1 // SPDX-License-Identifier: GPL-2.0-only
3 * A fairly generic DMA-API to IOMMU-API glue layer.
5 * Copyright (C) 2014-2015 ARM Ltd.
7 * based in part on arch/arm/mm/dma-mapping.c:
8 * Copyright (C) 2000-2004 Russell King
11 #include <linux/acpi_iort.h>
12 #include <linux/atomic.h>
13 #include <linux/crash_dump.h>
14 #include <linux/device.h>
15 #include <linux/dma-direct.h>
16 #include <linux/dma-map-ops.h>
17 #include <linux/gfp.h>
18 #include <linux/huge_mm.h>
19 #include <linux/iommu.h>
20 #include <linux/iova.h>
21 #include <linux/irq.h>
22 #include <linux/list_sort.h>
23 #include <linux/memremap.h>
25 #include <linux/mutex.h>
26 #include <linux/pci.h>
27 #include <linux/scatterlist.h>
28 #include <linux/spinlock.h>
29 #include <linux/swiotlb.h>
30 #include <linux/vmalloc.h>
32 #include "dma-iommu.h"
34 struct iommu_dma_msi_page {
35 struct list_head list;
40 enum iommu_dma_cookie_type {
41 IOMMU_DMA_IOVA_COOKIE,
45 struct iommu_dma_cookie {
46 enum iommu_dma_cookie_type type;
48 /* Full allocator for IOMMU_DMA_IOVA_COOKIE */
50 struct iova_domain iovad;
52 struct iova_fq __percpu *fq; /* Flush queue */
53 /* Number of TLB flushes that have been started */
54 atomic64_t fq_flush_start_cnt;
55 /* Number of TLB flushes that have been finished */
56 atomic64_t fq_flush_finish_cnt;
57 /* Timer to regularily empty the flush queues */
58 struct timer_list fq_timer;
59 /* 1 when timer is active, 0 when not */
62 /* Trivial linear page allocator for IOMMU_DMA_MSI_COOKIE */
65 struct list_head msi_page_list;
67 /* Domain for flush queue callback; NULL if flush queue not in use */
68 struct iommu_domain *fq_domain;
72 static DEFINE_STATIC_KEY_FALSE(iommu_deferred_attach_enabled);
73 bool iommu_dma_forcedac __read_mostly;
75 static int __init iommu_dma_forcedac_setup(char *str)
77 int ret = kstrtobool(str, &iommu_dma_forcedac);
79 if (!ret && iommu_dma_forcedac)
80 pr_info("Forcing DAC for PCI devices\n");
83 early_param("iommu.forcedac", iommu_dma_forcedac_setup);
85 /* Number of entries per flush queue */
86 #define IOVA_FQ_SIZE 256
88 /* Timeout (in ms) after which entries are flushed from the queue */
89 #define IOVA_FQ_TIMEOUT 10
91 /* Flush queue entry for deferred flushing */
92 struct iova_fq_entry {
93 unsigned long iova_pfn;
95 struct list_head freelist;
96 u64 counter; /* Flush counter when this entry was added */
99 /* Per-CPU flush queue structure */
101 struct iova_fq_entry entries[IOVA_FQ_SIZE];
102 unsigned int head, tail;
106 #define fq_ring_for_each(i, fq) \
107 for ((i) = (fq)->head; (i) != (fq)->tail; (i) = ((i) + 1) % IOVA_FQ_SIZE)
109 static inline bool fq_full(struct iova_fq *fq)
111 assert_spin_locked(&fq->lock);
112 return (((fq->tail + 1) % IOVA_FQ_SIZE) == fq->head);
115 static inline unsigned int fq_ring_add(struct iova_fq *fq)
117 unsigned int idx = fq->tail;
119 assert_spin_locked(&fq->lock);
121 fq->tail = (idx + 1) % IOVA_FQ_SIZE;
126 static void fq_ring_free(struct iommu_dma_cookie *cookie, struct iova_fq *fq)
128 u64 counter = atomic64_read(&cookie->fq_flush_finish_cnt);
131 assert_spin_locked(&fq->lock);
133 fq_ring_for_each(idx, fq) {
135 if (fq->entries[idx].counter >= counter)
138 put_pages_list(&fq->entries[idx].freelist);
139 free_iova_fast(&cookie->iovad,
140 fq->entries[idx].iova_pfn,
141 fq->entries[idx].pages);
143 fq->head = (fq->head + 1) % IOVA_FQ_SIZE;
147 static void fq_flush_iotlb(struct iommu_dma_cookie *cookie)
149 atomic64_inc(&cookie->fq_flush_start_cnt);
150 cookie->fq_domain->ops->flush_iotlb_all(cookie->fq_domain);
151 atomic64_inc(&cookie->fq_flush_finish_cnt);
154 static void fq_flush_timeout(struct timer_list *t)
156 struct iommu_dma_cookie *cookie = from_timer(cookie, t, fq_timer);
159 atomic_set(&cookie->fq_timer_on, 0);
160 fq_flush_iotlb(cookie);
162 for_each_possible_cpu(cpu) {
166 fq = per_cpu_ptr(cookie->fq, cpu);
167 spin_lock_irqsave(&fq->lock, flags);
168 fq_ring_free(cookie, fq);
169 spin_unlock_irqrestore(&fq->lock, flags);
173 static void queue_iova(struct iommu_dma_cookie *cookie,
174 unsigned long pfn, unsigned long pages,
175 struct list_head *freelist)
182 * Order against the IOMMU driver's pagetable update from unmapping
183 * @pte, to guarantee that fq_flush_iotlb() observes that if called
184 * from a different CPU before we release the lock below. Full barrier
185 * so it also pairs with iommu_dma_init_fq() to avoid seeing partially
186 * written fq state here.
190 fq = raw_cpu_ptr(cookie->fq);
191 spin_lock_irqsave(&fq->lock, flags);
194 * First remove all entries from the flush queue that have already been
195 * flushed out on another CPU. This makes the fq_full() check below less
198 fq_ring_free(cookie, fq);
201 fq_flush_iotlb(cookie);
202 fq_ring_free(cookie, fq);
205 idx = fq_ring_add(fq);
207 fq->entries[idx].iova_pfn = pfn;
208 fq->entries[idx].pages = pages;
209 fq->entries[idx].counter = atomic64_read(&cookie->fq_flush_start_cnt);
210 list_splice(freelist, &fq->entries[idx].freelist);
212 spin_unlock_irqrestore(&fq->lock, flags);
214 /* Avoid false sharing as much as possible. */
215 if (!atomic_read(&cookie->fq_timer_on) &&
216 !atomic_xchg(&cookie->fq_timer_on, 1))
217 mod_timer(&cookie->fq_timer,
218 jiffies + msecs_to_jiffies(IOVA_FQ_TIMEOUT));
221 static void iommu_dma_free_fq(struct iommu_dma_cookie *cookie)
228 del_timer_sync(&cookie->fq_timer);
229 /* The IOVAs will be torn down separately, so just free our queued pages */
230 for_each_possible_cpu(cpu) {
231 struct iova_fq *fq = per_cpu_ptr(cookie->fq, cpu);
233 fq_ring_for_each(idx, fq)
234 put_pages_list(&fq->entries[idx].freelist);
237 free_percpu(cookie->fq);
240 /* sysfs updates are serialised by the mutex of the group owning @domain */
241 int iommu_dma_init_fq(struct iommu_domain *domain)
243 struct iommu_dma_cookie *cookie = domain->iova_cookie;
244 struct iova_fq __percpu *queue;
247 if (cookie->fq_domain)
250 atomic64_set(&cookie->fq_flush_start_cnt, 0);
251 atomic64_set(&cookie->fq_flush_finish_cnt, 0);
253 queue = alloc_percpu(struct iova_fq);
255 pr_warn("iova flush queue initialization failed\n");
259 for_each_possible_cpu(cpu) {
260 struct iova_fq *fq = per_cpu_ptr(queue, cpu);
265 spin_lock_init(&fq->lock);
267 for (i = 0; i < IOVA_FQ_SIZE; i++)
268 INIT_LIST_HEAD(&fq->entries[i].freelist);
273 timer_setup(&cookie->fq_timer, fq_flush_timeout, 0);
274 atomic_set(&cookie->fq_timer_on, 0);
276 * Prevent incomplete fq state being observable. Pairs with path from
277 * __iommu_dma_unmap() through iommu_dma_free_iova() to queue_iova()
280 WRITE_ONCE(cookie->fq_domain, domain);
284 static inline size_t cookie_msi_granule(struct iommu_dma_cookie *cookie)
286 if (cookie->type == IOMMU_DMA_IOVA_COOKIE)
287 return cookie->iovad.granule;
291 static struct iommu_dma_cookie *cookie_alloc(enum iommu_dma_cookie_type type)
293 struct iommu_dma_cookie *cookie;
295 cookie = kzalloc(sizeof(*cookie), GFP_KERNEL);
297 INIT_LIST_HEAD(&cookie->msi_page_list);
304 * iommu_get_dma_cookie - Acquire DMA-API resources for a domain
305 * @domain: IOMMU domain to prepare for DMA-API usage
307 int iommu_get_dma_cookie(struct iommu_domain *domain)
309 if (domain->iova_cookie)
312 domain->iova_cookie = cookie_alloc(IOMMU_DMA_IOVA_COOKIE);
313 if (!domain->iova_cookie)
316 mutex_init(&domain->iova_cookie->mutex);
321 * iommu_get_msi_cookie - Acquire just MSI remapping resources
322 * @domain: IOMMU domain to prepare
323 * @base: Start address of IOVA region for MSI mappings
325 * Users who manage their own IOVA allocation and do not want DMA API support,
326 * but would still like to take advantage of automatic MSI remapping, can use
327 * this to initialise their own domain appropriately. Users should reserve a
328 * contiguous IOVA region, starting at @base, large enough to accommodate the
329 * number of PAGE_SIZE mappings necessary to cover every MSI doorbell address
330 * used by the devices attached to @domain.
332 int iommu_get_msi_cookie(struct iommu_domain *domain, dma_addr_t base)
334 struct iommu_dma_cookie *cookie;
336 if (domain->type != IOMMU_DOMAIN_UNMANAGED)
339 if (domain->iova_cookie)
342 cookie = cookie_alloc(IOMMU_DMA_MSI_COOKIE);
346 cookie->msi_iova = base;
347 domain->iova_cookie = cookie;
350 EXPORT_SYMBOL(iommu_get_msi_cookie);
353 * iommu_put_dma_cookie - Release a domain's DMA mapping resources
354 * @domain: IOMMU domain previously prepared by iommu_get_dma_cookie() or
355 * iommu_get_msi_cookie()
357 void iommu_put_dma_cookie(struct iommu_domain *domain)
359 struct iommu_dma_cookie *cookie = domain->iova_cookie;
360 struct iommu_dma_msi_page *msi, *tmp;
365 if (cookie->type == IOMMU_DMA_IOVA_COOKIE && cookie->iovad.granule) {
366 iommu_dma_free_fq(cookie);
367 put_iova_domain(&cookie->iovad);
370 list_for_each_entry_safe(msi, tmp, &cookie->msi_page_list, list) {
371 list_del(&msi->list);
375 domain->iova_cookie = NULL;
379 * iommu_dma_get_resv_regions - Reserved region driver helper
380 * @dev: Device from iommu_get_resv_regions()
381 * @list: Reserved region list from iommu_get_resv_regions()
383 * IOMMU drivers can use this to implement their .get_resv_regions callback
384 * for general non-IOMMU-specific reservations. Currently, this covers GICv3
385 * ITS region reservation on ACPI based ARM platforms that may require HW MSI
388 void iommu_dma_get_resv_regions(struct device *dev, struct list_head *list)
391 if (!is_of_node(dev_iommu_fwspec_get(dev)->iommu_fwnode))
392 iort_iommu_get_resv_regions(dev, list);
395 EXPORT_SYMBOL(iommu_dma_get_resv_regions);
397 static int cookie_init_hw_msi_region(struct iommu_dma_cookie *cookie,
398 phys_addr_t start, phys_addr_t end)
400 struct iova_domain *iovad = &cookie->iovad;
401 struct iommu_dma_msi_page *msi_page;
404 start -= iova_offset(iovad, start);
405 num_pages = iova_align(iovad, end - start) >> iova_shift(iovad);
407 for (i = 0; i < num_pages; i++) {
408 msi_page = kmalloc(sizeof(*msi_page), GFP_KERNEL);
412 msi_page->phys = start;
413 msi_page->iova = start;
414 INIT_LIST_HEAD(&msi_page->list);
415 list_add(&msi_page->list, &cookie->msi_page_list);
416 start += iovad->granule;
422 static int iommu_dma_ranges_sort(void *priv, const struct list_head *a,
423 const struct list_head *b)
425 struct resource_entry *res_a = list_entry(a, typeof(*res_a), node);
426 struct resource_entry *res_b = list_entry(b, typeof(*res_b), node);
428 return res_a->res->start > res_b->res->start;
431 static int iova_reserve_pci_windows(struct pci_dev *dev,
432 struct iova_domain *iovad)
434 struct pci_host_bridge *bridge = pci_find_host_bridge(dev->bus);
435 struct resource_entry *window;
436 unsigned long lo, hi;
437 phys_addr_t start = 0, end;
439 resource_list_for_each_entry(window, &bridge->windows) {
440 if (resource_type(window->res) != IORESOURCE_MEM)
443 lo = iova_pfn(iovad, window->res->start - window->offset);
444 hi = iova_pfn(iovad, window->res->end - window->offset);
445 reserve_iova(iovad, lo, hi);
448 /* Get reserved DMA windows from host bridge */
449 list_sort(NULL, &bridge->dma_ranges, iommu_dma_ranges_sort);
450 resource_list_for_each_entry(window, &bridge->dma_ranges) {
451 end = window->res->start - window->offset;
454 lo = iova_pfn(iovad, start);
455 hi = iova_pfn(iovad, end);
456 reserve_iova(iovad, lo, hi);
457 } else if (end < start) {
458 /* DMA ranges should be non-overlapping */
460 "Failed to reserve IOVA [%pa-%pa]\n",
465 start = window->res->end - window->offset + 1;
466 /* If window is last entry */
467 if (window->node.next == &bridge->dma_ranges &&
468 end != ~(phys_addr_t)0) {
469 end = ~(phys_addr_t)0;
477 static int iova_reserve_iommu_regions(struct device *dev,
478 struct iommu_domain *domain)
480 struct iommu_dma_cookie *cookie = domain->iova_cookie;
481 struct iova_domain *iovad = &cookie->iovad;
482 struct iommu_resv_region *region;
483 LIST_HEAD(resv_regions);
486 if (dev_is_pci(dev)) {
487 ret = iova_reserve_pci_windows(to_pci_dev(dev), iovad);
492 iommu_get_resv_regions(dev, &resv_regions);
493 list_for_each_entry(region, &resv_regions, list) {
494 unsigned long lo, hi;
496 /* We ARE the software that manages these! */
497 if (region->type == IOMMU_RESV_SW_MSI)
500 lo = iova_pfn(iovad, region->start);
501 hi = iova_pfn(iovad, region->start + region->length - 1);
502 reserve_iova(iovad, lo, hi);
504 if (region->type == IOMMU_RESV_MSI)
505 ret = cookie_init_hw_msi_region(cookie, region->start,
506 region->start + region->length);
510 iommu_put_resv_regions(dev, &resv_regions);
515 static bool dev_is_untrusted(struct device *dev)
517 return dev_is_pci(dev) && to_pci_dev(dev)->untrusted;
520 static bool dev_use_swiotlb(struct device *dev)
522 return IS_ENABLED(CONFIG_SWIOTLB) && dev_is_untrusted(dev);
526 * iommu_dma_init_domain - Initialise a DMA mapping domain
527 * @domain: IOMMU domain previously prepared by iommu_get_dma_cookie()
528 * @base: IOVA at which the mappable address space starts
529 * @limit: Last address of the IOVA space
530 * @dev: Device the domain is being initialised for
532 * @base and @limit + 1 should be exact multiples of IOMMU page granularity to
533 * avoid rounding surprises. If necessary, we reserve the page at address 0
534 * to ensure it is an invalid IOVA. It is safe to reinitialise a domain, but
535 * any change which could make prior IOVAs invalid will fail.
537 static int iommu_dma_init_domain(struct iommu_domain *domain, dma_addr_t base,
538 dma_addr_t limit, struct device *dev)
540 struct iommu_dma_cookie *cookie = domain->iova_cookie;
541 unsigned long order, base_pfn;
542 struct iova_domain *iovad;
545 if (!cookie || cookie->type != IOMMU_DMA_IOVA_COOKIE)
548 iovad = &cookie->iovad;
550 /* Use the smallest supported page size for IOVA granularity */
551 order = __ffs(domain->pgsize_bitmap);
552 base_pfn = max_t(unsigned long, 1, base >> order);
554 /* Check the domain allows at least some access to the device... */
555 if (domain->geometry.force_aperture) {
556 if (base > domain->geometry.aperture_end ||
557 limit < domain->geometry.aperture_start) {
558 pr_warn("specified DMA range outside IOMMU capability\n");
561 /* ...then finally give it a kicking to make sure it fits */
562 base_pfn = max_t(unsigned long, base_pfn,
563 domain->geometry.aperture_start >> order);
566 /* start_pfn is always nonzero for an already-initialised domain */
567 mutex_lock(&cookie->mutex);
568 if (iovad->start_pfn) {
569 if (1UL << order != iovad->granule ||
570 base_pfn != iovad->start_pfn) {
571 pr_warn("Incompatible range for DMA domain\n");
580 init_iova_domain(iovad, 1UL << order, base_pfn);
581 ret = iova_domain_init_rcaches(iovad);
585 /* If the FQ fails we can simply fall back to strict mode */
586 if (domain->type == IOMMU_DOMAIN_DMA_FQ && iommu_dma_init_fq(domain))
587 domain->type = IOMMU_DOMAIN_DMA;
589 ret = iova_reserve_iommu_regions(dev, domain);
592 mutex_unlock(&cookie->mutex);
597 * dma_info_to_prot - Translate DMA API directions and attributes to IOMMU API
599 * @dir: Direction of DMA transfer
600 * @coherent: Is the DMA master cache-coherent?
601 * @attrs: DMA attributes for the mapping
603 * Return: corresponding IOMMU API page protection flags
605 static int dma_info_to_prot(enum dma_data_direction dir, bool coherent,
608 int prot = coherent ? IOMMU_CACHE : 0;
610 if (attrs & DMA_ATTR_PRIVILEGED)
614 case DMA_BIDIRECTIONAL:
615 return prot | IOMMU_READ | IOMMU_WRITE;
617 return prot | IOMMU_READ;
618 case DMA_FROM_DEVICE:
619 return prot | IOMMU_WRITE;
625 static dma_addr_t iommu_dma_alloc_iova(struct iommu_domain *domain,
626 size_t size, u64 dma_limit, struct device *dev)
628 struct iommu_dma_cookie *cookie = domain->iova_cookie;
629 struct iova_domain *iovad = &cookie->iovad;
630 unsigned long shift, iova_len, iova = 0;
632 if (cookie->type == IOMMU_DMA_MSI_COOKIE) {
633 cookie->msi_iova += size;
634 return cookie->msi_iova - size;
637 shift = iova_shift(iovad);
638 iova_len = size >> shift;
640 dma_limit = min_not_zero(dma_limit, dev->bus_dma_limit);
642 if (domain->geometry.force_aperture)
643 dma_limit = min(dma_limit, (u64)domain->geometry.aperture_end);
645 /* Try to get PCI devices a SAC address */
646 if (dma_limit > DMA_BIT_MASK(32) && !iommu_dma_forcedac && dev_is_pci(dev))
647 iova = alloc_iova_fast(iovad, iova_len,
648 DMA_BIT_MASK(32) >> shift, false);
651 iova = alloc_iova_fast(iovad, iova_len, dma_limit >> shift,
654 return (dma_addr_t)iova << shift;
657 static void iommu_dma_free_iova(struct iommu_dma_cookie *cookie,
658 dma_addr_t iova, size_t size, struct iommu_iotlb_gather *gather)
660 struct iova_domain *iovad = &cookie->iovad;
662 /* The MSI case is only ever cleaning up its most recent allocation */
663 if (cookie->type == IOMMU_DMA_MSI_COOKIE)
664 cookie->msi_iova -= size;
665 else if (gather && gather->queued)
666 queue_iova(cookie, iova_pfn(iovad, iova),
667 size >> iova_shift(iovad),
670 free_iova_fast(iovad, iova_pfn(iovad, iova),
671 size >> iova_shift(iovad));
674 static void __iommu_dma_unmap(struct device *dev, dma_addr_t dma_addr,
677 struct iommu_domain *domain = iommu_get_dma_domain(dev);
678 struct iommu_dma_cookie *cookie = domain->iova_cookie;
679 struct iova_domain *iovad = &cookie->iovad;
680 size_t iova_off = iova_offset(iovad, dma_addr);
681 struct iommu_iotlb_gather iotlb_gather;
684 dma_addr -= iova_off;
685 size = iova_align(iovad, size + iova_off);
686 iommu_iotlb_gather_init(&iotlb_gather);
687 iotlb_gather.queued = READ_ONCE(cookie->fq_domain);
689 unmapped = iommu_unmap_fast(domain, dma_addr, size, &iotlb_gather);
690 WARN_ON(unmapped != size);
692 if (!iotlb_gather.queued)
693 iommu_iotlb_sync(domain, &iotlb_gather);
694 iommu_dma_free_iova(cookie, dma_addr, size, &iotlb_gather);
697 static dma_addr_t __iommu_dma_map(struct device *dev, phys_addr_t phys,
698 size_t size, int prot, u64 dma_mask)
700 struct iommu_domain *domain = iommu_get_dma_domain(dev);
701 struct iommu_dma_cookie *cookie = domain->iova_cookie;
702 struct iova_domain *iovad = &cookie->iovad;
703 size_t iova_off = iova_offset(iovad, phys);
706 if (static_branch_unlikely(&iommu_deferred_attach_enabled) &&
707 iommu_deferred_attach(dev, domain))
708 return DMA_MAPPING_ERROR;
710 size = iova_align(iovad, size + iova_off);
712 iova = iommu_dma_alloc_iova(domain, size, dma_mask, dev);
714 return DMA_MAPPING_ERROR;
716 if (iommu_map_atomic(domain, iova, phys - iova_off, size, prot)) {
717 iommu_dma_free_iova(cookie, iova, size, NULL);
718 return DMA_MAPPING_ERROR;
720 return iova + iova_off;
723 static void __iommu_dma_free_pages(struct page **pages, int count)
726 __free_page(pages[count]);
730 static struct page **__iommu_dma_alloc_pages(struct device *dev,
731 unsigned int count, unsigned long order_mask, gfp_t gfp)
734 unsigned int i = 0, nid = dev_to_node(dev);
736 order_mask &= (2U << MAX_ORDER) - 1;
740 pages = kvcalloc(count, sizeof(*pages), GFP_KERNEL);
744 /* IOMMU can map any pages, so himem can also be used here */
745 gfp |= __GFP_NOWARN | __GFP_HIGHMEM;
747 /* It makes no sense to muck about with huge pages */
751 struct page *page = NULL;
752 unsigned int order_size;
755 * Higher-order allocations are a convenience rather
756 * than a necessity, hence using __GFP_NORETRY until
757 * falling back to minimum-order allocations.
759 for (order_mask &= (2U << __fls(count)) - 1;
760 order_mask; order_mask &= ~order_size) {
761 unsigned int order = __fls(order_mask);
762 gfp_t alloc_flags = gfp;
764 order_size = 1U << order;
765 if (order_mask > order_size)
766 alloc_flags |= __GFP_NORETRY;
767 page = alloc_pages_node(nid, alloc_flags, order);
771 split_page(page, order);
775 __iommu_dma_free_pages(pages, i);
786 * If size is less than PAGE_SIZE, then a full CPU page will be allocated,
787 * but an IOMMU which supports smaller pages might not map the whole thing.
789 static struct page **__iommu_dma_alloc_noncontiguous(struct device *dev,
790 size_t size, struct sg_table *sgt, gfp_t gfp, pgprot_t prot,
793 struct iommu_domain *domain = iommu_get_dma_domain(dev);
794 struct iommu_dma_cookie *cookie = domain->iova_cookie;
795 struct iova_domain *iovad = &cookie->iovad;
796 bool coherent = dev_is_dma_coherent(dev);
797 int ioprot = dma_info_to_prot(DMA_BIDIRECTIONAL, coherent, attrs);
798 unsigned int count, min_size, alloc_sizes = domain->pgsize_bitmap;
803 if (static_branch_unlikely(&iommu_deferred_attach_enabled) &&
804 iommu_deferred_attach(dev, domain))
807 min_size = alloc_sizes & -alloc_sizes;
808 if (min_size < PAGE_SIZE) {
809 min_size = PAGE_SIZE;
810 alloc_sizes |= PAGE_SIZE;
812 size = ALIGN(size, min_size);
814 if (attrs & DMA_ATTR_ALLOC_SINGLE_PAGES)
815 alloc_sizes = min_size;
817 count = PAGE_ALIGN(size) >> PAGE_SHIFT;
818 pages = __iommu_dma_alloc_pages(dev, count, alloc_sizes >> PAGE_SHIFT,
823 size = iova_align(iovad, size);
824 iova = iommu_dma_alloc_iova(domain, size, dev->coherent_dma_mask, dev);
828 if (sg_alloc_table_from_pages(sgt, pages, count, 0, size, GFP_KERNEL))
831 if (!(ioprot & IOMMU_CACHE)) {
832 struct scatterlist *sg;
835 for_each_sg(sgt->sgl, sg, sgt->orig_nents, i)
836 arch_dma_prep_coherent(sg_page(sg), sg->length);
839 ret = iommu_map_sg_atomic(domain, iova, sgt->sgl, sgt->orig_nents, ioprot);
840 if (ret < 0 || ret < size)
843 sgt->sgl->dma_address = iova;
844 sgt->sgl->dma_length = size;
850 iommu_dma_free_iova(cookie, iova, size, NULL);
852 __iommu_dma_free_pages(pages, count);
856 static void *iommu_dma_alloc_remap(struct device *dev, size_t size,
857 dma_addr_t *dma_handle, gfp_t gfp, pgprot_t prot,
864 pages = __iommu_dma_alloc_noncontiguous(dev, size, &sgt, gfp, prot,
868 *dma_handle = sgt.sgl->dma_address;
870 vaddr = dma_common_pages_remap(pages, size, prot,
871 __builtin_return_address(0));
877 __iommu_dma_unmap(dev, *dma_handle, size);
878 __iommu_dma_free_pages(pages, PAGE_ALIGN(size) >> PAGE_SHIFT);
882 static struct sg_table *iommu_dma_alloc_noncontiguous(struct device *dev,
883 size_t size, enum dma_data_direction dir, gfp_t gfp,
886 struct dma_sgt_handle *sh;
888 sh = kmalloc(sizeof(*sh), gfp);
892 sh->pages = __iommu_dma_alloc_noncontiguous(dev, size, &sh->sgt, gfp,
901 static void iommu_dma_free_noncontiguous(struct device *dev, size_t size,
902 struct sg_table *sgt, enum dma_data_direction dir)
904 struct dma_sgt_handle *sh = sgt_handle(sgt);
906 __iommu_dma_unmap(dev, sgt->sgl->dma_address, size);
907 __iommu_dma_free_pages(sh->pages, PAGE_ALIGN(size) >> PAGE_SHIFT);
908 sg_free_table(&sh->sgt);
912 static void iommu_dma_sync_single_for_cpu(struct device *dev,
913 dma_addr_t dma_handle, size_t size, enum dma_data_direction dir)
917 if (dev_is_dma_coherent(dev) && !dev_use_swiotlb(dev))
920 phys = iommu_iova_to_phys(iommu_get_dma_domain(dev), dma_handle);
921 if (!dev_is_dma_coherent(dev))
922 arch_sync_dma_for_cpu(phys, size, dir);
924 if (is_swiotlb_buffer(dev, phys))
925 swiotlb_sync_single_for_cpu(dev, phys, size, dir);
928 static void iommu_dma_sync_single_for_device(struct device *dev,
929 dma_addr_t dma_handle, size_t size, enum dma_data_direction dir)
933 if (dev_is_dma_coherent(dev) && !dev_use_swiotlb(dev))
936 phys = iommu_iova_to_phys(iommu_get_dma_domain(dev), dma_handle);
937 if (is_swiotlb_buffer(dev, phys))
938 swiotlb_sync_single_for_device(dev, phys, size, dir);
940 if (!dev_is_dma_coherent(dev))
941 arch_sync_dma_for_device(phys, size, dir);
944 static void iommu_dma_sync_sg_for_cpu(struct device *dev,
945 struct scatterlist *sgl, int nelems,
946 enum dma_data_direction dir)
948 struct scatterlist *sg;
951 if (dev_use_swiotlb(dev))
952 for_each_sg(sgl, sg, nelems, i)
953 iommu_dma_sync_single_for_cpu(dev, sg_dma_address(sg),
955 else if (!dev_is_dma_coherent(dev))
956 for_each_sg(sgl, sg, nelems, i)
957 arch_sync_dma_for_cpu(sg_phys(sg), sg->length, dir);
960 static void iommu_dma_sync_sg_for_device(struct device *dev,
961 struct scatterlist *sgl, int nelems,
962 enum dma_data_direction dir)
964 struct scatterlist *sg;
967 if (dev_use_swiotlb(dev))
968 for_each_sg(sgl, sg, nelems, i)
969 iommu_dma_sync_single_for_device(dev,
972 else if (!dev_is_dma_coherent(dev))
973 for_each_sg(sgl, sg, nelems, i)
974 arch_sync_dma_for_device(sg_phys(sg), sg->length, dir);
977 static dma_addr_t iommu_dma_map_page(struct device *dev, struct page *page,
978 unsigned long offset, size_t size, enum dma_data_direction dir,
981 phys_addr_t phys = page_to_phys(page) + offset;
982 bool coherent = dev_is_dma_coherent(dev);
983 int prot = dma_info_to_prot(dir, coherent, attrs);
984 struct iommu_domain *domain = iommu_get_dma_domain(dev);
985 struct iommu_dma_cookie *cookie = domain->iova_cookie;
986 struct iova_domain *iovad = &cookie->iovad;
987 dma_addr_t iova, dma_mask = dma_get_mask(dev);
990 * If both the physical buffer start address and size are
991 * page aligned, we don't need to use a bounce page.
993 if (dev_use_swiotlb(dev) && iova_offset(iovad, phys | size)) {
995 size_t padding_size, aligned_size;
997 if (!is_swiotlb_active(dev)) {
998 dev_warn_once(dev, "DMA bounce buffers are inactive, unable to map unaligned transaction.\n");
999 return DMA_MAPPING_ERROR;
1002 aligned_size = iova_align(iovad, size);
1003 phys = swiotlb_tbl_map_single(dev, phys, size, aligned_size,
1004 iova_mask(iovad), dir, attrs);
1006 if (phys == DMA_MAPPING_ERROR)
1007 return DMA_MAPPING_ERROR;
1009 /* Cleanup the padding area. */
1010 padding_start = phys_to_virt(phys);
1011 padding_size = aligned_size;
1013 if (!(attrs & DMA_ATTR_SKIP_CPU_SYNC) &&
1014 (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL)) {
1015 padding_start += size;
1016 padding_size -= size;
1019 memset(padding_start, 0, padding_size);
1022 if (!coherent && !(attrs & DMA_ATTR_SKIP_CPU_SYNC))
1023 arch_sync_dma_for_device(phys, size, dir);
1025 iova = __iommu_dma_map(dev, phys, size, prot, dma_mask);
1026 if (iova == DMA_MAPPING_ERROR && is_swiotlb_buffer(dev, phys))
1027 swiotlb_tbl_unmap_single(dev, phys, size, dir, attrs);
1031 static void iommu_dma_unmap_page(struct device *dev, dma_addr_t dma_handle,
1032 size_t size, enum dma_data_direction dir, unsigned long attrs)
1034 struct iommu_domain *domain = iommu_get_dma_domain(dev);
1037 phys = iommu_iova_to_phys(domain, dma_handle);
1041 if (!(attrs & DMA_ATTR_SKIP_CPU_SYNC) && !dev_is_dma_coherent(dev))
1042 arch_sync_dma_for_cpu(phys, size, dir);
1044 __iommu_dma_unmap(dev, dma_handle, size);
1046 if (unlikely(is_swiotlb_buffer(dev, phys)))
1047 swiotlb_tbl_unmap_single(dev, phys, size, dir, attrs);
1051 * Prepare a successfully-mapped scatterlist to give back to the caller.
1053 * At this point the segments are already laid out by iommu_dma_map_sg() to
1054 * avoid individually crossing any boundaries, so we merely need to check a
1055 * segment's start address to avoid concatenating across one.
1057 static int __finalise_sg(struct device *dev, struct scatterlist *sg, int nents,
1058 dma_addr_t dma_addr)
1060 struct scatterlist *s, *cur = sg;
1061 unsigned long seg_mask = dma_get_seg_boundary(dev);
1062 unsigned int cur_len = 0, max_len = dma_get_max_seg_size(dev);
1065 for_each_sg(sg, s, nents, i) {
1066 /* Restore this segment's original unaligned fields first */
1067 dma_addr_t s_dma_addr = sg_dma_address(s);
1068 unsigned int s_iova_off = sg_dma_address(s);
1069 unsigned int s_length = sg_dma_len(s);
1070 unsigned int s_iova_len = s->length;
1072 sg_dma_address(s) = DMA_MAPPING_ERROR;
1075 if (sg_is_dma_bus_address(s)) {
1079 sg_dma_unmark_bus_address(s);
1080 sg_dma_address(cur) = s_dma_addr;
1081 sg_dma_len(cur) = s_length;
1082 sg_dma_mark_bus_address(cur);
1088 s->offset += s_iova_off;
1089 s->length = s_length;
1092 * Now fill in the real DMA data. If...
1093 * - there is a valid output segment to append to
1094 * - and this segment starts on an IOVA page boundary
1095 * - but doesn't fall at a segment boundary
1096 * - and wouldn't make the resulting output segment too long
1098 if (cur_len && !s_iova_off && (dma_addr & seg_mask) &&
1099 (max_len - cur_len >= s_length)) {
1100 /* ...then concatenate it with the previous one */
1101 cur_len += s_length;
1103 /* Otherwise start the next output segment */
1109 sg_dma_address(cur) = dma_addr + s_iova_off;
1112 sg_dma_len(cur) = cur_len;
1113 dma_addr += s_iova_len;
1115 if (s_length + s_iova_off < s_iova_len)
1122 * If mapping failed, then just restore the original list,
1123 * but making sure the DMA fields are invalidated.
1125 static void __invalidate_sg(struct scatterlist *sg, int nents)
1127 struct scatterlist *s;
1130 for_each_sg(sg, s, nents, i) {
1131 if (sg_is_dma_bus_address(s)) {
1132 sg_dma_unmark_bus_address(s);
1134 if (sg_dma_address(s) != DMA_MAPPING_ERROR)
1135 s->offset += sg_dma_address(s);
1137 s->length = sg_dma_len(s);
1139 sg_dma_address(s) = DMA_MAPPING_ERROR;
1144 static void iommu_dma_unmap_sg_swiotlb(struct device *dev, struct scatterlist *sg,
1145 int nents, enum dma_data_direction dir, unsigned long attrs)
1147 struct scatterlist *s;
1150 for_each_sg(sg, s, nents, i)
1151 iommu_dma_unmap_page(dev, sg_dma_address(s),
1152 sg_dma_len(s), dir, attrs);
1155 static int iommu_dma_map_sg_swiotlb(struct device *dev, struct scatterlist *sg,
1156 int nents, enum dma_data_direction dir, unsigned long attrs)
1158 struct scatterlist *s;
1161 for_each_sg(sg, s, nents, i) {
1162 sg_dma_address(s) = iommu_dma_map_page(dev, sg_page(s),
1163 s->offset, s->length, dir, attrs);
1164 if (sg_dma_address(s) == DMA_MAPPING_ERROR)
1166 sg_dma_len(s) = s->length;
1172 iommu_dma_unmap_sg_swiotlb(dev, sg, i, dir, attrs | DMA_ATTR_SKIP_CPU_SYNC);
1177 * The DMA API client is passing in a scatterlist which could describe
1178 * any old buffer layout, but the IOMMU API requires everything to be
1179 * aligned to IOMMU pages. Hence the need for this complicated bit of
1180 * impedance-matching, to be able to hand off a suitably-aligned list,
1181 * but still preserve the original offsets and sizes for the caller.
1183 static int iommu_dma_map_sg(struct device *dev, struct scatterlist *sg,
1184 int nents, enum dma_data_direction dir, unsigned long attrs)
1186 struct iommu_domain *domain = iommu_get_dma_domain(dev);
1187 struct iommu_dma_cookie *cookie = domain->iova_cookie;
1188 struct iova_domain *iovad = &cookie->iovad;
1189 struct scatterlist *s, *prev = NULL;
1190 int prot = dma_info_to_prot(dir, dev_is_dma_coherent(dev), attrs);
1191 struct pci_p2pdma_map_state p2pdma_state = {};
1192 enum pci_p2pdma_map_type map;
1194 size_t iova_len = 0;
1195 unsigned long mask = dma_get_seg_boundary(dev);
1199 if (static_branch_unlikely(&iommu_deferred_attach_enabled)) {
1200 ret = iommu_deferred_attach(dev, domain);
1205 if (dev_use_swiotlb(dev))
1206 return iommu_dma_map_sg_swiotlb(dev, sg, nents, dir, attrs);
1208 if (!(attrs & DMA_ATTR_SKIP_CPU_SYNC))
1209 iommu_dma_sync_sg_for_device(dev, sg, nents, dir);
1212 * Work out how much IOVA space we need, and align the segments to
1213 * IOVA granules for the IOMMU driver to handle. With some clever
1214 * trickery we can modify the list in-place, but reversibly, by
1215 * stashing the unaligned parts in the as-yet-unused DMA fields.
1217 for_each_sg(sg, s, nents, i) {
1218 size_t s_iova_off = iova_offset(iovad, s->offset);
1219 size_t s_length = s->length;
1220 size_t pad_len = (mask - iova_len + 1) & mask;
1222 if (is_pci_p2pdma_page(sg_page(s))) {
1223 map = pci_p2pdma_map_segment(&p2pdma_state, dev, s);
1225 case PCI_P2PDMA_MAP_BUS_ADDR:
1227 * iommu_map_sg() will skip this segment as
1228 * it is marked as a bus address,
1229 * __finalise_sg() will copy the dma address
1230 * into the output segment.
1233 case PCI_P2PDMA_MAP_THRU_HOST_BRIDGE:
1235 * Mapping through host bridge should be
1236 * mapped with regular IOVAs, thus we
1237 * do nothing here and continue below.
1242 goto out_restore_sg;
1246 sg_dma_address(s) = s_iova_off;
1247 sg_dma_len(s) = s_length;
1248 s->offset -= s_iova_off;
1249 s_length = iova_align(iovad, s_length + s_iova_off);
1250 s->length = s_length;
1253 * Due to the alignment of our single IOVA allocation, we can
1254 * depend on these assumptions about the segment boundary mask:
1255 * - If mask size >= IOVA size, then the IOVA range cannot
1256 * possibly fall across a boundary, so we don't care.
1257 * - If mask size < IOVA size, then the IOVA range must start
1258 * exactly on a boundary, therefore we can lay things out
1259 * based purely on segment lengths without needing to know
1260 * the actual addresses beforehand.
1261 * - The mask must be a power of 2, so pad_len == 0 if
1262 * iova_len == 0, thus we cannot dereference prev the first
1263 * time through here (i.e. before it has a meaningful value).
1265 if (pad_len && pad_len < s_length - 1) {
1266 prev->length += pad_len;
1267 iova_len += pad_len;
1270 iova_len += s_length;
1275 return __finalise_sg(dev, sg, nents, 0);
1277 iova = iommu_dma_alloc_iova(domain, iova_len, dma_get_mask(dev), dev);
1280 goto out_restore_sg;
1284 * We'll leave any physical concatenation to the IOMMU driver's
1285 * implementation - it knows better than we do.
1287 ret = iommu_map_sg_atomic(domain, iova, sg, nents, prot);
1288 if (ret < 0 || ret < iova_len)
1291 return __finalise_sg(dev, sg, nents, iova);
1294 iommu_dma_free_iova(cookie, iova, iova_len, NULL);
1296 __invalidate_sg(sg, nents);
1298 if (ret != -ENOMEM && ret != -EREMOTEIO)
1303 static void iommu_dma_unmap_sg(struct device *dev, struct scatterlist *sg,
1304 int nents, enum dma_data_direction dir, unsigned long attrs)
1306 dma_addr_t end = 0, start;
1307 struct scatterlist *tmp;
1310 if (dev_use_swiotlb(dev)) {
1311 iommu_dma_unmap_sg_swiotlb(dev, sg, nents, dir, attrs);
1315 if (!(attrs & DMA_ATTR_SKIP_CPU_SYNC))
1316 iommu_dma_sync_sg_for_cpu(dev, sg, nents, dir);
1319 * The scatterlist segments are mapped into a single
1320 * contiguous IOVA allocation, the start and end points
1321 * just have to be determined.
1323 for_each_sg(sg, tmp, nents, i) {
1324 if (sg_is_dma_bus_address(tmp)) {
1325 sg_dma_unmark_bus_address(tmp);
1329 if (sg_dma_len(tmp) == 0)
1332 start = sg_dma_address(tmp);
1337 for_each_sg(tmp, tmp, nents, i) {
1338 if (sg_is_dma_bus_address(tmp)) {
1339 sg_dma_unmark_bus_address(tmp);
1343 if (sg_dma_len(tmp) == 0)
1346 end = sg_dma_address(tmp) + sg_dma_len(tmp);
1350 __iommu_dma_unmap(dev, start, end - start);
1353 static dma_addr_t iommu_dma_map_resource(struct device *dev, phys_addr_t phys,
1354 size_t size, enum dma_data_direction dir, unsigned long attrs)
1356 return __iommu_dma_map(dev, phys, size,
1357 dma_info_to_prot(dir, false, attrs) | IOMMU_MMIO,
1361 static void iommu_dma_unmap_resource(struct device *dev, dma_addr_t handle,
1362 size_t size, enum dma_data_direction dir, unsigned long attrs)
1364 __iommu_dma_unmap(dev, handle, size);
1367 static void __iommu_dma_free(struct device *dev, size_t size, void *cpu_addr)
1369 size_t alloc_size = PAGE_ALIGN(size);
1370 int count = alloc_size >> PAGE_SHIFT;
1371 struct page *page = NULL, **pages = NULL;
1373 /* Non-coherent atomic allocation? Easy */
1374 if (IS_ENABLED(CONFIG_DMA_DIRECT_REMAP) &&
1375 dma_free_from_pool(dev, cpu_addr, alloc_size))
1378 if (is_vmalloc_addr(cpu_addr)) {
1380 * If it the address is remapped, then it's either non-coherent
1381 * or highmem CMA, or an iommu_dma_alloc_remap() construction.
1383 pages = dma_common_find_pages(cpu_addr);
1385 page = vmalloc_to_page(cpu_addr);
1386 dma_common_free_remap(cpu_addr, alloc_size);
1388 /* Lowmem means a coherent atomic or CMA allocation */
1389 page = virt_to_page(cpu_addr);
1393 __iommu_dma_free_pages(pages, count);
1395 dma_free_contiguous(dev, page, alloc_size);
1398 static void iommu_dma_free(struct device *dev, size_t size, void *cpu_addr,
1399 dma_addr_t handle, unsigned long attrs)
1401 __iommu_dma_unmap(dev, handle, size);
1402 __iommu_dma_free(dev, size, cpu_addr);
1405 static void *iommu_dma_alloc_pages(struct device *dev, size_t size,
1406 struct page **pagep, gfp_t gfp, unsigned long attrs)
1408 bool coherent = dev_is_dma_coherent(dev);
1409 size_t alloc_size = PAGE_ALIGN(size);
1410 int node = dev_to_node(dev);
1411 struct page *page = NULL;
1414 page = dma_alloc_contiguous(dev, alloc_size, gfp);
1416 page = alloc_pages_node(node, gfp, get_order(alloc_size));
1420 if (!coherent || PageHighMem(page)) {
1421 pgprot_t prot = dma_pgprot(dev, PAGE_KERNEL, attrs);
1423 cpu_addr = dma_common_contiguous_remap(page, alloc_size,
1424 prot, __builtin_return_address(0));
1426 goto out_free_pages;
1429 arch_dma_prep_coherent(page, size);
1431 cpu_addr = page_address(page);
1435 memset(cpu_addr, 0, alloc_size);
1438 dma_free_contiguous(dev, page, alloc_size);
1442 static void *iommu_dma_alloc(struct device *dev, size_t size,
1443 dma_addr_t *handle, gfp_t gfp, unsigned long attrs)
1445 bool coherent = dev_is_dma_coherent(dev);
1446 int ioprot = dma_info_to_prot(DMA_BIDIRECTIONAL, coherent, attrs);
1447 struct page *page = NULL;
1452 if (gfpflags_allow_blocking(gfp) &&
1453 !(attrs & DMA_ATTR_FORCE_CONTIGUOUS)) {
1454 return iommu_dma_alloc_remap(dev, size, handle, gfp,
1455 dma_pgprot(dev, PAGE_KERNEL, attrs), attrs);
1458 if (IS_ENABLED(CONFIG_DMA_DIRECT_REMAP) &&
1459 !gfpflags_allow_blocking(gfp) && !coherent)
1460 page = dma_alloc_from_pool(dev, PAGE_ALIGN(size), &cpu_addr,
1463 cpu_addr = iommu_dma_alloc_pages(dev, size, &page, gfp, attrs);
1467 *handle = __iommu_dma_map(dev, page_to_phys(page), size, ioprot,
1468 dev->coherent_dma_mask);
1469 if (*handle == DMA_MAPPING_ERROR) {
1470 __iommu_dma_free(dev, size, cpu_addr);
1477 static int iommu_dma_mmap(struct device *dev, struct vm_area_struct *vma,
1478 void *cpu_addr, dma_addr_t dma_addr, size_t size,
1479 unsigned long attrs)
1481 unsigned long nr_pages = PAGE_ALIGN(size) >> PAGE_SHIFT;
1482 unsigned long pfn, off = vma->vm_pgoff;
1485 vma->vm_page_prot = dma_pgprot(dev, vma->vm_page_prot, attrs);
1487 if (dma_mmap_from_dev_coherent(dev, vma, cpu_addr, size, &ret))
1490 if (off >= nr_pages || vma_pages(vma) > nr_pages - off)
1493 if (is_vmalloc_addr(cpu_addr)) {
1494 struct page **pages = dma_common_find_pages(cpu_addr);
1497 return vm_map_pages(vma, pages, nr_pages);
1498 pfn = vmalloc_to_pfn(cpu_addr);
1500 pfn = page_to_pfn(virt_to_page(cpu_addr));
1503 return remap_pfn_range(vma, vma->vm_start, pfn + off,
1504 vma->vm_end - vma->vm_start,
1508 static int iommu_dma_get_sgtable(struct device *dev, struct sg_table *sgt,
1509 void *cpu_addr, dma_addr_t dma_addr, size_t size,
1510 unsigned long attrs)
1515 if (is_vmalloc_addr(cpu_addr)) {
1516 struct page **pages = dma_common_find_pages(cpu_addr);
1519 return sg_alloc_table_from_pages(sgt, pages,
1520 PAGE_ALIGN(size) >> PAGE_SHIFT,
1521 0, size, GFP_KERNEL);
1524 page = vmalloc_to_page(cpu_addr);
1526 page = virt_to_page(cpu_addr);
1529 ret = sg_alloc_table(sgt, 1, GFP_KERNEL);
1531 sg_set_page(sgt->sgl, page, PAGE_ALIGN(size), 0);
1535 static unsigned long iommu_dma_get_merge_boundary(struct device *dev)
1537 struct iommu_domain *domain = iommu_get_dma_domain(dev);
1539 return (1UL << __ffs(domain->pgsize_bitmap)) - 1;
1542 static size_t iommu_dma_opt_mapping_size(void)
1544 return iova_rcache_range();
1547 static const struct dma_map_ops iommu_dma_ops = {
1548 .flags = DMA_F_PCI_P2PDMA_SUPPORTED,
1549 .alloc = iommu_dma_alloc,
1550 .free = iommu_dma_free,
1551 .alloc_pages = dma_common_alloc_pages,
1552 .free_pages = dma_common_free_pages,
1553 .alloc_noncontiguous = iommu_dma_alloc_noncontiguous,
1554 .free_noncontiguous = iommu_dma_free_noncontiguous,
1555 .mmap = iommu_dma_mmap,
1556 .get_sgtable = iommu_dma_get_sgtable,
1557 .map_page = iommu_dma_map_page,
1558 .unmap_page = iommu_dma_unmap_page,
1559 .map_sg = iommu_dma_map_sg,
1560 .unmap_sg = iommu_dma_unmap_sg,
1561 .sync_single_for_cpu = iommu_dma_sync_single_for_cpu,
1562 .sync_single_for_device = iommu_dma_sync_single_for_device,
1563 .sync_sg_for_cpu = iommu_dma_sync_sg_for_cpu,
1564 .sync_sg_for_device = iommu_dma_sync_sg_for_device,
1565 .map_resource = iommu_dma_map_resource,
1566 .unmap_resource = iommu_dma_unmap_resource,
1567 .get_merge_boundary = iommu_dma_get_merge_boundary,
1568 .opt_mapping_size = iommu_dma_opt_mapping_size,
1572 * The IOMMU core code allocates the default DMA domain, which the underlying
1573 * IOMMU driver needs to support via the dma-iommu layer.
1575 void iommu_setup_dma_ops(struct device *dev, u64 dma_base, u64 dma_limit)
1577 struct iommu_domain *domain = iommu_get_domain_for_dev(dev);
1583 * The IOMMU core code allocates the default DMA domain, which the
1584 * underlying IOMMU driver needs to support via the dma-iommu layer.
1586 if (iommu_is_dma_domain(domain)) {
1587 if (iommu_dma_init_domain(domain, dma_base, dma_limit, dev))
1589 dev->dma_ops = &iommu_dma_ops;
1594 pr_warn("Failed to set up IOMMU for device %s; retaining platform DMA ops\n",
1597 EXPORT_SYMBOL_GPL(iommu_setup_dma_ops);
1599 static struct iommu_dma_msi_page *iommu_dma_get_msi_page(struct device *dev,
1600 phys_addr_t msi_addr, struct iommu_domain *domain)
1602 struct iommu_dma_cookie *cookie = domain->iova_cookie;
1603 struct iommu_dma_msi_page *msi_page;
1605 int prot = IOMMU_WRITE | IOMMU_NOEXEC | IOMMU_MMIO;
1606 size_t size = cookie_msi_granule(cookie);
1608 msi_addr &= ~(phys_addr_t)(size - 1);
1609 list_for_each_entry(msi_page, &cookie->msi_page_list, list)
1610 if (msi_page->phys == msi_addr)
1613 msi_page = kzalloc(sizeof(*msi_page), GFP_KERNEL);
1617 iova = iommu_dma_alloc_iova(domain, size, dma_get_mask(dev), dev);
1621 if (iommu_map(domain, iova, msi_addr, size, prot))
1624 INIT_LIST_HEAD(&msi_page->list);
1625 msi_page->phys = msi_addr;
1626 msi_page->iova = iova;
1627 list_add(&msi_page->list, &cookie->msi_page_list);
1631 iommu_dma_free_iova(cookie, iova, size, NULL);
1638 * iommu_dma_prepare_msi() - Map the MSI page in the IOMMU domain
1639 * @desc: MSI descriptor, will store the MSI page
1640 * @msi_addr: MSI target address to be mapped
1642 * Return: 0 on success or negative error code if the mapping failed.
1644 int iommu_dma_prepare_msi(struct msi_desc *desc, phys_addr_t msi_addr)
1646 struct device *dev = msi_desc_to_dev(desc);
1647 struct iommu_domain *domain = iommu_get_domain_for_dev(dev);
1648 struct iommu_dma_msi_page *msi_page;
1649 static DEFINE_MUTEX(msi_prepare_lock); /* see below */
1651 if (!domain || !domain->iova_cookie) {
1652 desc->iommu_cookie = NULL;
1657 * In fact the whole prepare operation should already be serialised by
1658 * irq_domain_mutex further up the callchain, but that's pretty subtle
1659 * on its own, so consider this locking as failsafe documentation...
1661 mutex_lock(&msi_prepare_lock);
1662 msi_page = iommu_dma_get_msi_page(dev, msi_addr, domain);
1663 mutex_unlock(&msi_prepare_lock);
1665 msi_desc_set_iommu_cookie(desc, msi_page);
1673 * iommu_dma_compose_msi_msg() - Apply translation to an MSI message
1674 * @desc: MSI descriptor prepared by iommu_dma_prepare_msi()
1675 * @msg: MSI message containing target physical address
1677 void iommu_dma_compose_msi_msg(struct msi_desc *desc, struct msi_msg *msg)
1679 struct device *dev = msi_desc_to_dev(desc);
1680 const struct iommu_domain *domain = iommu_get_domain_for_dev(dev);
1681 const struct iommu_dma_msi_page *msi_page;
1683 msi_page = msi_desc_get_iommu_cookie(desc);
1685 if (!domain || !domain->iova_cookie || WARN_ON(!msi_page))
1688 msg->address_hi = upper_32_bits(msi_page->iova);
1689 msg->address_lo &= cookie_msi_granule(domain->iova_cookie) - 1;
1690 msg->address_lo += lower_32_bits(msi_page->iova);
1693 static int iommu_dma_init(void)
1695 if (is_kdump_kernel())
1696 static_branch_enable(&iommu_deferred_attach_enabled);
1698 return iova_cache_get();
1700 arch_initcall(iommu_dma_init);