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-iommu.h>
17 #include <linux/dma-map-ops.h>
18 #include <linux/gfp.h>
19 #include <linux/huge_mm.h>
20 #include <linux/iommu.h>
21 #include <linux/iova.h>
22 #include <linux/irq.h>
23 #include <linux/list_sort.h>
24 #include <linux/memremap.h>
26 #include <linux/mutex.h>
27 #include <linux/pci.h>
28 #include <linux/scatterlist.h>
29 #include <linux/spinlock.h>
30 #include <linux/swiotlb.h>
31 #include <linux/vmalloc.h>
33 struct iommu_dma_msi_page {
34 struct list_head list;
39 enum iommu_dma_cookie_type {
40 IOMMU_DMA_IOVA_COOKIE,
44 struct iommu_dma_cookie {
45 enum iommu_dma_cookie_type type;
47 /* Full allocator for IOMMU_DMA_IOVA_COOKIE */
49 struct iova_domain iovad;
51 struct iova_fq __percpu *fq; /* Flush queue */
52 /* Number of TLB flushes that have been started */
53 atomic64_t fq_flush_start_cnt;
54 /* Number of TLB flushes that have been finished */
55 atomic64_t fq_flush_finish_cnt;
56 /* Timer to regularily empty the flush queues */
57 struct timer_list fq_timer;
58 /* 1 when timer is active, 0 when not */
61 /* Trivial linear page allocator for IOMMU_DMA_MSI_COOKIE */
64 struct list_head msi_page_list;
66 /* Domain for flush queue callback; NULL if flush queue not in use */
67 struct iommu_domain *fq_domain;
71 static DEFINE_STATIC_KEY_FALSE(iommu_deferred_attach_enabled);
72 bool iommu_dma_forcedac __read_mostly;
74 static int __init iommu_dma_forcedac_setup(char *str)
76 int ret = kstrtobool(str, &iommu_dma_forcedac);
78 if (!ret && iommu_dma_forcedac)
79 pr_info("Forcing DAC for PCI devices\n");
82 early_param("iommu.forcedac", iommu_dma_forcedac_setup);
84 /* Number of entries per flush queue */
85 #define IOVA_FQ_SIZE 256
87 /* Timeout (in ms) after which entries are flushed from the queue */
88 #define IOVA_FQ_TIMEOUT 10
90 /* Flush queue entry for deferred flushing */
91 struct iova_fq_entry {
92 unsigned long iova_pfn;
94 struct list_head freelist;
95 u64 counter; /* Flush counter when this entry was added */
98 /* Per-CPU flush queue structure */
100 struct iova_fq_entry entries[IOVA_FQ_SIZE];
101 unsigned int head, tail;
105 #define fq_ring_for_each(i, fq) \
106 for ((i) = (fq)->head; (i) != (fq)->tail; (i) = ((i) + 1) % IOVA_FQ_SIZE)
108 static inline bool fq_full(struct iova_fq *fq)
110 assert_spin_locked(&fq->lock);
111 return (((fq->tail + 1) % IOVA_FQ_SIZE) == fq->head);
114 static inline unsigned int fq_ring_add(struct iova_fq *fq)
116 unsigned int idx = fq->tail;
118 assert_spin_locked(&fq->lock);
120 fq->tail = (idx + 1) % IOVA_FQ_SIZE;
125 static void fq_ring_free(struct iommu_dma_cookie *cookie, struct iova_fq *fq)
127 u64 counter = atomic64_read(&cookie->fq_flush_finish_cnt);
130 assert_spin_locked(&fq->lock);
132 fq_ring_for_each(idx, fq) {
134 if (fq->entries[idx].counter >= counter)
137 put_pages_list(&fq->entries[idx].freelist);
138 free_iova_fast(&cookie->iovad,
139 fq->entries[idx].iova_pfn,
140 fq->entries[idx].pages);
142 fq->head = (fq->head + 1) % IOVA_FQ_SIZE;
146 static void fq_flush_iotlb(struct iommu_dma_cookie *cookie)
148 atomic64_inc(&cookie->fq_flush_start_cnt);
149 cookie->fq_domain->ops->flush_iotlb_all(cookie->fq_domain);
150 atomic64_inc(&cookie->fq_flush_finish_cnt);
153 static void fq_flush_timeout(struct timer_list *t)
155 struct iommu_dma_cookie *cookie = from_timer(cookie, t, fq_timer);
158 atomic_set(&cookie->fq_timer_on, 0);
159 fq_flush_iotlb(cookie);
161 for_each_possible_cpu(cpu) {
165 fq = per_cpu_ptr(cookie->fq, cpu);
166 spin_lock_irqsave(&fq->lock, flags);
167 fq_ring_free(cookie, fq);
168 spin_unlock_irqrestore(&fq->lock, flags);
172 static void queue_iova(struct iommu_dma_cookie *cookie,
173 unsigned long pfn, unsigned long pages,
174 struct list_head *freelist)
181 * Order against the IOMMU driver's pagetable update from unmapping
182 * @pte, to guarantee that fq_flush_iotlb() observes that if called
183 * from a different CPU before we release the lock below. Full barrier
184 * so it also pairs with iommu_dma_init_fq() to avoid seeing partially
185 * written fq state here.
189 fq = raw_cpu_ptr(cookie->fq);
190 spin_lock_irqsave(&fq->lock, flags);
193 * First remove all entries from the flush queue that have already been
194 * flushed out on another CPU. This makes the fq_full() check below less
197 fq_ring_free(cookie, fq);
200 fq_flush_iotlb(cookie);
201 fq_ring_free(cookie, fq);
204 idx = fq_ring_add(fq);
206 fq->entries[idx].iova_pfn = pfn;
207 fq->entries[idx].pages = pages;
208 fq->entries[idx].counter = atomic64_read(&cookie->fq_flush_start_cnt);
209 list_splice(freelist, &fq->entries[idx].freelist);
211 spin_unlock_irqrestore(&fq->lock, flags);
213 /* Avoid false sharing as much as possible. */
214 if (!atomic_read(&cookie->fq_timer_on) &&
215 !atomic_xchg(&cookie->fq_timer_on, 1))
216 mod_timer(&cookie->fq_timer,
217 jiffies + msecs_to_jiffies(IOVA_FQ_TIMEOUT));
220 static void iommu_dma_free_fq(struct iommu_dma_cookie *cookie)
227 del_timer_sync(&cookie->fq_timer);
228 /* The IOVAs will be torn down separately, so just free our queued pages */
229 for_each_possible_cpu(cpu) {
230 struct iova_fq *fq = per_cpu_ptr(cookie->fq, cpu);
232 fq_ring_for_each(idx, fq)
233 put_pages_list(&fq->entries[idx].freelist);
236 free_percpu(cookie->fq);
239 /* sysfs updates are serialised by the mutex of the group owning @domain */
240 int iommu_dma_init_fq(struct iommu_domain *domain)
242 struct iommu_dma_cookie *cookie = domain->iova_cookie;
243 struct iova_fq __percpu *queue;
246 if (cookie->fq_domain)
249 atomic64_set(&cookie->fq_flush_start_cnt, 0);
250 atomic64_set(&cookie->fq_flush_finish_cnt, 0);
252 queue = alloc_percpu(struct iova_fq);
254 pr_warn("iova flush queue initialization failed\n");
258 for_each_possible_cpu(cpu) {
259 struct iova_fq *fq = per_cpu_ptr(queue, cpu);
264 spin_lock_init(&fq->lock);
266 for (i = 0; i < IOVA_FQ_SIZE; i++)
267 INIT_LIST_HEAD(&fq->entries[i].freelist);
272 timer_setup(&cookie->fq_timer, fq_flush_timeout, 0);
273 atomic_set(&cookie->fq_timer_on, 0);
275 * Prevent incomplete fq state being observable. Pairs with path from
276 * __iommu_dma_unmap() through iommu_dma_free_iova() to queue_iova()
279 WRITE_ONCE(cookie->fq_domain, domain);
283 static inline size_t cookie_msi_granule(struct iommu_dma_cookie *cookie)
285 if (cookie->type == IOMMU_DMA_IOVA_COOKIE)
286 return cookie->iovad.granule;
290 static struct iommu_dma_cookie *cookie_alloc(enum iommu_dma_cookie_type type)
292 struct iommu_dma_cookie *cookie;
294 cookie = kzalloc(sizeof(*cookie), GFP_KERNEL);
296 INIT_LIST_HEAD(&cookie->msi_page_list);
303 * iommu_get_dma_cookie - Acquire DMA-API resources for a domain
304 * @domain: IOMMU domain to prepare for DMA-API usage
306 int iommu_get_dma_cookie(struct iommu_domain *domain)
308 if (domain->iova_cookie)
311 domain->iova_cookie = cookie_alloc(IOMMU_DMA_IOVA_COOKIE);
312 if (!domain->iova_cookie)
315 mutex_init(&domain->iova_cookie->mutex);
320 * iommu_get_msi_cookie - Acquire just MSI remapping resources
321 * @domain: IOMMU domain to prepare
322 * @base: Start address of IOVA region for MSI mappings
324 * Users who manage their own IOVA allocation and do not want DMA API support,
325 * but would still like to take advantage of automatic MSI remapping, can use
326 * this to initialise their own domain appropriately. Users should reserve a
327 * contiguous IOVA region, starting at @base, large enough to accommodate the
328 * number of PAGE_SIZE mappings necessary to cover every MSI doorbell address
329 * used by the devices attached to @domain.
331 int iommu_get_msi_cookie(struct iommu_domain *domain, dma_addr_t base)
333 struct iommu_dma_cookie *cookie;
335 if (domain->type != IOMMU_DOMAIN_UNMANAGED)
338 if (domain->iova_cookie)
341 cookie = cookie_alloc(IOMMU_DMA_MSI_COOKIE);
345 cookie->msi_iova = base;
346 domain->iova_cookie = cookie;
349 EXPORT_SYMBOL(iommu_get_msi_cookie);
352 * iommu_put_dma_cookie - Release a domain's DMA mapping resources
353 * @domain: IOMMU domain previously prepared by iommu_get_dma_cookie() or
354 * iommu_get_msi_cookie()
356 void iommu_put_dma_cookie(struct iommu_domain *domain)
358 struct iommu_dma_cookie *cookie = domain->iova_cookie;
359 struct iommu_dma_msi_page *msi, *tmp;
364 if (cookie->type == IOMMU_DMA_IOVA_COOKIE && cookie->iovad.granule) {
365 iommu_dma_free_fq(cookie);
366 put_iova_domain(&cookie->iovad);
369 list_for_each_entry_safe(msi, tmp, &cookie->msi_page_list, list) {
370 list_del(&msi->list);
374 domain->iova_cookie = NULL;
378 * iommu_dma_get_resv_regions - Reserved region driver helper
379 * @dev: Device from iommu_get_resv_regions()
380 * @list: Reserved region list from iommu_get_resv_regions()
382 * IOMMU drivers can use this to implement their .get_resv_regions callback
383 * for general non-IOMMU-specific reservations. Currently, this covers GICv3
384 * ITS region reservation on ACPI based ARM platforms that may require HW MSI
387 void iommu_dma_get_resv_regions(struct device *dev, struct list_head *list)
390 if (!is_of_node(dev_iommu_fwspec_get(dev)->iommu_fwnode))
391 iort_iommu_get_resv_regions(dev, list);
394 EXPORT_SYMBOL(iommu_dma_get_resv_regions);
396 static int cookie_init_hw_msi_region(struct iommu_dma_cookie *cookie,
397 phys_addr_t start, phys_addr_t end)
399 struct iova_domain *iovad = &cookie->iovad;
400 struct iommu_dma_msi_page *msi_page;
403 start -= iova_offset(iovad, start);
404 num_pages = iova_align(iovad, end - start) >> iova_shift(iovad);
406 for (i = 0; i < num_pages; i++) {
407 msi_page = kmalloc(sizeof(*msi_page), GFP_KERNEL);
411 msi_page->phys = start;
412 msi_page->iova = start;
413 INIT_LIST_HEAD(&msi_page->list);
414 list_add(&msi_page->list, &cookie->msi_page_list);
415 start += iovad->granule;
421 static int iommu_dma_ranges_sort(void *priv, const struct list_head *a,
422 const struct list_head *b)
424 struct resource_entry *res_a = list_entry(a, typeof(*res_a), node);
425 struct resource_entry *res_b = list_entry(b, typeof(*res_b), node);
427 return res_a->res->start > res_b->res->start;
430 static int iova_reserve_pci_windows(struct pci_dev *dev,
431 struct iova_domain *iovad)
433 struct pci_host_bridge *bridge = pci_find_host_bridge(dev->bus);
434 struct resource_entry *window;
435 unsigned long lo, hi;
436 phys_addr_t start = 0, end;
438 resource_list_for_each_entry(window, &bridge->windows) {
439 if (resource_type(window->res) != IORESOURCE_MEM)
442 lo = iova_pfn(iovad, window->res->start - window->offset);
443 hi = iova_pfn(iovad, window->res->end - window->offset);
444 reserve_iova(iovad, lo, hi);
447 /* Get reserved DMA windows from host bridge */
448 list_sort(NULL, &bridge->dma_ranges, iommu_dma_ranges_sort);
449 resource_list_for_each_entry(window, &bridge->dma_ranges) {
450 end = window->res->start - window->offset;
453 lo = iova_pfn(iovad, start);
454 hi = iova_pfn(iovad, end);
455 reserve_iova(iovad, lo, hi);
456 } else if (end < start) {
457 /* DMA ranges should be non-overlapping */
459 "Failed to reserve IOVA [%pa-%pa]\n",
464 start = window->res->end - window->offset + 1;
465 /* If window is last entry */
466 if (window->node.next == &bridge->dma_ranges &&
467 end != ~(phys_addr_t)0) {
468 end = ~(phys_addr_t)0;
476 static int iova_reserve_iommu_regions(struct device *dev,
477 struct iommu_domain *domain)
479 struct iommu_dma_cookie *cookie = domain->iova_cookie;
480 struct iova_domain *iovad = &cookie->iovad;
481 struct iommu_resv_region *region;
482 LIST_HEAD(resv_regions);
485 if (dev_is_pci(dev)) {
486 ret = iova_reserve_pci_windows(to_pci_dev(dev), iovad);
491 iommu_get_resv_regions(dev, &resv_regions);
492 list_for_each_entry(region, &resv_regions, list) {
493 unsigned long lo, hi;
495 /* We ARE the software that manages these! */
496 if (region->type == IOMMU_RESV_SW_MSI)
499 lo = iova_pfn(iovad, region->start);
500 hi = iova_pfn(iovad, region->start + region->length - 1);
501 reserve_iova(iovad, lo, hi);
503 if (region->type == IOMMU_RESV_MSI)
504 ret = cookie_init_hw_msi_region(cookie, region->start,
505 region->start + region->length);
509 iommu_put_resv_regions(dev, &resv_regions);
514 static bool dev_is_untrusted(struct device *dev)
516 return dev_is_pci(dev) && to_pci_dev(dev)->untrusted;
519 static bool dev_use_swiotlb(struct device *dev)
521 return IS_ENABLED(CONFIG_SWIOTLB) && dev_is_untrusted(dev);
525 * iommu_dma_init_domain - Initialise a DMA mapping domain
526 * @domain: IOMMU domain previously prepared by iommu_get_dma_cookie()
527 * @base: IOVA at which the mappable address space starts
528 * @limit: Last address of the IOVA space
529 * @dev: Device the domain is being initialised for
531 * @base and @limit + 1 should be exact multiples of IOMMU page granularity to
532 * avoid rounding surprises. If necessary, we reserve the page at address 0
533 * to ensure it is an invalid IOVA. It is safe to reinitialise a domain, but
534 * any change which could make prior IOVAs invalid will fail.
536 static int iommu_dma_init_domain(struct iommu_domain *domain, dma_addr_t base,
537 dma_addr_t limit, struct device *dev)
539 struct iommu_dma_cookie *cookie = domain->iova_cookie;
540 unsigned long order, base_pfn;
541 struct iova_domain *iovad;
544 if (!cookie || cookie->type != IOMMU_DMA_IOVA_COOKIE)
547 iovad = &cookie->iovad;
549 /* Use the smallest supported page size for IOVA granularity */
550 order = __ffs(domain->pgsize_bitmap);
551 base_pfn = max_t(unsigned long, 1, base >> order);
553 /* Check the domain allows at least some access to the device... */
554 if (domain->geometry.force_aperture) {
555 if (base > domain->geometry.aperture_end ||
556 limit < domain->geometry.aperture_start) {
557 pr_warn("specified DMA range outside IOMMU capability\n");
560 /* ...then finally give it a kicking to make sure it fits */
561 base_pfn = max_t(unsigned long, base_pfn,
562 domain->geometry.aperture_start >> order);
565 /* start_pfn is always nonzero for an already-initialised domain */
566 mutex_lock(&cookie->mutex);
567 if (iovad->start_pfn) {
568 if (1UL << order != iovad->granule ||
569 base_pfn != iovad->start_pfn) {
570 pr_warn("Incompatible range for DMA domain\n");
579 init_iova_domain(iovad, 1UL << order, base_pfn);
580 ret = iova_domain_init_rcaches(iovad);
584 /* If the FQ fails we can simply fall back to strict mode */
585 if (domain->type == IOMMU_DOMAIN_DMA_FQ && iommu_dma_init_fq(domain))
586 domain->type = IOMMU_DOMAIN_DMA;
588 ret = iova_reserve_iommu_regions(dev, domain);
591 mutex_unlock(&cookie->mutex);
596 * dma_info_to_prot - Translate DMA API directions and attributes to IOMMU API
598 * @dir: Direction of DMA transfer
599 * @coherent: Is the DMA master cache-coherent?
600 * @attrs: DMA attributes for the mapping
602 * Return: corresponding IOMMU API page protection flags
604 static int dma_info_to_prot(enum dma_data_direction dir, bool coherent,
607 int prot = coherent ? IOMMU_CACHE : 0;
609 if (attrs & DMA_ATTR_PRIVILEGED)
613 case DMA_BIDIRECTIONAL:
614 return prot | IOMMU_READ | IOMMU_WRITE;
616 return prot | IOMMU_READ;
617 case DMA_FROM_DEVICE:
618 return prot | IOMMU_WRITE;
624 static dma_addr_t iommu_dma_alloc_iova(struct iommu_domain *domain,
625 size_t size, u64 dma_limit, struct device *dev)
627 struct iommu_dma_cookie *cookie = domain->iova_cookie;
628 struct iova_domain *iovad = &cookie->iovad;
629 unsigned long shift, iova_len, iova = 0;
631 if (cookie->type == IOMMU_DMA_MSI_COOKIE) {
632 cookie->msi_iova += size;
633 return cookie->msi_iova - size;
636 shift = iova_shift(iovad);
637 iova_len = size >> shift;
639 dma_limit = min_not_zero(dma_limit, dev->bus_dma_limit);
641 if (domain->geometry.force_aperture)
642 dma_limit = min(dma_limit, (u64)domain->geometry.aperture_end);
644 /* Try to get PCI devices a SAC address */
645 if (dma_limit > DMA_BIT_MASK(32) && !iommu_dma_forcedac && dev_is_pci(dev))
646 iova = alloc_iova_fast(iovad, iova_len,
647 DMA_BIT_MASK(32) >> shift, false);
650 iova = alloc_iova_fast(iovad, iova_len, dma_limit >> shift,
653 return (dma_addr_t)iova << shift;
656 static void iommu_dma_free_iova(struct iommu_dma_cookie *cookie,
657 dma_addr_t iova, size_t size, struct iommu_iotlb_gather *gather)
659 struct iova_domain *iovad = &cookie->iovad;
661 /* The MSI case is only ever cleaning up its most recent allocation */
662 if (cookie->type == IOMMU_DMA_MSI_COOKIE)
663 cookie->msi_iova -= size;
664 else if (gather && gather->queued)
665 queue_iova(cookie, iova_pfn(iovad, iova),
666 size >> iova_shift(iovad),
669 free_iova_fast(iovad, iova_pfn(iovad, iova),
670 size >> iova_shift(iovad));
673 static void __iommu_dma_unmap(struct device *dev, dma_addr_t dma_addr,
676 struct iommu_domain *domain = iommu_get_dma_domain(dev);
677 struct iommu_dma_cookie *cookie = domain->iova_cookie;
678 struct iova_domain *iovad = &cookie->iovad;
679 size_t iova_off = iova_offset(iovad, dma_addr);
680 struct iommu_iotlb_gather iotlb_gather;
683 dma_addr -= iova_off;
684 size = iova_align(iovad, size + iova_off);
685 iommu_iotlb_gather_init(&iotlb_gather);
686 iotlb_gather.queued = READ_ONCE(cookie->fq_domain);
688 unmapped = iommu_unmap_fast(domain, dma_addr, size, &iotlb_gather);
689 WARN_ON(unmapped != size);
691 if (!iotlb_gather.queued)
692 iommu_iotlb_sync(domain, &iotlb_gather);
693 iommu_dma_free_iova(cookie, dma_addr, size, &iotlb_gather);
696 static dma_addr_t __iommu_dma_map(struct device *dev, phys_addr_t phys,
697 size_t size, int prot, u64 dma_mask)
699 struct iommu_domain *domain = iommu_get_dma_domain(dev);
700 struct iommu_dma_cookie *cookie = domain->iova_cookie;
701 struct iova_domain *iovad = &cookie->iovad;
702 size_t iova_off = iova_offset(iovad, phys);
705 if (static_branch_unlikely(&iommu_deferred_attach_enabled) &&
706 iommu_deferred_attach(dev, domain))
707 return DMA_MAPPING_ERROR;
709 size = iova_align(iovad, size + iova_off);
711 iova = iommu_dma_alloc_iova(domain, size, dma_mask, dev);
713 return DMA_MAPPING_ERROR;
715 if (iommu_map_atomic(domain, iova, phys - iova_off, size, prot)) {
716 iommu_dma_free_iova(cookie, iova, size, NULL);
717 return DMA_MAPPING_ERROR;
719 return iova + iova_off;
722 static void __iommu_dma_free_pages(struct page **pages, int count)
725 __free_page(pages[count]);
729 static struct page **__iommu_dma_alloc_pages(struct device *dev,
730 unsigned int count, unsigned long order_mask, gfp_t gfp)
733 unsigned int i = 0, nid = dev_to_node(dev);
735 order_mask &= (2U << MAX_ORDER) - 1;
739 pages = kvcalloc(count, sizeof(*pages), GFP_KERNEL);
743 /* IOMMU can map any pages, so himem can also be used here */
744 gfp |= __GFP_NOWARN | __GFP_HIGHMEM;
746 /* It makes no sense to muck about with huge pages */
750 struct page *page = NULL;
751 unsigned int order_size;
754 * Higher-order allocations are a convenience rather
755 * than a necessity, hence using __GFP_NORETRY until
756 * falling back to minimum-order allocations.
758 for (order_mask &= (2U << __fls(count)) - 1;
759 order_mask; order_mask &= ~order_size) {
760 unsigned int order = __fls(order_mask);
761 gfp_t alloc_flags = gfp;
763 order_size = 1U << order;
764 if (order_mask > order_size)
765 alloc_flags |= __GFP_NORETRY;
766 page = alloc_pages_node(nid, alloc_flags, order);
770 split_page(page, order);
774 __iommu_dma_free_pages(pages, i);
785 * If size is less than PAGE_SIZE, then a full CPU page will be allocated,
786 * but an IOMMU which supports smaller pages might not map the whole thing.
788 static struct page **__iommu_dma_alloc_noncontiguous(struct device *dev,
789 size_t size, struct sg_table *sgt, gfp_t gfp, pgprot_t prot,
792 struct iommu_domain *domain = iommu_get_dma_domain(dev);
793 struct iommu_dma_cookie *cookie = domain->iova_cookie;
794 struct iova_domain *iovad = &cookie->iovad;
795 bool coherent = dev_is_dma_coherent(dev);
796 int ioprot = dma_info_to_prot(DMA_BIDIRECTIONAL, coherent, attrs);
797 unsigned int count, min_size, alloc_sizes = domain->pgsize_bitmap;
802 if (static_branch_unlikely(&iommu_deferred_attach_enabled) &&
803 iommu_deferred_attach(dev, domain))
806 min_size = alloc_sizes & -alloc_sizes;
807 if (min_size < PAGE_SIZE) {
808 min_size = PAGE_SIZE;
809 alloc_sizes |= PAGE_SIZE;
811 size = ALIGN(size, min_size);
813 if (attrs & DMA_ATTR_ALLOC_SINGLE_PAGES)
814 alloc_sizes = min_size;
816 count = PAGE_ALIGN(size) >> PAGE_SHIFT;
817 pages = __iommu_dma_alloc_pages(dev, count, alloc_sizes >> PAGE_SHIFT,
822 size = iova_align(iovad, size);
823 iova = iommu_dma_alloc_iova(domain, size, dev->coherent_dma_mask, dev);
827 if (sg_alloc_table_from_pages(sgt, pages, count, 0, size, GFP_KERNEL))
830 if (!(ioprot & IOMMU_CACHE)) {
831 struct scatterlist *sg;
834 for_each_sg(sgt->sgl, sg, sgt->orig_nents, i)
835 arch_dma_prep_coherent(sg_page(sg), sg->length);
838 ret = iommu_map_sg_atomic(domain, iova, sgt->sgl, sgt->orig_nents, ioprot);
839 if (ret < 0 || ret < size)
842 sgt->sgl->dma_address = iova;
843 sgt->sgl->dma_length = size;
849 iommu_dma_free_iova(cookie, iova, size, NULL);
851 __iommu_dma_free_pages(pages, count);
855 static void *iommu_dma_alloc_remap(struct device *dev, size_t size,
856 dma_addr_t *dma_handle, gfp_t gfp, pgprot_t prot,
863 pages = __iommu_dma_alloc_noncontiguous(dev, size, &sgt, gfp, prot,
867 *dma_handle = sgt.sgl->dma_address;
869 vaddr = dma_common_pages_remap(pages, size, prot,
870 __builtin_return_address(0));
876 __iommu_dma_unmap(dev, *dma_handle, size);
877 __iommu_dma_free_pages(pages, PAGE_ALIGN(size) >> PAGE_SHIFT);
881 static struct sg_table *iommu_dma_alloc_noncontiguous(struct device *dev,
882 size_t size, enum dma_data_direction dir, gfp_t gfp,
885 struct dma_sgt_handle *sh;
887 sh = kmalloc(sizeof(*sh), gfp);
891 sh->pages = __iommu_dma_alloc_noncontiguous(dev, size, &sh->sgt, gfp,
900 static void iommu_dma_free_noncontiguous(struct device *dev, size_t size,
901 struct sg_table *sgt, enum dma_data_direction dir)
903 struct dma_sgt_handle *sh = sgt_handle(sgt);
905 __iommu_dma_unmap(dev, sgt->sgl->dma_address, size);
906 __iommu_dma_free_pages(sh->pages, PAGE_ALIGN(size) >> PAGE_SHIFT);
907 sg_free_table(&sh->sgt);
911 static void iommu_dma_sync_single_for_cpu(struct device *dev,
912 dma_addr_t dma_handle, size_t size, enum dma_data_direction dir)
916 if (dev_is_dma_coherent(dev) && !dev_use_swiotlb(dev))
919 phys = iommu_iova_to_phys(iommu_get_dma_domain(dev), dma_handle);
920 if (!dev_is_dma_coherent(dev))
921 arch_sync_dma_for_cpu(phys, size, dir);
923 if (is_swiotlb_buffer(dev, phys))
924 swiotlb_sync_single_for_cpu(dev, phys, size, dir);
927 static void iommu_dma_sync_single_for_device(struct device *dev,
928 dma_addr_t dma_handle, size_t size, enum dma_data_direction dir)
932 if (dev_is_dma_coherent(dev) && !dev_use_swiotlb(dev))
935 phys = iommu_iova_to_phys(iommu_get_dma_domain(dev), dma_handle);
936 if (is_swiotlb_buffer(dev, phys))
937 swiotlb_sync_single_for_device(dev, phys, size, dir);
939 if (!dev_is_dma_coherent(dev))
940 arch_sync_dma_for_device(phys, size, dir);
943 static void iommu_dma_sync_sg_for_cpu(struct device *dev,
944 struct scatterlist *sgl, int nelems,
945 enum dma_data_direction dir)
947 struct scatterlist *sg;
950 if (dev_use_swiotlb(dev))
951 for_each_sg(sgl, sg, nelems, i)
952 iommu_dma_sync_single_for_cpu(dev, sg_dma_address(sg),
954 else if (!dev_is_dma_coherent(dev))
955 for_each_sg(sgl, sg, nelems, i)
956 arch_sync_dma_for_cpu(sg_phys(sg), sg->length, dir);
959 static void iommu_dma_sync_sg_for_device(struct device *dev,
960 struct scatterlist *sgl, int nelems,
961 enum dma_data_direction dir)
963 struct scatterlist *sg;
966 if (dev_use_swiotlb(dev))
967 for_each_sg(sgl, sg, nelems, i)
968 iommu_dma_sync_single_for_device(dev,
971 else if (!dev_is_dma_coherent(dev))
972 for_each_sg(sgl, sg, nelems, i)
973 arch_sync_dma_for_device(sg_phys(sg), sg->length, dir);
976 static dma_addr_t iommu_dma_map_page(struct device *dev, struct page *page,
977 unsigned long offset, size_t size, enum dma_data_direction dir,
980 phys_addr_t phys = page_to_phys(page) + offset;
981 bool coherent = dev_is_dma_coherent(dev);
982 int prot = dma_info_to_prot(dir, coherent, attrs);
983 struct iommu_domain *domain = iommu_get_dma_domain(dev);
984 struct iommu_dma_cookie *cookie = domain->iova_cookie;
985 struct iova_domain *iovad = &cookie->iovad;
986 dma_addr_t iova, dma_mask = dma_get_mask(dev);
989 * If both the physical buffer start address and size are
990 * page aligned, we don't need to use a bounce page.
992 if (dev_use_swiotlb(dev) && iova_offset(iovad, phys | size)) {
994 size_t padding_size, aligned_size;
996 if (!is_swiotlb_active(dev)) {
997 dev_warn_once(dev, "DMA bounce buffers are inactive, unable to map unaligned transaction.\n");
998 return DMA_MAPPING_ERROR;
1001 aligned_size = iova_align(iovad, size);
1002 phys = swiotlb_tbl_map_single(dev, phys, size, aligned_size,
1003 iova_mask(iovad), dir, attrs);
1005 if (phys == DMA_MAPPING_ERROR)
1006 return DMA_MAPPING_ERROR;
1008 /* Cleanup the padding area. */
1009 padding_start = phys_to_virt(phys);
1010 padding_size = aligned_size;
1012 if (!(attrs & DMA_ATTR_SKIP_CPU_SYNC) &&
1013 (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL)) {
1014 padding_start += size;
1015 padding_size -= size;
1018 memset(padding_start, 0, padding_size);
1021 if (!coherent && !(attrs & DMA_ATTR_SKIP_CPU_SYNC))
1022 arch_sync_dma_for_device(phys, size, dir);
1024 iova = __iommu_dma_map(dev, phys, size, prot, dma_mask);
1025 if (iova == DMA_MAPPING_ERROR && is_swiotlb_buffer(dev, phys))
1026 swiotlb_tbl_unmap_single(dev, phys, size, dir, attrs);
1030 static void iommu_dma_unmap_page(struct device *dev, dma_addr_t dma_handle,
1031 size_t size, enum dma_data_direction dir, unsigned long attrs)
1033 struct iommu_domain *domain = iommu_get_dma_domain(dev);
1036 phys = iommu_iova_to_phys(domain, dma_handle);
1040 if (!(attrs & DMA_ATTR_SKIP_CPU_SYNC) && !dev_is_dma_coherent(dev))
1041 arch_sync_dma_for_cpu(phys, size, dir);
1043 __iommu_dma_unmap(dev, dma_handle, size);
1045 if (unlikely(is_swiotlb_buffer(dev, phys)))
1046 swiotlb_tbl_unmap_single(dev, phys, size, dir, attrs);
1050 * Prepare a successfully-mapped scatterlist to give back to the caller.
1052 * At this point the segments are already laid out by iommu_dma_map_sg() to
1053 * avoid individually crossing any boundaries, so we merely need to check a
1054 * segment's start address to avoid concatenating across one.
1056 static int __finalise_sg(struct device *dev, struct scatterlist *sg, int nents,
1057 dma_addr_t dma_addr)
1059 struct scatterlist *s, *cur = sg;
1060 unsigned long seg_mask = dma_get_seg_boundary(dev);
1061 unsigned int cur_len = 0, max_len = dma_get_max_seg_size(dev);
1064 for_each_sg(sg, s, nents, i) {
1065 /* Restore this segment's original unaligned fields first */
1066 dma_addr_t s_dma_addr = sg_dma_address(s);
1067 unsigned int s_iova_off = sg_dma_address(s);
1068 unsigned int s_length = sg_dma_len(s);
1069 unsigned int s_iova_len = s->length;
1071 sg_dma_address(s) = DMA_MAPPING_ERROR;
1074 if (sg_is_dma_bus_address(s)) {
1078 sg_dma_unmark_bus_address(s);
1079 sg_dma_address(cur) = s_dma_addr;
1080 sg_dma_len(cur) = s_length;
1081 sg_dma_mark_bus_address(cur);
1087 s->offset += s_iova_off;
1088 s->length = s_length;
1091 * Now fill in the real DMA data. If...
1092 * - there is a valid output segment to append to
1093 * - and this segment starts on an IOVA page boundary
1094 * - but doesn't fall at a segment boundary
1095 * - and wouldn't make the resulting output segment too long
1097 if (cur_len && !s_iova_off && (dma_addr & seg_mask) &&
1098 (max_len - cur_len >= s_length)) {
1099 /* ...then concatenate it with the previous one */
1100 cur_len += s_length;
1102 /* Otherwise start the next output segment */
1108 sg_dma_address(cur) = dma_addr + s_iova_off;
1111 sg_dma_len(cur) = cur_len;
1112 dma_addr += s_iova_len;
1114 if (s_length + s_iova_off < s_iova_len)
1121 * If mapping failed, then just restore the original list,
1122 * but making sure the DMA fields are invalidated.
1124 static void __invalidate_sg(struct scatterlist *sg, int nents)
1126 struct scatterlist *s;
1129 for_each_sg(sg, s, nents, i) {
1130 if (sg_is_dma_bus_address(s)) {
1131 sg_dma_unmark_bus_address(s);
1133 if (sg_dma_address(s) != DMA_MAPPING_ERROR)
1134 s->offset += sg_dma_address(s);
1136 s->length = sg_dma_len(s);
1138 sg_dma_address(s) = DMA_MAPPING_ERROR;
1143 static void iommu_dma_unmap_sg_swiotlb(struct device *dev, struct scatterlist *sg,
1144 int nents, enum dma_data_direction dir, unsigned long attrs)
1146 struct scatterlist *s;
1149 for_each_sg(sg, s, nents, i)
1150 iommu_dma_unmap_page(dev, sg_dma_address(s),
1151 sg_dma_len(s), dir, attrs);
1154 static int iommu_dma_map_sg_swiotlb(struct device *dev, struct scatterlist *sg,
1155 int nents, enum dma_data_direction dir, unsigned long attrs)
1157 struct scatterlist *s;
1160 for_each_sg(sg, s, nents, i) {
1161 sg_dma_address(s) = iommu_dma_map_page(dev, sg_page(s),
1162 s->offset, s->length, dir, attrs);
1163 if (sg_dma_address(s) == DMA_MAPPING_ERROR)
1165 sg_dma_len(s) = s->length;
1171 iommu_dma_unmap_sg_swiotlb(dev, sg, i, dir, attrs | DMA_ATTR_SKIP_CPU_SYNC);
1176 * The DMA API client is passing in a scatterlist which could describe
1177 * any old buffer layout, but the IOMMU API requires everything to be
1178 * aligned to IOMMU pages. Hence the need for this complicated bit of
1179 * impedance-matching, to be able to hand off a suitably-aligned list,
1180 * but still preserve the original offsets and sizes for the caller.
1182 static int iommu_dma_map_sg(struct device *dev, struct scatterlist *sg,
1183 int nents, enum dma_data_direction dir, unsigned long attrs)
1185 struct iommu_domain *domain = iommu_get_dma_domain(dev);
1186 struct iommu_dma_cookie *cookie = domain->iova_cookie;
1187 struct iova_domain *iovad = &cookie->iovad;
1188 struct scatterlist *s, *prev = NULL;
1189 int prot = dma_info_to_prot(dir, dev_is_dma_coherent(dev), attrs);
1190 struct pci_p2pdma_map_state p2pdma_state = {};
1191 enum pci_p2pdma_map_type map;
1193 size_t iova_len = 0;
1194 unsigned long mask = dma_get_seg_boundary(dev);
1198 if (static_branch_unlikely(&iommu_deferred_attach_enabled)) {
1199 ret = iommu_deferred_attach(dev, domain);
1204 if (dev_use_swiotlb(dev))
1205 return iommu_dma_map_sg_swiotlb(dev, sg, nents, dir, attrs);
1207 if (!(attrs & DMA_ATTR_SKIP_CPU_SYNC))
1208 iommu_dma_sync_sg_for_device(dev, sg, nents, dir);
1211 * Work out how much IOVA space we need, and align the segments to
1212 * IOVA granules for the IOMMU driver to handle. With some clever
1213 * trickery we can modify the list in-place, but reversibly, by
1214 * stashing the unaligned parts in the as-yet-unused DMA fields.
1216 for_each_sg(sg, s, nents, i) {
1217 size_t s_iova_off = iova_offset(iovad, s->offset);
1218 size_t s_length = s->length;
1219 size_t pad_len = (mask - iova_len + 1) & mask;
1221 if (is_pci_p2pdma_page(sg_page(s))) {
1222 map = pci_p2pdma_map_segment(&p2pdma_state, dev, s);
1224 case PCI_P2PDMA_MAP_BUS_ADDR:
1226 * iommu_map_sg() will skip this segment as
1227 * it is marked as a bus address,
1228 * __finalise_sg() will copy the dma address
1229 * into the output segment.
1232 case PCI_P2PDMA_MAP_THRU_HOST_BRIDGE:
1234 * Mapping through host bridge should be
1235 * mapped with regular IOVAs, thus we
1236 * do nothing here and continue below.
1241 goto out_restore_sg;
1245 sg_dma_address(s) = s_iova_off;
1246 sg_dma_len(s) = s_length;
1247 s->offset -= s_iova_off;
1248 s_length = iova_align(iovad, s_length + s_iova_off);
1249 s->length = s_length;
1252 * Due to the alignment of our single IOVA allocation, we can
1253 * depend on these assumptions about the segment boundary mask:
1254 * - If mask size >= IOVA size, then the IOVA range cannot
1255 * possibly fall across a boundary, so we don't care.
1256 * - If mask size < IOVA size, then the IOVA range must start
1257 * exactly on a boundary, therefore we can lay things out
1258 * based purely on segment lengths without needing to know
1259 * the actual addresses beforehand.
1260 * - The mask must be a power of 2, so pad_len == 0 if
1261 * iova_len == 0, thus we cannot dereference prev the first
1262 * time through here (i.e. before it has a meaningful value).
1264 if (pad_len && pad_len < s_length - 1) {
1265 prev->length += pad_len;
1266 iova_len += pad_len;
1269 iova_len += s_length;
1274 return __finalise_sg(dev, sg, nents, 0);
1276 iova = iommu_dma_alloc_iova(domain, iova_len, dma_get_mask(dev), dev);
1279 goto out_restore_sg;
1283 * We'll leave any physical concatenation to the IOMMU driver's
1284 * implementation - it knows better than we do.
1286 ret = iommu_map_sg_atomic(domain, iova, sg, nents, prot);
1287 if (ret < 0 || ret < iova_len)
1290 return __finalise_sg(dev, sg, nents, iova);
1293 iommu_dma_free_iova(cookie, iova, iova_len, NULL);
1295 __invalidate_sg(sg, nents);
1297 if (ret != -ENOMEM && ret != -EREMOTEIO)
1302 static void iommu_dma_unmap_sg(struct device *dev, struct scatterlist *sg,
1303 int nents, enum dma_data_direction dir, unsigned long attrs)
1305 dma_addr_t end = 0, start;
1306 struct scatterlist *tmp;
1309 if (dev_use_swiotlb(dev)) {
1310 iommu_dma_unmap_sg_swiotlb(dev, sg, nents, dir, attrs);
1314 if (!(attrs & DMA_ATTR_SKIP_CPU_SYNC))
1315 iommu_dma_sync_sg_for_cpu(dev, sg, nents, dir);
1318 * The scatterlist segments are mapped into a single
1319 * contiguous IOVA allocation, the start and end points
1320 * just have to be determined.
1322 for_each_sg(sg, tmp, nents, i) {
1323 if (sg_is_dma_bus_address(tmp)) {
1324 sg_dma_unmark_bus_address(tmp);
1328 if (sg_dma_len(tmp) == 0)
1331 start = sg_dma_address(tmp);
1336 for_each_sg(tmp, tmp, nents, i) {
1337 if (sg_is_dma_bus_address(tmp)) {
1338 sg_dma_unmark_bus_address(tmp);
1342 if (sg_dma_len(tmp) == 0)
1345 end = sg_dma_address(tmp) + sg_dma_len(tmp);
1349 __iommu_dma_unmap(dev, start, end - start);
1352 static dma_addr_t iommu_dma_map_resource(struct device *dev, phys_addr_t phys,
1353 size_t size, enum dma_data_direction dir, unsigned long attrs)
1355 return __iommu_dma_map(dev, phys, size,
1356 dma_info_to_prot(dir, false, attrs) | IOMMU_MMIO,
1360 static void iommu_dma_unmap_resource(struct device *dev, dma_addr_t handle,
1361 size_t size, enum dma_data_direction dir, unsigned long attrs)
1363 __iommu_dma_unmap(dev, handle, size);
1366 static void __iommu_dma_free(struct device *dev, size_t size, void *cpu_addr)
1368 size_t alloc_size = PAGE_ALIGN(size);
1369 int count = alloc_size >> PAGE_SHIFT;
1370 struct page *page = NULL, **pages = NULL;
1372 /* Non-coherent atomic allocation? Easy */
1373 if (IS_ENABLED(CONFIG_DMA_DIRECT_REMAP) &&
1374 dma_free_from_pool(dev, cpu_addr, alloc_size))
1377 if (is_vmalloc_addr(cpu_addr)) {
1379 * If it the address is remapped, then it's either non-coherent
1380 * or highmem CMA, or an iommu_dma_alloc_remap() construction.
1382 pages = dma_common_find_pages(cpu_addr);
1384 page = vmalloc_to_page(cpu_addr);
1385 dma_common_free_remap(cpu_addr, alloc_size);
1387 /* Lowmem means a coherent atomic or CMA allocation */
1388 page = virt_to_page(cpu_addr);
1392 __iommu_dma_free_pages(pages, count);
1394 dma_free_contiguous(dev, page, alloc_size);
1397 static void iommu_dma_free(struct device *dev, size_t size, void *cpu_addr,
1398 dma_addr_t handle, unsigned long attrs)
1400 __iommu_dma_unmap(dev, handle, size);
1401 __iommu_dma_free(dev, size, cpu_addr);
1404 static void *iommu_dma_alloc_pages(struct device *dev, size_t size,
1405 struct page **pagep, gfp_t gfp, unsigned long attrs)
1407 bool coherent = dev_is_dma_coherent(dev);
1408 size_t alloc_size = PAGE_ALIGN(size);
1409 int node = dev_to_node(dev);
1410 struct page *page = NULL;
1413 page = dma_alloc_contiguous(dev, alloc_size, gfp);
1415 page = alloc_pages_node(node, gfp, get_order(alloc_size));
1419 if (!coherent || PageHighMem(page)) {
1420 pgprot_t prot = dma_pgprot(dev, PAGE_KERNEL, attrs);
1422 cpu_addr = dma_common_contiguous_remap(page, alloc_size,
1423 prot, __builtin_return_address(0));
1425 goto out_free_pages;
1428 arch_dma_prep_coherent(page, size);
1430 cpu_addr = page_address(page);
1434 memset(cpu_addr, 0, alloc_size);
1437 dma_free_contiguous(dev, page, alloc_size);
1441 static void *iommu_dma_alloc(struct device *dev, size_t size,
1442 dma_addr_t *handle, gfp_t gfp, unsigned long attrs)
1444 bool coherent = dev_is_dma_coherent(dev);
1445 int ioprot = dma_info_to_prot(DMA_BIDIRECTIONAL, coherent, attrs);
1446 struct page *page = NULL;
1451 if (gfpflags_allow_blocking(gfp) &&
1452 !(attrs & DMA_ATTR_FORCE_CONTIGUOUS)) {
1453 return iommu_dma_alloc_remap(dev, size, handle, gfp,
1454 dma_pgprot(dev, PAGE_KERNEL, attrs), attrs);
1457 if (IS_ENABLED(CONFIG_DMA_DIRECT_REMAP) &&
1458 !gfpflags_allow_blocking(gfp) && !coherent)
1459 page = dma_alloc_from_pool(dev, PAGE_ALIGN(size), &cpu_addr,
1462 cpu_addr = iommu_dma_alloc_pages(dev, size, &page, gfp, attrs);
1466 *handle = __iommu_dma_map(dev, page_to_phys(page), size, ioprot,
1467 dev->coherent_dma_mask);
1468 if (*handle == DMA_MAPPING_ERROR) {
1469 __iommu_dma_free(dev, size, cpu_addr);
1476 static int iommu_dma_mmap(struct device *dev, struct vm_area_struct *vma,
1477 void *cpu_addr, dma_addr_t dma_addr, size_t size,
1478 unsigned long attrs)
1480 unsigned long nr_pages = PAGE_ALIGN(size) >> PAGE_SHIFT;
1481 unsigned long pfn, off = vma->vm_pgoff;
1484 vma->vm_page_prot = dma_pgprot(dev, vma->vm_page_prot, attrs);
1486 if (dma_mmap_from_dev_coherent(dev, vma, cpu_addr, size, &ret))
1489 if (off >= nr_pages || vma_pages(vma) > nr_pages - off)
1492 if (is_vmalloc_addr(cpu_addr)) {
1493 struct page **pages = dma_common_find_pages(cpu_addr);
1496 return vm_map_pages(vma, pages, nr_pages);
1497 pfn = vmalloc_to_pfn(cpu_addr);
1499 pfn = page_to_pfn(virt_to_page(cpu_addr));
1502 return remap_pfn_range(vma, vma->vm_start, pfn + off,
1503 vma->vm_end - vma->vm_start,
1507 static int iommu_dma_get_sgtable(struct device *dev, struct sg_table *sgt,
1508 void *cpu_addr, dma_addr_t dma_addr, size_t size,
1509 unsigned long attrs)
1514 if (is_vmalloc_addr(cpu_addr)) {
1515 struct page **pages = dma_common_find_pages(cpu_addr);
1518 return sg_alloc_table_from_pages(sgt, pages,
1519 PAGE_ALIGN(size) >> PAGE_SHIFT,
1520 0, size, GFP_KERNEL);
1523 page = vmalloc_to_page(cpu_addr);
1525 page = virt_to_page(cpu_addr);
1528 ret = sg_alloc_table(sgt, 1, GFP_KERNEL);
1530 sg_set_page(sgt->sgl, page, PAGE_ALIGN(size), 0);
1534 static unsigned long iommu_dma_get_merge_boundary(struct device *dev)
1536 struct iommu_domain *domain = iommu_get_dma_domain(dev);
1538 return (1UL << __ffs(domain->pgsize_bitmap)) - 1;
1541 static size_t iommu_dma_opt_mapping_size(void)
1543 return iova_rcache_range();
1546 static const struct dma_map_ops iommu_dma_ops = {
1547 .flags = DMA_F_PCI_P2PDMA_SUPPORTED,
1548 .alloc = iommu_dma_alloc,
1549 .free = iommu_dma_free,
1550 .alloc_pages = dma_common_alloc_pages,
1551 .free_pages = dma_common_free_pages,
1552 .alloc_noncontiguous = iommu_dma_alloc_noncontiguous,
1553 .free_noncontiguous = iommu_dma_free_noncontiguous,
1554 .mmap = iommu_dma_mmap,
1555 .get_sgtable = iommu_dma_get_sgtable,
1556 .map_page = iommu_dma_map_page,
1557 .unmap_page = iommu_dma_unmap_page,
1558 .map_sg = iommu_dma_map_sg,
1559 .unmap_sg = iommu_dma_unmap_sg,
1560 .sync_single_for_cpu = iommu_dma_sync_single_for_cpu,
1561 .sync_single_for_device = iommu_dma_sync_single_for_device,
1562 .sync_sg_for_cpu = iommu_dma_sync_sg_for_cpu,
1563 .sync_sg_for_device = iommu_dma_sync_sg_for_device,
1564 .map_resource = iommu_dma_map_resource,
1565 .unmap_resource = iommu_dma_unmap_resource,
1566 .get_merge_boundary = iommu_dma_get_merge_boundary,
1567 .opt_mapping_size = iommu_dma_opt_mapping_size,
1571 * The IOMMU core code allocates the default DMA domain, which the underlying
1572 * IOMMU driver needs to support via the dma-iommu layer.
1574 void iommu_setup_dma_ops(struct device *dev, u64 dma_base, u64 dma_limit)
1576 struct iommu_domain *domain = iommu_get_domain_for_dev(dev);
1582 * The IOMMU core code allocates the default DMA domain, which the
1583 * underlying IOMMU driver needs to support via the dma-iommu layer.
1585 if (iommu_is_dma_domain(domain)) {
1586 if (iommu_dma_init_domain(domain, dma_base, dma_limit, dev))
1588 dev->dma_ops = &iommu_dma_ops;
1593 pr_warn("Failed to set up IOMMU for device %s; retaining platform DMA ops\n",
1596 EXPORT_SYMBOL_GPL(iommu_setup_dma_ops);
1598 static struct iommu_dma_msi_page *iommu_dma_get_msi_page(struct device *dev,
1599 phys_addr_t msi_addr, struct iommu_domain *domain)
1601 struct iommu_dma_cookie *cookie = domain->iova_cookie;
1602 struct iommu_dma_msi_page *msi_page;
1604 int prot = IOMMU_WRITE | IOMMU_NOEXEC | IOMMU_MMIO;
1605 size_t size = cookie_msi_granule(cookie);
1607 msi_addr &= ~(phys_addr_t)(size - 1);
1608 list_for_each_entry(msi_page, &cookie->msi_page_list, list)
1609 if (msi_page->phys == msi_addr)
1612 msi_page = kzalloc(sizeof(*msi_page), GFP_KERNEL);
1616 iova = iommu_dma_alloc_iova(domain, size, dma_get_mask(dev), dev);
1620 if (iommu_map(domain, iova, msi_addr, size, prot))
1623 INIT_LIST_HEAD(&msi_page->list);
1624 msi_page->phys = msi_addr;
1625 msi_page->iova = iova;
1626 list_add(&msi_page->list, &cookie->msi_page_list);
1630 iommu_dma_free_iova(cookie, iova, size, NULL);
1636 int iommu_dma_prepare_msi(struct msi_desc *desc, phys_addr_t msi_addr)
1638 struct device *dev = msi_desc_to_dev(desc);
1639 struct iommu_domain *domain = iommu_get_domain_for_dev(dev);
1640 struct iommu_dma_msi_page *msi_page;
1641 static DEFINE_MUTEX(msi_prepare_lock); /* see below */
1643 if (!domain || !domain->iova_cookie) {
1644 desc->iommu_cookie = NULL;
1649 * In fact the whole prepare operation should already be serialised by
1650 * irq_domain_mutex further up the callchain, but that's pretty subtle
1651 * on its own, so consider this locking as failsafe documentation...
1653 mutex_lock(&msi_prepare_lock);
1654 msi_page = iommu_dma_get_msi_page(dev, msi_addr, domain);
1655 mutex_unlock(&msi_prepare_lock);
1657 msi_desc_set_iommu_cookie(desc, msi_page);
1664 void iommu_dma_compose_msi_msg(struct msi_desc *desc,
1665 struct msi_msg *msg)
1667 struct device *dev = msi_desc_to_dev(desc);
1668 const struct iommu_domain *domain = iommu_get_domain_for_dev(dev);
1669 const struct iommu_dma_msi_page *msi_page;
1671 msi_page = msi_desc_get_iommu_cookie(desc);
1673 if (!domain || !domain->iova_cookie || WARN_ON(!msi_page))
1676 msg->address_hi = upper_32_bits(msi_page->iova);
1677 msg->address_lo &= cookie_msi_granule(domain->iova_cookie) - 1;
1678 msg->address_lo += lower_32_bits(msi_page->iova);
1681 static int iommu_dma_init(void)
1683 if (is_kdump_kernel())
1684 static_branch_enable(&iommu_deferred_attach_enabled);
1686 return iova_cache_get();
1688 arch_initcall(iommu_dma_init);