid_setup_pass_through(struct intel_iommu *iommu,
503 				   struct device *dev, u32 pasid)
504 {
505 	u16 did = FLPT_DEFAULT_DID;
506 	struct pasid_entry *pte;
507 
508 	spin_lock(&iommu->lock);
509 	pte = intel_pasid_get_entry(dev, pasid);
510 	if (!pte) {
511 		spin_unlock(&iommu->lock);
512 		return -ENODEV;
513 	}
514 
515 	if (pasid_pte_is_present(pte)) {
516 		spin_unlock(&iommu->lock);
517 		return -EBUSY;
518 	}
519 
520 	pasid_clear_entry(pte);
521 	pasid_set_domain_id(pte, did);
522 	pasid_set_address_width(pte, iommu->agaw);
523 	pasid_set_translation_type(pte, PASID_ENTRY_PGTT_PT);
524 	pasid_set_fault_enable(pte);
525 	pasid_set_page_snoop(pte, !!ecap_smpwc(iommu->ecap));
526 	pasid_set_present(pte);
527 	spin_unlock(&iommu->lock);
528 
529 	pasid_flush_caches(iommu, pte, pasid, did);
530 
531 	return 0;
532 }
533 
534 /*
535  * Set the page snoop control for a pasid entry which has been set up.
536  */
intel_pasid_setup_page_snoop_control(struct intel_iommu * iommu,struct device * dev,u32 pasid)537 void intel_pasid_setup_page_snoop_control(struct intel_iommu *iommu,
538 					  struct device *dev, u32 pasid)
539 {
540 	struct pasid_entry *pte;
541 	u16 did;
542 
543 	spin_lock(&iommu->lock);
544 	pte = intel_pasid_get_entry(dev, pasid);
545 	if (WARN_ON(!pte || !pasid_pte_is_present(pte))) {
546 		spin_unlock(&iommu->lock);
547 		return;
548 	}
549 
550 	pasid_set_pgsnp(pte);
551 	did = pasid_get_domain_id(pte);
552 	spin_unlock(&iommu->lock);
553 
554 	if (!ecap_coherent(iommu->ecap))
555 		clflush_cache_range(pte, sizeof(*pte));
556 
557 	/*
558 	 * VT-d spec 3.4 table23 states guides for cache invalidation:
559 	 *
560 	 * - PASID-selective-within-Domain PASID-cache invalidation
561 	 * - PASID-selective PASID-based IOTLB invalidation
562 	 * - If (pasid is RID_PASID)
563 	 *    - Global Device-TLB invalidation to affected functions
564 	 *   Else
565 	 *    - PASID-based Device-TLB invalidation (with S=1 and
566 	 *      Addr[63:12]=0x7FFFFFFF_FFFFF) to affected functions
567 	 */
568 	pasid_cache_invalidation_with_pasid(iommu, did, pasid);
569 	qi_flush_piotlb(iommu, did, pasid, 0, -1, 0);
570 
571 	devtlb_invalidation_with_pasid(iommu, dev, pasid);
572 }
573 
574 /**
575  * intel_pasid_setup_nested() - Set up PASID entry for nested translation.
576  * @iommu:      IOMMU which the device belong to
577  * @dev:        Device to be set up for translation
578  * @pasid:      PASID to be programmed in the device PASID table
579  * @domain:     User stage-1 domain nested on a stage-2 domain
580  *
581  * This is used for nested translation. The input domain should be
582  * nested type and nested on a parent with 'is_nested_parent' flag
583  * set.
584  */
intel_pasid_setup_nested(struct intel_iommu * iommu,struct device * dev,u32 pasid,struct dmar_domain * domain)585 int intel_pasid_setup_nested(struct intel_iommu *iommu, struct device *dev,
586 			     u32 pasid, struct dmar_domain *domain)
587 {
588 	struct iommu_hwpt_vtd_s1 *s1_cfg = &domain->s1_cfg;
589 	pgd_t *s1_gpgd = (pgd_t *)(uintptr_t)domain->s1_pgtbl;
590 	struct dmar_domain *s2_domain = domain->s2_domain;
591 	u16 did = domain_id_iommu(domain, iommu);
592 	struct dma_pte *pgd = s2_domain->pgd;
593 	struct pasid_entry *pte;
594 
595 	/* Address width should match the address width supported by hardware */
596 	switch (s1_cfg->addr_width) {
597 	case ADDR_WIDTH_4LEVEL:
598 		break;
599 	case ADDR_WIDTH_5LEVEL:
600 		if (!cap_fl5lp_support(iommu->cap)) {
601 			dev_err_ratelimited(dev,
602 					    "5-level paging not supported\n");
603 			return -EINVAL;
604 		}
605 		break;
606 	default:
607 		dev_err_ratelimited(dev, "Invalid stage-1 address width %d\n",
608 				    s1_cfg->addr_width);
609 		return -EINVAL;
610 	}
611 
612 	if ((s1_cfg->flags & IOMMU_VTD_S1_SRE) && !ecap_srs(iommu->ecap)) {
613 		pr_err_ratelimited("No supervisor request support on %s\n",
614 				   iommu->name);
615 		return -EINVAL;
616 	}
617 
618 	if ((s1_cfg->flags & IOMMU_VTD_S1_EAFE) && !ecap_eafs(iommu->ecap)) {
619 		pr_err_ratelimited("No extended access flag support on %s\n",
620 				   iommu->name);
621 		return -EINVAL;
622 	}
623 
624 	spin_lock(&iommu->lock);
625 	pte = intel_pasid_get_entry(dev, pasid);
626 	if (!pte) {
627 		spin_unlock(&iommu->lock);
628 		return -ENODEV;
629 	}
630 	if (pasid_pte_is_present(pte)) {
631 		spin_unlock(&iommu->lock);
632 		return -EBUSY;
633 	}
634 
635 	pasid_clear_entry(pte);
636 
637 	if (s1_cfg->addr_width == ADDR_WIDTH_5LEVEL)
638 		pasid_set_flpm(pte, 1);
639 
640 	pasid_set_flptr(pte, (uintptr_t)s1_gpgd);
641 
642 	if (s1_cfg->flags & IOMMU_VTD_S1_SRE) {
643 		pasid_set_sre(pte);
644 		if (s1_cfg->flags & IOMMU_VTD_S1_WPE)
645 			pasid_set_wpe(pte);
646 	}
647 
648 	if (s1_cfg->flags & IOMMU_VTD_S1_EAFE)
649 		pasid_set_eafe(pte);
650 
651 	if (s2_domain->force_snooping)
652 		pasid_set_pgsnp(pte);
653 
654 	pasid_set_slptr(pte, virt_to_phys(pgd));
655 	pasid_set_fault_enable(pte);
656 	pasid_set_domain_id(pte, did);
657 	pasid_set_address_width(pte, s2_domain->agaw);
658 	pasid_set_page_snoop(pte, !!ecap_smpwc(iommu->ecap));
659 	if (s2_domain->dirty_tracking)
660 		pasid_set_ssade(pte);
661 	pasid_set_translation_type(pte, PASID_ENTRY_PGTT_NESTED);
662 	pasid_set_present(pte);
663 	spin_unlock(&iommu->lock);
664 
665 	pasid_flush_caches(iommu, pte, pasid, did);
666 
667 	return 0;
668 }
669 
670 /*
671  * Interfaces to setup or teardown a pasid table to the scalable-mode
672  * context table entry:
673  */
674 
device_pasid_table_teardown(struct device * dev,u8 bus,u8 devfn)675 static void device_pasid_table_teardown(struct device *dev, u8 bus, u8 devfn)
676 {
677 	struct device_domain_info *info = dev_iommu_priv_get(dev);
678 	struct intel_iommu *iommu = info->iommu;
679 	struct context_entry *context;
680 	u16 did;
681 
682 	spin_lock(&iommu->lock);
683 	context = iommu_context_addr(iommu, bus, devfn, false);
684 	if (!context) {
685 		spin_unlock(&iommu->lock);
686 		return;
687 	}
688 
689 	did = context_domain_id(context);
690 	context_clear_entry(context);
691 	__iommu_flush_cache(iommu, context, sizeof(*context));
692 	spin_unlock(&iommu->lock);
693 	intel_context_flush_present(info, context, did, false);
694 }
695 
pci_pasid_table_teardown(struct pci_dev * pdev,u16 alias,void * data)696 static int pci_pasid_table_teardown(struct pci_dev *pdev, u16 alias, void *data)
697 {
698 	struct device *dev = data;
699 
700 	if (dev == &pdev->dev)
701 		device_pasid_table_teardown(dev, PCI_BUS_NUM(alias), alias & 0xff);
702 
703 	return 0;
704 }
705 
intel_pasid_teardown_sm_context(struct device * dev)706 void intel_pasid_teardown_sm_context(struct device *dev)
707 {
708 	struct device_domain_info *info = dev_iommu_priv_get(dev);
709 
710 	if (!dev_is_pci(dev)) {
711 		device_pasid_table_teardown(dev, info->bus, info->devfn);
712 		return;
713 	}
714 
715 	pci_for_each_dma_alias(to_pci_dev(dev), pci_pasid_table_teardown, dev);
716 }
717 
718 /*
719  * Get the PASID directory size for scalable mode context entry.
720  * Value of X in the PDTS field of a scalable mode context entry
721  * indicates PASID directory with 2^(X + 7) entries.
722  */
context_get_sm_pds(struct pasid_table * table)723 static unsigned long context_get_sm_pds(struct pasid_table *table)
724 {
725 	unsigned long pds, max_pde;
726 
727 	max_pde = table->max_pasid >> PASID_PDE_SHIFT;
728 	pds = find_first_bit(&max_pde, MAX_NR_PASID_BITS);
729 	if (pds < 7)
730 		return 0;
731 
732 	return pds - 7;
733 }
734 
context_entry_set_pasid_table(struct context_entry * context,struct device * dev)735 static int context_entry_set_pasid_table(struct context_entry *context,
736 					 struct device *dev)
737 {
738 	struct device_domain_info *info = dev_iommu_priv_get(dev);
739 	struct pasid_table *table = info->pasid_table;
740 	struct intel_iommu *iommu = info->iommu;
741 	unsigned long pds;
742 
743 	context_clear_entry(context);
744 
745 	pds = context_get_sm_pds(table);
746 	context->lo = (u64)virt_to_phys(table->table) | context_pdts(pds);
747 	context_set_sm_rid2pasid(context, IOMMU_NO_PASID);
748 
749 	if (info->ats_supported)
750 		context_set_sm_dte(context);
751 	if (info->pasid_supported)
752 		context_set_pasid(context);
753 
754 	context_set_fault_enable(context);
755 	context_set_present(context);
756 	__iommu_flush_cache(iommu, context, sizeof(*context));
757 
758 	return 0;
759 }
760 
device_pasid_table_setup(struct device * dev,u8 bus,u8 devfn)761 static int device_pasid_table_setup(struct device *dev, u8 bus, u8 devfn)
762 {
763 	struct device_domain_info *info = dev_iommu_priv_get(dev);
764 	struct intel_iommu *iommu = info->iommu;
765 	struct context_entry *context;
766 
767 	spin_lock(&iommu->lock);
768 	context = iommu_context_addr(iommu, bus, devfn, true);
769 	if (!context) {
770 		spin_unlock(&iommu->lock);
771 		return -ENOMEM;
772 	}
773 
774 	if (context_present(context) && !context_copied(iommu, bus, devfn)) {
775 		spin_unlock(&iommu->lock);
776 		return 0;
777 	}
778 
779 	if (context_copied(iommu, bus, devfn)) {
780 		context_clear_entry(context);
781 		__iommu_flush_cache(iommu, context, sizeof(*context));
782 
783 		/*
784 		 * For kdump cases, old valid entries may be cached due to
785 		 * the in-flight DMA and copied pgtable, but there is no
786 		 * unmapping behaviour for them, thus we need explicit cache
787 		 * flushes for all affected domain IDs and PASIDs used in
788 		 * the copied PASID table. Given that we have no idea about
789 		 * which domain IDs and PASIDs were used in the copied tables,
790 		 * upgrade them to global PASID and IOTLB cache invalidation.
791 		 */
792 		iommu->flush.flush_context(iommu, 0,
793 					   PCI_DEVID(bus, devfn),
794 					   DMA_CCMD_MASK_NOBIT,
795 					   DMA_CCMD_DEVICE_INVL);
796 		qi_flush_pasid_cache(iommu, 0, QI_PC_GLOBAL, 0);
797 		iommu->flush.flush_iotlb(iommu, 0, 0, 0, DMA_TLB_GLOBAL_FLUSH);
798 		devtlb_invalidation_with_pasid(iommu, dev, IOMMU_NO_PASID);
799 
800 		/*
801 		 * At this point, the device is supposed to finish reset at
802 		 * its driver probe stage, so no in-flight DMA will exist,
803 		 * and we don't need to worry anymore hereafter.
804 		 */
805 		clear_context_copied(iommu, bus, devfn);
806 	}
807 
808 	context_entry_set_pasid_table(context, dev);
809 	spin_unlock(&iommu->lock);
810 
811 	/*
812 	 * It's a non-present to present mapping. If hardware doesn't cache
813 	 * non-present entry we don't need to flush the caches. If it does
814 	 * cache non-present entries, then it does so in the special
815 	 * domain #0, which we have to flush:
816 	 */
817 	if (cap_caching_mode(iommu->cap)) {
818 		iommu->flush.flush_context(iommu, 0,
819 					   PCI_DEVID(bus, devfn),
820 					   DMA_CCMD_MASK_NOBIT,
821 					   DMA_CCMD_DEVICE_INVL);
822 		iommu->flush.flush_iotlb(iommu, 0, 0, 0, DMA_TLB_DSI_FLUSH);
823 	}
824 
825 	return 0;
826 }
827 
pci_pasid_table_setup(struct pci_dev * pdev,u16 alias,void * data)828 static int pci_pasid_table_setup(struct pci_dev *pdev, u16 alias, void *data)
829 {
830 	struct device *dev = data;
831 
832 	if (dev != &pdev->dev)
833 		return 0;
834 
835 	return device_pasid_table_setup(dev, PCI_BUS_NUM(alias), alias & 0xff);
836 }
837 
838 /*
839  * Set the device's PASID table to its context table entry.
840  *
841  * The PASID table is set to the context entries of both device itself
842  * and its alias requester ID for DMA.
843  */
intel_pasid_setup_sm_context(struct device * dev)844 int intel_pasid_setup_sm_context(struct device *dev)
845 {
846 	struct device_domain_info *info = dev_iommu_priv_get(dev);
847 
848 	if (!dev_is_pci(dev))
849 		return device_pasid_table_setup(dev, info->bus, info->devfn);
850 
851 	return pci_for_each_dma_alias(to_pci_dev(dev), pci_pasid_table_setup, dev);
852 }
853 
854 /*
855  * Global Device-TLB invalidation following changes in a context entry which
856  * was present.
857  */
__context_flush_dev_iotlb(struct device_domain_info * info)858 static void __context_flush_dev_iotlb(struct device_domain_info *info)
859 {
860 	if (!info->ats_enabled)
861 		return;
862 
863 	qi_flush_dev_iotlb(info->iommu, PCI_DEVID(info->bus, info->devfn),
864 			   info->pfsid, info->ats_qdep, 0, MAX_AGAW_PFN_WIDTH);
865 
866 	/*
867 	 * There is no guarantee that the device DMA is stopped when it reaches
868 	 * here. Therefore, always attempt the extra device TLB invalidation
869 	 * quirk. The impact on performance is acceptable since this is not a
870 	 * performance-critical path.
871 	 */
872 	quirk_extra_dev_tlb_flush(info, 0, MAX_AGAW_PFN_WIDTH, IOMMU_NO_PASID,
873 				  info->ats_qdep);
874 }
875 
876 /*
877  * Cache invalidations after change in a context table entry that was present
878  * according to the Spec 6.5.3.3 (Guidance to Software for Invalidations). If
879  * IOMMU is in scalable mode and all PASID table entries of the device were
880  * non-present, set flush_domains to false. Otherwise, true.
881  */
intel_context_flush_present(struct device_domain_info * info,struct context_entry * context,u16 did,bool flush_domains)882 void intel_context_flush_present(struct device_domain_info *info,
883 				 struct context_entry *context,
884 				 u16 did, bool flush_domains)
885 {
886 	struct intel_iommu *iommu = info->iommu;
887 	struct pasid_entry *pte;
888 	int i;
889 
890 	/*
891 	 * Device-selective context-cache invalidation. The Domain-ID field
892 	 * of the Context-cache Invalidate Descriptor is ignored by hardware
893 	 * when operating in scalable mode. Therefore the @did value doesn't
894 	 * matter in scalable mode.
895 	 */
896 	iommu->flush.flush_context(iommu, did, PCI_DEVID(info->bus, info->devfn),
897 				   DMA_CCMD_MASK_NOBIT, DMA_CCMD_DEVICE_INVL);
898 
899 	/*
900 	 * For legacy mode:
901 	 * - Domain-selective IOTLB invalidation
902 	 * - Global Device-TLB invalidation to all affected functions
903 	 */
904 	if (!sm_supported(iommu)) {
905 		iommu->flush.flush_iotlb(iommu, did, 0, 0, DMA_TLB_DSI_FLUSH);
906 		__context_flush_dev_iotlb(info);
907 
908 		return;
909 	}
910 
911 	/*
912 	 * For scalable mode:
913 	 * - Domain-selective PASID-cache invalidation to affected domains
914 	 * - Domain-selective IOTLB invalidation to affected domains
915 	 * - Global Device-TLB invalidation to affected functions
916 	 */
917 	if (flush_domains) {
918 		/*
919 		 * If the IOMMU is running in scalable mode and there might
920 		 * be potential PASID translations, the caller should hold
921 		 * the lock to ensure that context changes and cache flushes
922 		 * are atomic.
923 		 */
924 		assert_spin_locked(&iommu->lock);
925 		for (i = 0; i < info->pasid_table->max_pasid; i++) {
926 			pte = intel_pasid_get_entry(info->dev, i);
927 			if (!pte || !pasid_pte_is_present(pte))
928 				continue;
929 
930 			did = pasid_get_domain_id(pte);
931 			qi_flush_pasid_cache(iommu, did, QI_PC_ALL_PASIDS, 0);
932 			iommu->flush.flush_iotlb(iommu, did, 0, 0, DMA_TLB_DSI_FLUSH);
933 		}
934 	}
935 
936 	__context_flush_dev_iotlb(info);
937 }
938