1 // SPDX-License-Identifier: GPL-2.0
3 * intel-pasid.c - PASID idr, table and entry manipulation
5 * Copyright (C) 2018 Intel Corporation
10 #define pr_fmt(fmt) "DMAR: " fmt
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>
23 #include "../iommu-pages.h"
26 * Intel IOMMU system wide PASID name space:
28 u32 intel_pasid_max_id = PASID_MAX;
31 * Per device pasid table management:
35 * Allocate a pasid table for @dev. It should be called in a
36 * single-thread context.
38 int intel_pasid_alloc_table(struct device *dev)
40 struct device_domain_info *info;
41 struct pasid_table *pasid_table;
42 struct pasid_dir_entry *dir;
47 info = dev_iommu_priv_get(dev);
48 if (WARN_ON(!info || !dev_is_pci(dev)))
50 if (WARN_ON(info->pasid_table))
53 pasid_table = kzalloc(sizeof(*pasid_table), GFP_KERNEL);
57 if (info->pasid_supported)
58 max_pasid = min_t(u32, pci_max_pasids(to_pci_dev(dev)),
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);
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;
74 if (!ecap_coherent(info->iommu->ecap))
75 clflush_cache_range(pasid_table->table, (1 << order) * PAGE_SIZE);
80 void intel_pasid_free_table(struct device *dev)
82 struct device_domain_info *info;
83 struct pasid_table *pasid_table;
84 struct pasid_dir_entry *dir;
85 struct pasid_entry *table;
88 info = dev_iommu_priv_get(dev);
89 if (!info || !dev_is_pci(dev) || !info->pasid_table)
92 pasid_table = info->pasid_table;
93 info->pasid_table = NULL;
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);
103 iommu_free_pages(pasid_table->table, pasid_table->order);
107 struct pasid_table *intel_pasid_get_table(struct device *dev)
109 struct device_domain_info *info;
111 info = dev_iommu_priv_get(dev);
115 return info->pasid_table;
118 static int intel_pasid_get_dev_max_id(struct device *dev)
120 struct device_domain_info *info;
122 info = dev_iommu_priv_get(dev);
123 if (!info || !info->pasid_table)
126 return info->pasid_table->max_pasid;
129 static struct pasid_entry *intel_pasid_get_entry(struct device *dev, u32 pasid)
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;
137 pasid_table = intel_pasid_get_table(dev);
138 if (WARN_ON(!pasid_table || pasid >= intel_pasid_get_dev_max_id(dev)))
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;
147 entries = get_pasid_table_from_pde(&dir[dir_index]);
151 entries = iommu_alloc_page_node(info->iommu->node, GFP_ATOMIC);
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.
162 if (!try_cmpxchg64(&dir[dir_index].val, &tmp,
163 (u64)virt_to_phys(entries) | PASID_PTE_PRESENT)) {
164 iommu_free_page(entries);
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));
173 return &entries[index];
177 * Interfaces for PASID table entry manipulation:
180 intel_pasid_clear_entry(struct device *dev, u32 pasid, bool fault_ignore)
182 struct pasid_entry *pe;
184 pe = intel_pasid_get_entry(dev, pasid);
188 if (fault_ignore && pasid_pte_is_present(pe))
189 pasid_clear_entry_with_fpd(pe);
191 pasid_clear_entry(pe);
195 pasid_cache_invalidation_with_pasid(struct intel_iommu *iommu,
200 desc.qw0 = QI_PC_DID(did) | QI_PC_GRAN(QI_PC_PASID_SEL) |
201 QI_PC_PASID(pasid) | QI_PC_TYPE;
206 qi_submit_sync(iommu, &desc, 1, 0);
210 devtlb_invalidation_with_pasid(struct intel_iommu *iommu,
211 struct device *dev, u32 pasid)
213 struct device_domain_info *info;
214 u16 sid, qdep, pfsid;
216 info = dev_iommu_priv_get(dev);
217 if (!info || !info->ats_enabled)
220 if (pci_dev_is_disconnected(to_pci_dev(dev)))
223 sid = info->bus << 8 | info->devfn;
224 qdep = info->ats_qdep;
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.
233 if (pasid == IOMMU_NO_PASID)
234 qi_flush_dev_iotlb(iommu, sid, pfsid, qdep, 0, 64 - VTD_PAGE_SHIFT);
236 qi_flush_dev_iotlb_pasid(iommu, sid, pfsid, pasid, qdep, 0, 64 - VTD_PAGE_SHIFT);
239 void intel_pasid_tear_down_entry(struct intel_iommu *iommu, struct device *dev,
240 u32 pasid, bool fault_ignore)
242 struct pasid_entry *pte;
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);
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);
257 if (!ecap_coherent(iommu->ecap))
258 clflush_cache_range(pte, sizeof(*pte));
260 pasid_cache_invalidation_with_pasid(iommu, did, pasid);
262 if (pgtt == PASID_ENTRY_PGTT_PT || pgtt == PASID_ENTRY_PGTT_FL_ONLY)
263 qi_flush_piotlb(iommu, did, pasid, 0, -1, 0);
265 iommu->flush.flush_iotlb(iommu, did, 0, 0, DMA_TLB_DSI_FLUSH);
267 /* Device IOTLB doesn't need to be flushed in caching mode. */
268 if (!cap_caching_mode(iommu->cap))
269 devtlb_invalidation_with_pasid(iommu, dev, pasid);
273 * This function flushes cache for a newly setup pasid table entry.
274 * Caller of it should not modify the in-use pasid table entries.
276 static void pasid_flush_caches(struct intel_iommu *iommu,
277 struct pasid_entry *pte,
280 if (!ecap_coherent(iommu->ecap))
281 clflush_cache_range(pte, sizeof(*pte));
283 if (cap_caching_mode(iommu->cap)) {
284 pasid_cache_invalidation_with_pasid(iommu, did, pasid);
285 qi_flush_piotlb(iommu, did, pasid, 0, -1, 0);
287 iommu_flush_write_buffer(iommu);
292 * Set up the scalable mode pasid table entry for first only
295 int intel_pasid_setup_first_level(struct intel_iommu *iommu,
296 struct device *dev, pgd_t *pgd,
297 u32 pasid, u16 did, int flags)
299 struct pasid_entry *pte;
301 if (!ecap_flts(iommu->ecap)) {
302 pr_err("No first level translation support on %s\n",
307 if ((flags & PASID_FLAG_FL5LP) && !cap_fl5lp_support(iommu->cap)) {
308 pr_err("No 5-level paging support for first-level on %s\n",
313 spin_lock(&iommu->lock);
314 pte = intel_pasid_get_entry(dev, pasid);
316 spin_unlock(&iommu->lock);
320 if (pasid_pte_is_present(pte)) {
321 spin_unlock(&iommu->lock);
325 pasid_clear_entry(pte);
327 /* Setup the first level page table pointer: */
328 pasid_set_flptr(pte, (u64)__pa(pgd));
330 if (flags & PASID_FLAG_FL5LP)
331 pasid_set_flpm(pte, 1);
333 if (flags & PASID_FLAG_PAGE_SNOOP)
334 pasid_set_pgsnp(pte);
336 pasid_set_domain_id(pte, did);
337 pasid_set_address_width(pte, iommu->agaw);
338 pasid_set_page_snoop(pte, !!ecap_smpwc(iommu->ecap));
340 /* Setup Present and PASID Granular Transfer Type: */
341 pasid_set_translation_type(pte, PASID_ENTRY_PGTT_FL_ONLY);
342 pasid_set_present(pte);
343 spin_unlock(&iommu->lock);
345 pasid_flush_caches(iommu, pte, pasid, did);
351 * Skip top levels of page tables for iommu which has less agaw
352 * than default. Unnecessary for PT mode.
354 static int iommu_skip_agaw(struct dmar_domain *domain,
355 struct intel_iommu *iommu,
356 struct dma_pte **pgd)
360 for (agaw = domain->agaw; agaw > iommu->agaw; agaw--) {
361 *pgd = phys_to_virt(dma_pte_addr(*pgd));
362 if (!dma_pte_present(*pgd))
370 * Set up the scalable mode pasid entry for second only translation type.
372 int intel_pasid_setup_second_level(struct intel_iommu *iommu,
373 struct dmar_domain *domain,
374 struct device *dev, u32 pasid)
376 struct pasid_entry *pte;
383 * If hardware advertises no support for second level
384 * translation, return directly.
386 if (!ecap_slts(iommu->ecap)) {
387 pr_err("No second level translation support on %s\n",
393 agaw = iommu_skip_agaw(domain, iommu, &pgd);
395 dev_err(dev, "Invalid domain page table\n");
399 pgd_val = virt_to_phys(pgd);
400 did = domain_id_iommu(domain, iommu);
402 spin_lock(&iommu->lock);
403 pte = intel_pasid_get_entry(dev, pasid);
405 spin_unlock(&iommu->lock);
409 if (pasid_pte_is_present(pte)) {
410 spin_unlock(&iommu->lock);
414 pasid_clear_entry(pte);
415 pasid_set_domain_id(pte, did);
416 pasid_set_slptr(pte, pgd_val);
417 pasid_set_address_width(pte, agaw);
418 pasid_set_translation_type(pte, PASID_ENTRY_PGTT_SL_ONLY);
419 pasid_set_fault_enable(pte);
420 pasid_set_page_snoop(pte, !!ecap_smpwc(iommu->ecap));
421 if (domain->dirty_tracking)
422 pasid_set_ssade(pte);
424 pasid_set_present(pte);
425 spin_unlock(&iommu->lock);
427 pasid_flush_caches(iommu, pte, pasid, did);
433 * Set up dirty tracking on a second only or nested translation type.
435 int intel_pasid_setup_dirty_tracking(struct intel_iommu *iommu,
436 struct device *dev, u32 pasid,
439 struct pasid_entry *pte;
442 spin_lock(&iommu->lock);
444 pte = intel_pasid_get_entry(dev, pasid);
446 spin_unlock(&iommu->lock);
448 dev, "Failed to get pasid entry of PASID %d\n", pasid);
452 did = pasid_get_domain_id(pte);
453 pgtt = pasid_pte_get_pgtt(pte);
454 if (pgtt != PASID_ENTRY_PGTT_SL_ONLY &&
455 pgtt != PASID_ENTRY_PGTT_NESTED) {
456 spin_unlock(&iommu->lock);
459 "Dirty tracking not supported on translation type %d\n",
464 if (pasid_get_ssade(pte) == enabled) {
465 spin_unlock(&iommu->lock);
470 pasid_set_ssade(pte);
472 pasid_clear_ssade(pte);
473 spin_unlock(&iommu->lock);
475 if (!ecap_coherent(iommu->ecap))
476 clflush_cache_range(pte, sizeof(*pte));
479 * From VT-d spec table 25 "Guidance to Software for Invalidations":
481 * - PASID-selective-within-Domain PASID-cache invalidation
482 * If (PGTT=SS or Nested)
483 * - Domain-selective IOTLB invalidation
485 * - PASID-selective PASID-based IOTLB invalidation
486 * - If (pasid is RID_PASID)
487 * - Global Device-TLB invalidation to affected functions
489 * - PASID-based Device-TLB invalidation (with S=1 and
490 * Addr[63:12]=0x7FFFFFFF_FFFFF) to affected functions
492 pasid_cache_invalidation_with_pasid(iommu, did, pasid);
494 iommu->flush.flush_iotlb(iommu, did, 0, 0, DMA_TLB_DSI_FLUSH);
496 /* Device IOTLB doesn't need to be flushed in caching mode. */
497 if (!cap_caching_mode(iommu->cap))
498 devtlb_invalidation_with_pasid(iommu, dev, pasid);
504 * Set up the scalable mode pasid entry for passthrough translation type.
506 int intel_pasid_setup_pass_through(struct intel_iommu *iommu,
507 struct device *dev, u32 pasid)
509 u16 did = FLPT_DEFAULT_DID;
510 struct pasid_entry *pte;
512 spin_lock(&iommu->lock);
513 pte = intel_pasid_get_entry(dev, pasid);
515 spin_unlock(&iommu->lock);
519 if (pasid_pte_is_present(pte)) {
520 spin_unlock(&iommu->lock);
524 pasid_clear_entry(pte);
525 pasid_set_domain_id(pte, did);
526 pasid_set_address_width(pte, iommu->agaw);
527 pasid_set_translation_type(pte, PASID_ENTRY_PGTT_PT);
528 pasid_set_fault_enable(pte);
529 pasid_set_page_snoop(pte, !!ecap_smpwc(iommu->ecap));
530 pasid_set_present(pte);
531 spin_unlock(&iommu->lock);
533 pasid_flush_caches(iommu, pte, pasid, did);
539 * Set the page snoop control for a pasid entry which has been set up.
541 void intel_pasid_setup_page_snoop_control(struct intel_iommu *iommu,
542 struct device *dev, u32 pasid)
544 struct pasid_entry *pte;
547 spin_lock(&iommu->lock);
548 pte = intel_pasid_get_entry(dev, pasid);
549 if (WARN_ON(!pte || !pasid_pte_is_present(pte))) {
550 spin_unlock(&iommu->lock);
554 pasid_set_pgsnp(pte);
555 did = pasid_get_domain_id(pte);
556 spin_unlock(&iommu->lock);
558 if (!ecap_coherent(iommu->ecap))
559 clflush_cache_range(pte, sizeof(*pte));
562 * VT-d spec 3.4 table23 states guides for cache invalidation:
564 * - PASID-selective-within-Domain PASID-cache invalidation
565 * - PASID-selective PASID-based IOTLB invalidation
566 * - If (pasid is RID_PASID)
567 * - Global Device-TLB invalidation to affected functions
569 * - PASID-based Device-TLB invalidation (with S=1 and
570 * Addr[63:12]=0x7FFFFFFF_FFFFF) to affected functions
572 pasid_cache_invalidation_with_pasid(iommu, did, pasid);
573 qi_flush_piotlb(iommu, did, pasid, 0, -1, 0);
575 /* Device IOTLB doesn't need to be flushed in caching mode. */
576 if (!cap_caching_mode(iommu->cap))
577 devtlb_invalidation_with_pasid(iommu, dev, pasid);
581 * intel_pasid_setup_nested() - Set up PASID entry for nested translation.
582 * @iommu: IOMMU which the device belong to
583 * @dev: Device to be set up for translation
584 * @pasid: PASID to be programmed in the device PASID table
585 * @domain: User stage-1 domain nested on a stage-2 domain
587 * This is used for nested translation. The input domain should be
588 * nested type and nested on a parent with 'is_nested_parent' flag
591 int intel_pasid_setup_nested(struct intel_iommu *iommu, struct device *dev,
592 u32 pasid, struct dmar_domain *domain)
594 struct iommu_hwpt_vtd_s1 *s1_cfg = &domain->s1_cfg;
595 pgd_t *s1_gpgd = (pgd_t *)(uintptr_t)domain->s1_pgtbl;
596 struct dmar_domain *s2_domain = domain->s2_domain;
597 u16 did = domain_id_iommu(domain, iommu);
598 struct dma_pte *pgd = s2_domain->pgd;
599 struct pasid_entry *pte;
601 /* Address width should match the address width supported by hardware */
602 switch (s1_cfg->addr_width) {
603 case ADDR_WIDTH_4LEVEL:
605 case ADDR_WIDTH_5LEVEL:
606 if (!cap_fl5lp_support(iommu->cap)) {
607 dev_err_ratelimited(dev,
608 "5-level paging not supported\n");
613 dev_err_ratelimited(dev, "Invalid stage-1 address width %d\n",
618 if ((s1_cfg->flags & IOMMU_VTD_S1_SRE) && !ecap_srs(iommu->ecap)) {
619 pr_err_ratelimited("No supervisor request support on %s\n",
624 if ((s1_cfg->flags & IOMMU_VTD_S1_EAFE) && !ecap_eafs(iommu->ecap)) {
625 pr_err_ratelimited("No extended access flag support on %s\n",
630 spin_lock(&iommu->lock);
631 pte = intel_pasid_get_entry(dev, pasid);
633 spin_unlock(&iommu->lock);
636 if (pasid_pte_is_present(pte)) {
637 spin_unlock(&iommu->lock);
641 pasid_clear_entry(pte);
643 if (s1_cfg->addr_width == ADDR_WIDTH_5LEVEL)
644 pasid_set_flpm(pte, 1);
646 pasid_set_flptr(pte, (uintptr_t)s1_gpgd);
648 if (s1_cfg->flags & IOMMU_VTD_S1_SRE) {
650 if (s1_cfg->flags & IOMMU_VTD_S1_WPE)
654 if (s1_cfg->flags & IOMMU_VTD_S1_EAFE)
657 if (s2_domain->force_snooping)
658 pasid_set_pgsnp(pte);
660 pasid_set_slptr(pte, virt_to_phys(pgd));
661 pasid_set_fault_enable(pte);
662 pasid_set_domain_id(pte, did);
663 pasid_set_address_width(pte, s2_domain->agaw);
664 pasid_set_page_snoop(pte, !!ecap_smpwc(iommu->ecap));
665 if (s2_domain->dirty_tracking)
666 pasid_set_ssade(pte);
667 pasid_set_translation_type(pte, PASID_ENTRY_PGTT_NESTED);
668 pasid_set_present(pte);
669 spin_unlock(&iommu->lock);
671 pasid_flush_caches(iommu, pte, pasid, did);
677 * Interfaces to setup or teardown a pasid table to the scalable-mode
678 * context table entry:
681 static void device_pasid_table_teardown(struct device *dev, u8 bus, u8 devfn)
683 struct device_domain_info *info = dev_iommu_priv_get(dev);
684 struct intel_iommu *iommu = info->iommu;
685 struct context_entry *context;
687 spin_lock(&iommu->lock);
688 context = iommu_context_addr(iommu, bus, devfn, false);
690 spin_unlock(&iommu->lock);
694 context_clear_entry(context);
695 __iommu_flush_cache(iommu, context, sizeof(*context));
696 spin_unlock(&iommu->lock);
697 intel_context_flush_present(info, context, false);
700 static int pci_pasid_table_teardown(struct pci_dev *pdev, u16 alias, void *data)
702 struct device *dev = data;
704 if (dev == &pdev->dev)
705 device_pasid_table_teardown(dev, PCI_BUS_NUM(alias), alias & 0xff);
710 void intel_pasid_teardown_sm_context(struct device *dev)
712 struct device_domain_info *info = dev_iommu_priv_get(dev);
714 if (!dev_is_pci(dev)) {
715 device_pasid_table_teardown(dev, info->bus, info->devfn);
719 pci_for_each_dma_alias(to_pci_dev(dev), pci_pasid_table_teardown, dev);
723 * Get the PASID directory size for scalable mode context entry.
724 * Value of X in the PDTS field of a scalable mode context entry
725 * indicates PASID directory with 2^(X + 7) entries.
727 static unsigned long context_get_sm_pds(struct pasid_table *table)
729 unsigned long pds, max_pde;
731 max_pde = table->max_pasid >> PASID_PDE_SHIFT;
732 pds = find_first_bit(&max_pde, MAX_NR_PASID_BITS);
739 static int context_entry_set_pasid_table(struct context_entry *context,
742 struct device_domain_info *info = dev_iommu_priv_get(dev);
743 struct pasid_table *table = info->pasid_table;
744 struct intel_iommu *iommu = info->iommu;
747 context_clear_entry(context);
749 pds = context_get_sm_pds(table);
750 context->lo = (u64)virt_to_phys(table->table) | context_pdts(pds);
751 context_set_sm_rid2pasid(context, IOMMU_NO_PASID);
753 if (info->ats_supported)
754 context_set_sm_dte(context);
755 if (info->pasid_supported)
756 context_set_pasid(context);
758 context_set_fault_enable(context);
759 context_set_present(context);
760 __iommu_flush_cache(iommu, context, sizeof(*context));
765 static int device_pasid_table_setup(struct device *dev, u8 bus, u8 devfn)
767 struct device_domain_info *info = dev_iommu_priv_get(dev);
768 struct intel_iommu *iommu = info->iommu;
769 struct context_entry *context;
771 spin_lock(&iommu->lock);
772 context = iommu_context_addr(iommu, bus, devfn, true);
774 spin_unlock(&iommu->lock);
778 if (context_present(context) && !context_copied(iommu, bus, devfn)) {
779 spin_unlock(&iommu->lock);
783 if (context_copied(iommu, bus, devfn)) {
784 context_clear_entry(context);
785 __iommu_flush_cache(iommu, context, sizeof(*context));
788 * For kdump cases, old valid entries may be cached due to
789 * the in-flight DMA and copied pgtable, but there is no
790 * unmapping behaviour for them, thus we need explicit cache
791 * flushes for all affected domain IDs and PASIDs used in
792 * the copied PASID table. Given that we have no idea about
793 * which domain IDs and PASIDs were used in the copied tables,
794 * upgrade them to global PASID and IOTLB cache invalidation.
796 iommu->flush.flush_context(iommu, 0,
797 PCI_DEVID(bus, devfn),
799 DMA_CCMD_DEVICE_INVL);
800 qi_flush_pasid_cache(iommu, 0, QI_PC_GLOBAL, 0);
801 iommu->flush.flush_iotlb(iommu, 0, 0, 0, DMA_TLB_GLOBAL_FLUSH);
802 devtlb_invalidation_with_pasid(iommu, dev, IOMMU_NO_PASID);
805 * At this point, the device is supposed to finish reset at
806 * its driver probe stage, so no in-flight DMA will exist,
807 * and we don't need to worry anymore hereafter.
809 clear_context_copied(iommu, bus, devfn);
812 context_entry_set_pasid_table(context, dev);
813 spin_unlock(&iommu->lock);
816 * It's a non-present to present mapping. If hardware doesn't cache
817 * non-present entry we don't need to flush the caches. If it does
818 * cache non-present entries, then it does so in the special
819 * domain #0, which we have to flush:
821 if (cap_caching_mode(iommu->cap)) {
822 iommu->flush.flush_context(iommu, 0,
823 PCI_DEVID(bus, devfn),
825 DMA_CCMD_DEVICE_INVL);
826 iommu->flush.flush_iotlb(iommu, 0, 0, 0, DMA_TLB_DSI_FLUSH);
832 static int pci_pasid_table_setup(struct pci_dev *pdev, u16 alias, void *data)
834 struct device *dev = data;
836 if (dev != &pdev->dev)
839 return device_pasid_table_setup(dev, PCI_BUS_NUM(alias), alias & 0xff);
843 * Set the device's PASID table to its context table entry.
845 * The PASID table is set to the context entries of both device itself
846 * and its alias requester ID for DMA.
848 int intel_pasid_setup_sm_context(struct device *dev)
850 struct device_domain_info *info = dev_iommu_priv_get(dev);
852 if (!dev_is_pci(dev))
853 return device_pasid_table_setup(dev, info->bus, info->devfn);
855 return pci_for_each_dma_alias(to_pci_dev(dev), pci_pasid_table_setup, dev);
859 * Global Device-TLB invalidation following changes in a context entry which
862 static void __context_flush_dev_iotlb(struct device_domain_info *info)
864 if (!info->ats_enabled)
867 qi_flush_dev_iotlb(info->iommu, PCI_DEVID(info->bus, info->devfn),
868 info->pfsid, info->ats_qdep, 0, MAX_AGAW_PFN_WIDTH);
871 * There is no guarantee that the device DMA is stopped when it reaches
872 * here. Therefore, always attempt the extra device TLB invalidation
873 * quirk. The impact on performance is acceptable since this is not a
874 * performance-critical path.
876 quirk_extra_dev_tlb_flush(info, 0, MAX_AGAW_PFN_WIDTH, IOMMU_NO_PASID,
881 * Cache invalidations after change in a context table entry that was present
882 * according to the Spec 6.5.3.3 (Guidance to Software for Invalidations). If
883 * IOMMU is in scalable mode and all PASID table entries of the device were
884 * non-present, set flush_domains to false. Otherwise, true.
886 void intel_context_flush_present(struct device_domain_info *info,
887 struct context_entry *context,
890 struct intel_iommu *iommu = info->iommu;
891 u16 did = context_domain_id(context);
892 struct pasid_entry *pte;
896 * Device-selective context-cache invalidation. The Domain-ID field
897 * of the Context-cache Invalidate Descriptor is ignored by hardware
898 * when operating in scalable mode. Therefore the @did value doesn't
899 * matter in scalable mode.
901 iommu->flush.flush_context(iommu, did, PCI_DEVID(info->bus, info->devfn),
902 DMA_CCMD_MASK_NOBIT, DMA_CCMD_DEVICE_INVL);
906 * - Domain-selective IOTLB invalidation
907 * - Global Device-TLB invalidation to all affected functions
909 if (!sm_supported(iommu)) {
910 iommu->flush.flush_iotlb(iommu, did, 0, 0, DMA_TLB_DSI_FLUSH);
911 __context_flush_dev_iotlb(info);
918 * - Domain-selective PASID-cache invalidation to affected domains
919 * - Domain-selective IOTLB invalidation to affected domains
920 * - Global Device-TLB invalidation to affected functions
924 * If the IOMMU is running in scalable mode and there might
925 * be potential PASID translations, the caller should hold
926 * the lock to ensure that context changes and cache flushes
929 assert_spin_locked(&iommu->lock);
930 for (i = 0; i < info->pasid_table->max_pasid; i++) {
931 pte = intel_pasid_get_entry(info->dev, i);
932 if (!pte || !pasid_pte_is_present(pte))
935 did = pasid_get_domain_id(pte);
936 qi_flush_pasid_cache(iommu, did, QI_PC_ALL_PASIDS, 0);
937 iommu->flush.flush_iotlb(iommu, did, 0, 0, DMA_TLB_DSI_FLUSH);
941 __context_flush_dev_iotlb(info);