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/of_iommu.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>
32 #include <trace/events/swiotlb.h>
34 #include "dma-iommu.h"
36 struct iommu_dma_msi_page {
37 struct list_head list;
42 enum iommu_dma_cookie_type {
43 IOMMU_DMA_IOVA_COOKIE,
47 enum iommu_dma_queue_type {
48 IOMMU_DMA_OPTS_PER_CPU_QUEUE,
49 IOMMU_DMA_OPTS_SINGLE_QUEUE,
52 struct iommu_dma_options {
53 enum iommu_dma_queue_type qt;
55 unsigned int fq_timeout;
58 struct iommu_dma_cookie {
59 enum iommu_dma_cookie_type type;
61 /* Full allocator for IOMMU_DMA_IOVA_COOKIE */
63 struct iova_domain iovad;
66 struct iova_fq *single_fq;
67 struct iova_fq __percpu *percpu_fq;
69 /* Number of TLB flushes that have been started */
70 atomic64_t fq_flush_start_cnt;
71 /* Number of TLB flushes that have been finished */
72 atomic64_t fq_flush_finish_cnt;
73 /* Timer to regularily empty the flush queues */
74 struct timer_list fq_timer;
75 /* 1 when timer is active, 0 when not */
78 /* Trivial linear page allocator for IOMMU_DMA_MSI_COOKIE */
81 struct list_head msi_page_list;
83 /* Domain for flush queue callback; NULL if flush queue not in use */
84 struct iommu_domain *fq_domain;
85 /* Options for dma-iommu use */
86 struct iommu_dma_options options;
90 static DEFINE_STATIC_KEY_FALSE(iommu_deferred_attach_enabled);
91 bool iommu_dma_forcedac __read_mostly;
93 static int __init iommu_dma_forcedac_setup(char *str)
95 int ret = kstrtobool(str, &iommu_dma_forcedac);
97 if (!ret && iommu_dma_forcedac)
98 pr_info("Forcing DAC for PCI devices\n");
101 early_param("iommu.forcedac", iommu_dma_forcedac_setup);
103 /* Number of entries per flush queue */
104 #define IOVA_DEFAULT_FQ_SIZE 256
105 #define IOVA_SINGLE_FQ_SIZE 32768
107 /* Timeout (in ms) after which entries are flushed from the queue */
108 #define IOVA_DEFAULT_FQ_TIMEOUT 10
109 #define IOVA_SINGLE_FQ_TIMEOUT 1000
111 /* Flush queue entry for deferred flushing */
112 struct iova_fq_entry {
113 unsigned long iova_pfn;
115 struct list_head freelist;
116 u64 counter; /* Flush counter when this entry was added */
119 /* Per-CPU flush queue structure */
122 unsigned int head, tail;
123 unsigned int mod_mask;
124 struct iova_fq_entry entries[];
127 #define fq_ring_for_each(i, fq) \
128 for ((i) = (fq)->head; (i) != (fq)->tail; (i) = ((i) + 1) & (fq)->mod_mask)
130 static inline bool fq_full(struct iova_fq *fq)
132 assert_spin_locked(&fq->lock);
133 return (((fq->tail + 1) & fq->mod_mask) == fq->head);
136 static inline unsigned int fq_ring_add(struct iova_fq *fq)
138 unsigned int idx = fq->tail;
140 assert_spin_locked(&fq->lock);
142 fq->tail = (idx + 1) & fq->mod_mask;
147 static void fq_ring_free_locked(struct iommu_dma_cookie *cookie, struct iova_fq *fq)
149 u64 counter = atomic64_read(&cookie->fq_flush_finish_cnt);
152 assert_spin_locked(&fq->lock);
154 fq_ring_for_each(idx, fq) {
156 if (fq->entries[idx].counter >= counter)
159 put_pages_list(&fq->entries[idx].freelist);
160 free_iova_fast(&cookie->iovad,
161 fq->entries[idx].iova_pfn,
162 fq->entries[idx].pages);
164 fq->head = (fq->head + 1) & fq->mod_mask;
168 static void fq_ring_free(struct iommu_dma_cookie *cookie, struct iova_fq *fq)
172 spin_lock_irqsave(&fq->lock, flags);
173 fq_ring_free_locked(cookie, fq);
174 spin_unlock_irqrestore(&fq->lock, flags);
177 static void fq_flush_iotlb(struct iommu_dma_cookie *cookie)
179 atomic64_inc(&cookie->fq_flush_start_cnt);
180 cookie->fq_domain->ops->flush_iotlb_all(cookie->fq_domain);
181 atomic64_inc(&cookie->fq_flush_finish_cnt);
184 static void fq_flush_timeout(struct timer_list *t)
186 struct iommu_dma_cookie *cookie = from_timer(cookie, t, fq_timer);
189 atomic_set(&cookie->fq_timer_on, 0);
190 fq_flush_iotlb(cookie);
192 if (cookie->options.qt == IOMMU_DMA_OPTS_SINGLE_QUEUE) {
193 fq_ring_free(cookie, cookie->single_fq);
195 for_each_possible_cpu(cpu)
196 fq_ring_free(cookie, per_cpu_ptr(cookie->percpu_fq, cpu));
200 static void queue_iova(struct iommu_dma_cookie *cookie,
201 unsigned long pfn, unsigned long pages,
202 struct list_head *freelist)
209 * Order against the IOMMU driver's pagetable update from unmapping
210 * @pte, to guarantee that fq_flush_iotlb() observes that if called
211 * from a different CPU before we release the lock below. Full barrier
212 * so it also pairs with iommu_dma_init_fq() to avoid seeing partially
213 * written fq state here.
217 if (cookie->options.qt == IOMMU_DMA_OPTS_SINGLE_QUEUE)
218 fq = cookie->single_fq;
220 fq = raw_cpu_ptr(cookie->percpu_fq);
222 spin_lock_irqsave(&fq->lock, flags);
225 * First remove all entries from the flush queue that have already been
226 * flushed out on another CPU. This makes the fq_full() check below less
229 fq_ring_free_locked(cookie, fq);
232 fq_flush_iotlb(cookie);
233 fq_ring_free_locked(cookie, fq);
236 idx = fq_ring_add(fq);
238 fq->entries[idx].iova_pfn = pfn;
239 fq->entries[idx].pages = pages;
240 fq->entries[idx].counter = atomic64_read(&cookie->fq_flush_start_cnt);
241 list_splice(freelist, &fq->entries[idx].freelist);
243 spin_unlock_irqrestore(&fq->lock, flags);
245 /* Avoid false sharing as much as possible. */
246 if (!atomic_read(&cookie->fq_timer_on) &&
247 !atomic_xchg(&cookie->fq_timer_on, 1))
248 mod_timer(&cookie->fq_timer,
249 jiffies + msecs_to_jiffies(cookie->options.fq_timeout));
252 static void iommu_dma_free_fq_single(struct iova_fq *fq)
256 fq_ring_for_each(idx, fq)
257 put_pages_list(&fq->entries[idx].freelist);
261 static void iommu_dma_free_fq_percpu(struct iova_fq __percpu *percpu_fq)
265 /* The IOVAs will be torn down separately, so just free our queued pages */
266 for_each_possible_cpu(cpu) {
267 struct iova_fq *fq = per_cpu_ptr(percpu_fq, cpu);
269 fq_ring_for_each(idx, fq)
270 put_pages_list(&fq->entries[idx].freelist);
273 free_percpu(percpu_fq);
276 static void iommu_dma_free_fq(struct iommu_dma_cookie *cookie)
278 if (!cookie->fq_domain)
281 del_timer_sync(&cookie->fq_timer);
282 if (cookie->options.qt == IOMMU_DMA_OPTS_SINGLE_QUEUE)
283 iommu_dma_free_fq_single(cookie->single_fq);
285 iommu_dma_free_fq_percpu(cookie->percpu_fq);
288 static void iommu_dma_init_one_fq(struct iova_fq *fq, size_t fq_size)
294 fq->mod_mask = fq_size - 1;
296 spin_lock_init(&fq->lock);
298 for (i = 0; i < fq_size; i++)
299 INIT_LIST_HEAD(&fq->entries[i].freelist);
302 static int iommu_dma_init_fq_single(struct iommu_dma_cookie *cookie)
304 size_t fq_size = cookie->options.fq_size;
305 struct iova_fq *queue;
307 queue = vmalloc(struct_size(queue, entries, fq_size));
310 iommu_dma_init_one_fq(queue, fq_size);
311 cookie->single_fq = queue;
316 static int iommu_dma_init_fq_percpu(struct iommu_dma_cookie *cookie)
318 size_t fq_size = cookie->options.fq_size;
319 struct iova_fq __percpu *queue;
322 queue = __alloc_percpu(struct_size(queue, entries, fq_size),
323 __alignof__(*queue));
327 for_each_possible_cpu(cpu)
328 iommu_dma_init_one_fq(per_cpu_ptr(queue, cpu), fq_size);
329 cookie->percpu_fq = queue;
333 /* sysfs updates are serialised by the mutex of the group owning @domain */
334 int iommu_dma_init_fq(struct iommu_domain *domain)
336 struct iommu_dma_cookie *cookie = domain->iova_cookie;
339 if (cookie->fq_domain)
342 atomic64_set(&cookie->fq_flush_start_cnt, 0);
343 atomic64_set(&cookie->fq_flush_finish_cnt, 0);
345 if (cookie->options.qt == IOMMU_DMA_OPTS_SINGLE_QUEUE)
346 rc = iommu_dma_init_fq_single(cookie);
348 rc = iommu_dma_init_fq_percpu(cookie);
351 pr_warn("iova flush queue initialization failed\n");
355 timer_setup(&cookie->fq_timer, fq_flush_timeout, 0);
356 atomic_set(&cookie->fq_timer_on, 0);
358 * Prevent incomplete fq state being observable. Pairs with path from
359 * __iommu_dma_unmap() through iommu_dma_free_iova() to queue_iova()
362 WRITE_ONCE(cookie->fq_domain, domain);
366 static inline size_t cookie_msi_granule(struct iommu_dma_cookie *cookie)
368 if (cookie->type == IOMMU_DMA_IOVA_COOKIE)
369 return cookie->iovad.granule;
373 static struct iommu_dma_cookie *cookie_alloc(enum iommu_dma_cookie_type type)
375 struct iommu_dma_cookie *cookie;
377 cookie = kzalloc(sizeof(*cookie), GFP_KERNEL);
379 INIT_LIST_HEAD(&cookie->msi_page_list);
386 * iommu_get_dma_cookie - Acquire DMA-API resources for a domain
387 * @domain: IOMMU domain to prepare for DMA-API usage
389 int iommu_get_dma_cookie(struct iommu_domain *domain)
391 if (domain->iova_cookie)
394 domain->iova_cookie = cookie_alloc(IOMMU_DMA_IOVA_COOKIE);
395 if (!domain->iova_cookie)
398 mutex_init(&domain->iova_cookie->mutex);
403 * iommu_get_msi_cookie - Acquire just MSI remapping resources
404 * @domain: IOMMU domain to prepare
405 * @base: Start address of IOVA region for MSI mappings
407 * Users who manage their own IOVA allocation and do not want DMA API support,
408 * but would still like to take advantage of automatic MSI remapping, can use
409 * this to initialise their own domain appropriately. Users should reserve a
410 * contiguous IOVA region, starting at @base, large enough to accommodate the
411 * number of PAGE_SIZE mappings necessary to cover every MSI doorbell address
412 * used by the devices attached to @domain.
414 int iommu_get_msi_cookie(struct iommu_domain *domain, dma_addr_t base)
416 struct iommu_dma_cookie *cookie;
418 if (domain->type != IOMMU_DOMAIN_UNMANAGED)
421 if (domain->iova_cookie)
424 cookie = cookie_alloc(IOMMU_DMA_MSI_COOKIE);
428 cookie->msi_iova = base;
429 domain->iova_cookie = cookie;
432 EXPORT_SYMBOL(iommu_get_msi_cookie);
435 * iommu_put_dma_cookie - Release a domain's DMA mapping resources
436 * @domain: IOMMU domain previously prepared by iommu_get_dma_cookie() or
437 * iommu_get_msi_cookie()
439 void iommu_put_dma_cookie(struct iommu_domain *domain)
441 struct iommu_dma_cookie *cookie = domain->iova_cookie;
442 struct iommu_dma_msi_page *msi, *tmp;
447 if (cookie->type == IOMMU_DMA_IOVA_COOKIE && cookie->iovad.granule) {
448 iommu_dma_free_fq(cookie);
449 put_iova_domain(&cookie->iovad);
452 list_for_each_entry_safe(msi, tmp, &cookie->msi_page_list, list) {
453 list_del(&msi->list);
457 domain->iova_cookie = NULL;
461 * iommu_dma_get_resv_regions - Reserved region driver helper
462 * @dev: Device from iommu_get_resv_regions()
463 * @list: Reserved region list from iommu_get_resv_regions()
465 * IOMMU drivers can use this to implement their .get_resv_regions callback
466 * for general non-IOMMU-specific reservations. Currently, this covers GICv3
467 * ITS region reservation on ACPI based ARM platforms that may require HW MSI
470 void iommu_dma_get_resv_regions(struct device *dev, struct list_head *list)
473 if (!is_of_node(dev_iommu_fwspec_get(dev)->iommu_fwnode))
474 iort_iommu_get_resv_regions(dev, list);
477 of_iommu_get_resv_regions(dev, list);
479 EXPORT_SYMBOL(iommu_dma_get_resv_regions);
481 static int cookie_init_hw_msi_region(struct iommu_dma_cookie *cookie,
482 phys_addr_t start, phys_addr_t end)
484 struct iova_domain *iovad = &cookie->iovad;
485 struct iommu_dma_msi_page *msi_page;
488 start -= iova_offset(iovad, start);
489 num_pages = iova_align(iovad, end - start) >> iova_shift(iovad);
491 for (i = 0; i < num_pages; i++) {
492 msi_page = kmalloc(sizeof(*msi_page), GFP_KERNEL);
496 msi_page->phys = start;
497 msi_page->iova = start;
498 INIT_LIST_HEAD(&msi_page->list);
499 list_add(&msi_page->list, &cookie->msi_page_list);
500 start += iovad->granule;
506 static int iommu_dma_ranges_sort(void *priv, const struct list_head *a,
507 const struct list_head *b)
509 struct resource_entry *res_a = list_entry(a, typeof(*res_a), node);
510 struct resource_entry *res_b = list_entry(b, typeof(*res_b), node);
512 return res_a->res->start > res_b->res->start;
515 static int iova_reserve_pci_windows(struct pci_dev *dev,
516 struct iova_domain *iovad)
518 struct pci_host_bridge *bridge = pci_find_host_bridge(dev->bus);
519 struct resource_entry *window;
520 unsigned long lo, hi;
521 phys_addr_t start = 0, end;
523 resource_list_for_each_entry(window, &bridge->windows) {
524 if (resource_type(window->res) != IORESOURCE_MEM)
527 lo = iova_pfn(iovad, window->res->start - window->offset);
528 hi = iova_pfn(iovad, window->res->end - window->offset);
529 reserve_iova(iovad, lo, hi);
532 /* Get reserved DMA windows from host bridge */
533 list_sort(NULL, &bridge->dma_ranges, iommu_dma_ranges_sort);
534 resource_list_for_each_entry(window, &bridge->dma_ranges) {
535 end = window->res->start - window->offset;
538 lo = iova_pfn(iovad, start);
539 hi = iova_pfn(iovad, end);
540 reserve_iova(iovad, lo, hi);
541 } else if (end < start) {
542 /* DMA ranges should be non-overlapping */
544 "Failed to reserve IOVA [%pa-%pa]\n",
549 start = window->res->end - window->offset + 1;
550 /* If window is last entry */
551 if (window->node.next == &bridge->dma_ranges &&
552 end != ~(phys_addr_t)0) {
553 end = ~(phys_addr_t)0;
561 static int iova_reserve_iommu_regions(struct device *dev,
562 struct iommu_domain *domain)
564 struct iommu_dma_cookie *cookie = domain->iova_cookie;
565 struct iova_domain *iovad = &cookie->iovad;
566 struct iommu_resv_region *region;
567 LIST_HEAD(resv_regions);
570 if (dev_is_pci(dev)) {
571 ret = iova_reserve_pci_windows(to_pci_dev(dev), iovad);
576 iommu_get_resv_regions(dev, &resv_regions);
577 list_for_each_entry(region, &resv_regions, list) {
578 unsigned long lo, hi;
580 /* We ARE the software that manages these! */
581 if (region->type == IOMMU_RESV_SW_MSI)
584 lo = iova_pfn(iovad, region->start);
585 hi = iova_pfn(iovad, region->start + region->length - 1);
586 reserve_iova(iovad, lo, hi);
588 if (region->type == IOMMU_RESV_MSI)
589 ret = cookie_init_hw_msi_region(cookie, region->start,
590 region->start + region->length);
594 iommu_put_resv_regions(dev, &resv_regions);
599 static bool dev_is_untrusted(struct device *dev)
601 return dev_is_pci(dev) && to_pci_dev(dev)->untrusted;
604 static bool dev_use_swiotlb(struct device *dev, size_t size,
605 enum dma_data_direction dir)
607 return IS_ENABLED(CONFIG_SWIOTLB) &&
608 (dev_is_untrusted(dev) ||
609 dma_kmalloc_needs_bounce(dev, size, dir));
612 static bool dev_use_sg_swiotlb(struct device *dev, struct scatterlist *sg,
613 int nents, enum dma_data_direction dir)
615 struct scatterlist *s;
618 if (!IS_ENABLED(CONFIG_SWIOTLB))
621 if (dev_is_untrusted(dev))
625 * If kmalloc() buffers are not DMA-safe for this device and
626 * direction, check the individual lengths in the sg list. If any
627 * element is deemed unsafe, use the swiotlb for bouncing.
629 if (!dma_kmalloc_safe(dev, dir)) {
630 for_each_sg(sg, s, nents, i)
631 if (!dma_kmalloc_size_aligned(s->length))
639 * iommu_dma_init_options - Initialize dma-iommu options
640 * @options: The options to be initialized
641 * @dev: Device the options are set for
643 * This allows tuning dma-iommu specific to device properties
645 static void iommu_dma_init_options(struct iommu_dma_options *options,
648 /* Shadowing IOTLB flushes do better with a single large queue */
649 if (dev->iommu->shadow_on_flush) {
650 options->qt = IOMMU_DMA_OPTS_SINGLE_QUEUE;
651 options->fq_timeout = IOVA_SINGLE_FQ_TIMEOUT;
652 options->fq_size = IOVA_SINGLE_FQ_SIZE;
654 options->qt = IOMMU_DMA_OPTS_PER_CPU_QUEUE;
655 options->fq_size = IOVA_DEFAULT_FQ_SIZE;
656 options->fq_timeout = IOVA_DEFAULT_FQ_TIMEOUT;
661 * iommu_dma_init_domain - Initialise a DMA mapping domain
662 * @domain: IOMMU domain previously prepared by iommu_get_dma_cookie()
663 * @base: IOVA at which the mappable address space starts
664 * @limit: Last address of the IOVA space
665 * @dev: Device the domain is being initialised for
667 * @base and @limit + 1 should be exact multiples of IOMMU page granularity to
668 * avoid rounding surprises. If necessary, we reserve the page at address 0
669 * to ensure it is an invalid IOVA. It is safe to reinitialise a domain, but
670 * any change which could make prior IOVAs invalid will fail.
672 static int iommu_dma_init_domain(struct iommu_domain *domain, dma_addr_t base,
673 dma_addr_t limit, struct device *dev)
675 struct iommu_dma_cookie *cookie = domain->iova_cookie;
676 unsigned long order, base_pfn;
677 struct iova_domain *iovad;
680 if (!cookie || cookie->type != IOMMU_DMA_IOVA_COOKIE)
683 iovad = &cookie->iovad;
685 /* Use the smallest supported page size for IOVA granularity */
686 order = __ffs(domain->pgsize_bitmap);
687 base_pfn = max_t(unsigned long, 1, base >> order);
689 /* Check the domain allows at least some access to the device... */
690 if (domain->geometry.force_aperture) {
691 if (base > domain->geometry.aperture_end ||
692 limit < domain->geometry.aperture_start) {
693 pr_warn("specified DMA range outside IOMMU capability\n");
696 /* ...then finally give it a kicking to make sure it fits */
697 base_pfn = max_t(unsigned long, base_pfn,
698 domain->geometry.aperture_start >> order);
701 /* start_pfn is always nonzero for an already-initialised domain */
702 mutex_lock(&cookie->mutex);
703 if (iovad->start_pfn) {
704 if (1UL << order != iovad->granule ||
705 base_pfn != iovad->start_pfn) {
706 pr_warn("Incompatible range for DMA domain\n");
715 init_iova_domain(iovad, 1UL << order, base_pfn);
716 ret = iova_domain_init_rcaches(iovad);
720 iommu_dma_init_options(&cookie->options, dev);
722 /* If the FQ fails we can simply fall back to strict mode */
723 if (domain->type == IOMMU_DOMAIN_DMA_FQ &&
724 (!device_iommu_capable(dev, IOMMU_CAP_DEFERRED_FLUSH) || iommu_dma_init_fq(domain)))
725 domain->type = IOMMU_DOMAIN_DMA;
727 ret = iova_reserve_iommu_regions(dev, domain);
730 mutex_unlock(&cookie->mutex);
735 * dma_info_to_prot - Translate DMA API directions and attributes to IOMMU API
737 * @dir: Direction of DMA transfer
738 * @coherent: Is the DMA master cache-coherent?
739 * @attrs: DMA attributes for the mapping
741 * Return: corresponding IOMMU API page protection flags
743 static int dma_info_to_prot(enum dma_data_direction dir, bool coherent,
746 int prot = coherent ? IOMMU_CACHE : 0;
748 if (attrs & DMA_ATTR_PRIVILEGED)
752 case DMA_BIDIRECTIONAL:
753 return prot | IOMMU_READ | IOMMU_WRITE;
755 return prot | IOMMU_READ;
756 case DMA_FROM_DEVICE:
757 return prot | IOMMU_WRITE;
763 static dma_addr_t iommu_dma_alloc_iova(struct iommu_domain *domain,
764 size_t size, u64 dma_limit, struct device *dev)
766 struct iommu_dma_cookie *cookie = domain->iova_cookie;
767 struct iova_domain *iovad = &cookie->iovad;
768 unsigned long shift, iova_len, iova;
770 if (cookie->type == IOMMU_DMA_MSI_COOKIE) {
771 cookie->msi_iova += size;
772 return cookie->msi_iova - size;
775 shift = iova_shift(iovad);
776 iova_len = size >> shift;
778 dma_limit = min_not_zero(dma_limit, dev->bus_dma_limit);
780 if (domain->geometry.force_aperture)
781 dma_limit = min(dma_limit, (u64)domain->geometry.aperture_end);
784 * Try to use all the 32-bit PCI addresses first. The original SAC vs.
785 * DAC reasoning loses relevance with PCIe, but enough hardware and
786 * firmware bugs are still lurking out there that it's safest not to
787 * venture into the 64-bit space until necessary.
789 * If your device goes wrong after seeing the notice then likely either
790 * its driver is not setting DMA masks accurately, the hardware has
791 * some inherent bug in handling >32-bit addresses, or not all the
792 * expected address bits are wired up between the device and the IOMMU.
794 if (dma_limit > DMA_BIT_MASK(32) && dev->iommu->pci_32bit_workaround) {
795 iova = alloc_iova_fast(iovad, iova_len,
796 DMA_BIT_MASK(32) >> shift, false);
800 dev->iommu->pci_32bit_workaround = false;
801 dev_notice(dev, "Using %d-bit DMA addresses\n", bits_per(dma_limit));
804 iova = alloc_iova_fast(iovad, iova_len, dma_limit >> shift, true);
806 return (dma_addr_t)iova << shift;
809 static void iommu_dma_free_iova(struct iommu_dma_cookie *cookie,
810 dma_addr_t iova, size_t size, struct iommu_iotlb_gather *gather)
812 struct iova_domain *iovad = &cookie->iovad;
814 /* The MSI case is only ever cleaning up its most recent allocation */
815 if (cookie->type == IOMMU_DMA_MSI_COOKIE)
816 cookie->msi_iova -= size;
817 else if (gather && gather->queued)
818 queue_iova(cookie, iova_pfn(iovad, iova),
819 size >> iova_shift(iovad),
822 free_iova_fast(iovad, iova_pfn(iovad, iova),
823 size >> iova_shift(iovad));
826 static void __iommu_dma_unmap(struct device *dev, dma_addr_t dma_addr,
829 struct iommu_domain *domain = iommu_get_dma_domain(dev);
830 struct iommu_dma_cookie *cookie = domain->iova_cookie;
831 struct iova_domain *iovad = &cookie->iovad;
832 size_t iova_off = iova_offset(iovad, dma_addr);
833 struct iommu_iotlb_gather iotlb_gather;
836 dma_addr -= iova_off;
837 size = iova_align(iovad, size + iova_off);
838 iommu_iotlb_gather_init(&iotlb_gather);
839 iotlb_gather.queued = READ_ONCE(cookie->fq_domain);
841 unmapped = iommu_unmap_fast(domain, dma_addr, size, &iotlb_gather);
842 WARN_ON(unmapped != size);
844 if (!iotlb_gather.queued)
845 iommu_iotlb_sync(domain, &iotlb_gather);
846 iommu_dma_free_iova(cookie, dma_addr, size, &iotlb_gather);
849 static dma_addr_t __iommu_dma_map(struct device *dev, phys_addr_t phys,
850 size_t size, int prot, u64 dma_mask)
852 struct iommu_domain *domain = iommu_get_dma_domain(dev);
853 struct iommu_dma_cookie *cookie = domain->iova_cookie;
854 struct iova_domain *iovad = &cookie->iovad;
855 size_t iova_off = iova_offset(iovad, phys);
858 if (static_branch_unlikely(&iommu_deferred_attach_enabled) &&
859 iommu_deferred_attach(dev, domain))
860 return DMA_MAPPING_ERROR;
862 size = iova_align(iovad, size + iova_off);
864 iova = iommu_dma_alloc_iova(domain, size, dma_mask, dev);
866 return DMA_MAPPING_ERROR;
868 if (iommu_map(domain, iova, phys - iova_off, size, prot, GFP_ATOMIC)) {
869 iommu_dma_free_iova(cookie, iova, size, NULL);
870 return DMA_MAPPING_ERROR;
872 return iova + iova_off;
875 static void __iommu_dma_free_pages(struct page **pages, int count)
878 __free_page(pages[count]);
882 static struct page **__iommu_dma_alloc_pages(struct device *dev,
883 unsigned int count, unsigned long order_mask, gfp_t gfp)
886 unsigned int i = 0, nid = dev_to_node(dev);
888 order_mask &= GENMASK(MAX_PAGE_ORDER, 0);
892 pages = kvcalloc(count, sizeof(*pages), GFP_KERNEL);
896 /* IOMMU can map any pages, so himem can also be used here */
897 gfp |= __GFP_NOWARN | __GFP_HIGHMEM;
900 struct page *page = NULL;
901 unsigned int order_size;
904 * Higher-order allocations are a convenience rather
905 * than a necessity, hence using __GFP_NORETRY until
906 * falling back to minimum-order allocations.
908 for (order_mask &= GENMASK(__fls(count), 0);
909 order_mask; order_mask &= ~order_size) {
910 unsigned int order = __fls(order_mask);
911 gfp_t alloc_flags = gfp;
913 order_size = 1U << order;
914 if (order_mask > order_size)
915 alloc_flags |= __GFP_NORETRY;
916 page = alloc_pages_node(nid, alloc_flags, order);
920 split_page(page, order);
924 __iommu_dma_free_pages(pages, i);
935 * If size is less than PAGE_SIZE, then a full CPU page will be allocated,
936 * but an IOMMU which supports smaller pages might not map the whole thing.
938 static struct page **__iommu_dma_alloc_noncontiguous(struct device *dev,
939 size_t size, struct sg_table *sgt, gfp_t gfp, pgprot_t prot,
942 struct iommu_domain *domain = iommu_get_dma_domain(dev);
943 struct iommu_dma_cookie *cookie = domain->iova_cookie;
944 struct iova_domain *iovad = &cookie->iovad;
945 bool coherent = dev_is_dma_coherent(dev);
946 int ioprot = dma_info_to_prot(DMA_BIDIRECTIONAL, coherent, attrs);
947 unsigned int count, min_size, alloc_sizes = domain->pgsize_bitmap;
952 if (static_branch_unlikely(&iommu_deferred_attach_enabled) &&
953 iommu_deferred_attach(dev, domain))
956 min_size = alloc_sizes & -alloc_sizes;
957 if (min_size < PAGE_SIZE) {
958 min_size = PAGE_SIZE;
959 alloc_sizes |= PAGE_SIZE;
961 size = ALIGN(size, min_size);
963 if (attrs & DMA_ATTR_ALLOC_SINGLE_PAGES)
964 alloc_sizes = min_size;
966 count = PAGE_ALIGN(size) >> PAGE_SHIFT;
967 pages = __iommu_dma_alloc_pages(dev, count, alloc_sizes >> PAGE_SHIFT,
972 size = iova_align(iovad, size);
973 iova = iommu_dma_alloc_iova(domain, size, dev->coherent_dma_mask, dev);
978 * Remove the zone/policy flags from the GFP - these are applied to the
979 * __iommu_dma_alloc_pages() but are not used for the supporting
980 * internal allocations that follow.
982 gfp &= ~(__GFP_DMA | __GFP_DMA32 | __GFP_HIGHMEM | __GFP_COMP);
984 if (sg_alloc_table_from_pages(sgt, pages, count, 0, size, gfp))
987 if (!(ioprot & IOMMU_CACHE)) {
988 struct scatterlist *sg;
991 for_each_sg(sgt->sgl, sg, sgt->orig_nents, i)
992 arch_dma_prep_coherent(sg_page(sg), sg->length);
995 ret = iommu_map_sg(domain, iova, sgt->sgl, sgt->orig_nents, ioprot,
997 if (ret < 0 || ret < size)
1000 sgt->sgl->dma_address = iova;
1001 sgt->sgl->dma_length = size;
1007 iommu_dma_free_iova(cookie, iova, size, NULL);
1009 __iommu_dma_free_pages(pages, count);
1013 static void *iommu_dma_alloc_remap(struct device *dev, size_t size,
1014 dma_addr_t *dma_handle, gfp_t gfp, pgprot_t prot,
1015 unsigned long attrs)
1017 struct page **pages;
1018 struct sg_table sgt;
1021 pages = __iommu_dma_alloc_noncontiguous(dev, size, &sgt, gfp, prot,
1025 *dma_handle = sgt.sgl->dma_address;
1026 sg_free_table(&sgt);
1027 vaddr = dma_common_pages_remap(pages, size, prot,
1028 __builtin_return_address(0));
1034 __iommu_dma_unmap(dev, *dma_handle, size);
1035 __iommu_dma_free_pages(pages, PAGE_ALIGN(size) >> PAGE_SHIFT);
1039 static struct sg_table *iommu_dma_alloc_noncontiguous(struct device *dev,
1040 size_t size, enum dma_data_direction dir, gfp_t gfp,
1041 unsigned long attrs)
1043 struct dma_sgt_handle *sh;
1045 sh = kmalloc(sizeof(*sh), gfp);
1049 sh->pages = __iommu_dma_alloc_noncontiguous(dev, size, &sh->sgt, gfp,
1050 PAGE_KERNEL, attrs);
1058 static void iommu_dma_free_noncontiguous(struct device *dev, size_t size,
1059 struct sg_table *sgt, enum dma_data_direction dir)
1061 struct dma_sgt_handle *sh = sgt_handle(sgt);
1063 __iommu_dma_unmap(dev, sgt->sgl->dma_address, size);
1064 __iommu_dma_free_pages(sh->pages, PAGE_ALIGN(size) >> PAGE_SHIFT);
1065 sg_free_table(&sh->sgt);
1069 static void iommu_dma_sync_single_for_cpu(struct device *dev,
1070 dma_addr_t dma_handle, size_t size, enum dma_data_direction dir)
1074 if (dev_is_dma_coherent(dev) && !dev_use_swiotlb(dev, size, dir))
1077 phys = iommu_iova_to_phys(iommu_get_dma_domain(dev), dma_handle);
1078 if (!dev_is_dma_coherent(dev))
1079 arch_sync_dma_for_cpu(phys, size, dir);
1081 if (is_swiotlb_buffer(dev, phys))
1082 swiotlb_sync_single_for_cpu(dev, phys, size, dir);
1085 static void iommu_dma_sync_single_for_device(struct device *dev,
1086 dma_addr_t dma_handle, size_t size, enum dma_data_direction dir)
1090 if (dev_is_dma_coherent(dev) && !dev_use_swiotlb(dev, size, dir))
1093 phys = iommu_iova_to_phys(iommu_get_dma_domain(dev), dma_handle);
1094 if (is_swiotlb_buffer(dev, phys))
1095 swiotlb_sync_single_for_device(dev, phys, size, dir);
1097 if (!dev_is_dma_coherent(dev))
1098 arch_sync_dma_for_device(phys, size, dir);
1101 static void iommu_dma_sync_sg_for_cpu(struct device *dev,
1102 struct scatterlist *sgl, int nelems,
1103 enum dma_data_direction dir)
1105 struct scatterlist *sg;
1108 if (sg_dma_is_swiotlb(sgl))
1109 for_each_sg(sgl, sg, nelems, i)
1110 iommu_dma_sync_single_for_cpu(dev, sg_dma_address(sg),
1112 else if (!dev_is_dma_coherent(dev))
1113 for_each_sg(sgl, sg, nelems, i)
1114 arch_sync_dma_for_cpu(sg_phys(sg), sg->length, dir);
1117 static void iommu_dma_sync_sg_for_device(struct device *dev,
1118 struct scatterlist *sgl, int nelems,
1119 enum dma_data_direction dir)
1121 struct scatterlist *sg;
1124 if (sg_dma_is_swiotlb(sgl))
1125 for_each_sg(sgl, sg, nelems, i)
1126 iommu_dma_sync_single_for_device(dev,
1129 else if (!dev_is_dma_coherent(dev))
1130 for_each_sg(sgl, sg, nelems, i)
1131 arch_sync_dma_for_device(sg_phys(sg), sg->length, dir);
1134 static dma_addr_t iommu_dma_map_page(struct device *dev, struct page *page,
1135 unsigned long offset, size_t size, enum dma_data_direction dir,
1136 unsigned long attrs)
1138 phys_addr_t phys = page_to_phys(page) + offset;
1139 bool coherent = dev_is_dma_coherent(dev);
1140 int prot = dma_info_to_prot(dir, coherent, attrs);
1141 struct iommu_domain *domain = iommu_get_dma_domain(dev);
1142 struct iommu_dma_cookie *cookie = domain->iova_cookie;
1143 struct iova_domain *iovad = &cookie->iovad;
1144 dma_addr_t iova, dma_mask = dma_get_mask(dev);
1147 * If both the physical buffer start address and size are
1148 * page aligned, we don't need to use a bounce page.
1150 if (dev_use_swiotlb(dev, size, dir) &&
1151 iova_offset(iovad, phys | size)) {
1152 void *padding_start;
1153 size_t padding_size, aligned_size;
1155 if (!is_swiotlb_active(dev)) {
1156 dev_warn_once(dev, "DMA bounce buffers are inactive, unable to map unaligned transaction.\n");
1157 return DMA_MAPPING_ERROR;
1160 trace_swiotlb_bounced(dev, phys, size);
1162 aligned_size = iova_align(iovad, size);
1163 phys = swiotlb_tbl_map_single(dev, phys, size, aligned_size,
1164 iova_mask(iovad), dir, attrs);
1166 if (phys == DMA_MAPPING_ERROR)
1167 return DMA_MAPPING_ERROR;
1169 /* Cleanup the padding area. */
1170 padding_start = phys_to_virt(phys);
1171 padding_size = aligned_size;
1173 if (!(attrs & DMA_ATTR_SKIP_CPU_SYNC) &&
1174 (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL)) {
1175 padding_start += size;
1176 padding_size -= size;
1179 memset(padding_start, 0, padding_size);
1182 if (!coherent && !(attrs & DMA_ATTR_SKIP_CPU_SYNC))
1183 arch_sync_dma_for_device(phys, size, dir);
1185 iova = __iommu_dma_map(dev, phys, size, prot, dma_mask);
1186 if (iova == DMA_MAPPING_ERROR && is_swiotlb_buffer(dev, phys))
1187 swiotlb_tbl_unmap_single(dev, phys, size, dir, attrs);
1191 static void iommu_dma_unmap_page(struct device *dev, dma_addr_t dma_handle,
1192 size_t size, enum dma_data_direction dir, unsigned long attrs)
1194 struct iommu_domain *domain = iommu_get_dma_domain(dev);
1197 phys = iommu_iova_to_phys(domain, dma_handle);
1201 if (!(attrs & DMA_ATTR_SKIP_CPU_SYNC) && !dev_is_dma_coherent(dev))
1202 arch_sync_dma_for_cpu(phys, size, dir);
1204 __iommu_dma_unmap(dev, dma_handle, size);
1206 if (unlikely(is_swiotlb_buffer(dev, phys)))
1207 swiotlb_tbl_unmap_single(dev, phys, size, dir, attrs);
1211 * Prepare a successfully-mapped scatterlist to give back to the caller.
1213 * At this point the segments are already laid out by iommu_dma_map_sg() to
1214 * avoid individually crossing any boundaries, so we merely need to check a
1215 * segment's start address to avoid concatenating across one.
1217 static int __finalise_sg(struct device *dev, struct scatterlist *sg, int nents,
1218 dma_addr_t dma_addr)
1220 struct scatterlist *s, *cur = sg;
1221 unsigned long seg_mask = dma_get_seg_boundary(dev);
1222 unsigned int cur_len = 0, max_len = dma_get_max_seg_size(dev);
1225 for_each_sg(sg, s, nents, i) {
1226 /* Restore this segment's original unaligned fields first */
1227 dma_addr_t s_dma_addr = sg_dma_address(s);
1228 unsigned int s_iova_off = sg_dma_address(s);
1229 unsigned int s_length = sg_dma_len(s);
1230 unsigned int s_iova_len = s->length;
1232 sg_dma_address(s) = DMA_MAPPING_ERROR;
1235 if (sg_dma_is_bus_address(s)) {
1239 sg_dma_unmark_bus_address(s);
1240 sg_dma_address(cur) = s_dma_addr;
1241 sg_dma_len(cur) = s_length;
1242 sg_dma_mark_bus_address(cur);
1248 s->offset += s_iova_off;
1249 s->length = s_length;
1252 * Now fill in the real DMA data. If...
1253 * - there is a valid output segment to append to
1254 * - and this segment starts on an IOVA page boundary
1255 * - but doesn't fall at a segment boundary
1256 * - and wouldn't make the resulting output segment too long
1258 if (cur_len && !s_iova_off && (dma_addr & seg_mask) &&
1259 (max_len - cur_len >= s_length)) {
1260 /* ...then concatenate it with the previous one */
1261 cur_len += s_length;
1263 /* Otherwise start the next output segment */
1269 sg_dma_address(cur) = dma_addr + s_iova_off;
1272 sg_dma_len(cur) = cur_len;
1273 dma_addr += s_iova_len;
1275 if (s_length + s_iova_off < s_iova_len)
1282 * If mapping failed, then just restore the original list,
1283 * but making sure the DMA fields are invalidated.
1285 static void __invalidate_sg(struct scatterlist *sg, int nents)
1287 struct scatterlist *s;
1290 for_each_sg(sg, s, nents, i) {
1291 if (sg_dma_is_bus_address(s)) {
1292 sg_dma_unmark_bus_address(s);
1294 if (sg_dma_address(s) != DMA_MAPPING_ERROR)
1295 s->offset += sg_dma_address(s);
1297 s->length = sg_dma_len(s);
1299 sg_dma_address(s) = DMA_MAPPING_ERROR;
1304 static void iommu_dma_unmap_sg_swiotlb(struct device *dev, struct scatterlist *sg,
1305 int nents, enum dma_data_direction dir, unsigned long attrs)
1307 struct scatterlist *s;
1310 for_each_sg(sg, s, nents, i)
1311 iommu_dma_unmap_page(dev, sg_dma_address(s),
1312 sg_dma_len(s), dir, attrs);
1315 static int iommu_dma_map_sg_swiotlb(struct device *dev, struct scatterlist *sg,
1316 int nents, enum dma_data_direction dir, unsigned long attrs)
1318 struct scatterlist *s;
1321 sg_dma_mark_swiotlb(sg);
1323 for_each_sg(sg, s, nents, i) {
1324 sg_dma_address(s) = iommu_dma_map_page(dev, sg_page(s),
1325 s->offset, s->length, dir, attrs);
1326 if (sg_dma_address(s) == DMA_MAPPING_ERROR)
1328 sg_dma_len(s) = s->length;
1334 iommu_dma_unmap_sg_swiotlb(dev, sg, i, dir, attrs | DMA_ATTR_SKIP_CPU_SYNC);
1339 * The DMA API client is passing in a scatterlist which could describe
1340 * any old buffer layout, but the IOMMU API requires everything to be
1341 * aligned to IOMMU pages. Hence the need for this complicated bit of
1342 * impedance-matching, to be able to hand off a suitably-aligned list,
1343 * but still preserve the original offsets and sizes for the caller.
1345 static int iommu_dma_map_sg(struct device *dev, struct scatterlist *sg,
1346 int nents, enum dma_data_direction dir, unsigned long attrs)
1348 struct iommu_domain *domain = iommu_get_dma_domain(dev);
1349 struct iommu_dma_cookie *cookie = domain->iova_cookie;
1350 struct iova_domain *iovad = &cookie->iovad;
1351 struct scatterlist *s, *prev = NULL;
1352 int prot = dma_info_to_prot(dir, dev_is_dma_coherent(dev), attrs);
1353 struct pci_p2pdma_map_state p2pdma_state = {};
1354 enum pci_p2pdma_map_type map;
1356 size_t iova_len = 0;
1357 unsigned long mask = dma_get_seg_boundary(dev);
1361 if (static_branch_unlikely(&iommu_deferred_attach_enabled)) {
1362 ret = iommu_deferred_attach(dev, domain);
1367 if (dev_use_sg_swiotlb(dev, sg, nents, dir))
1368 return iommu_dma_map_sg_swiotlb(dev, sg, nents, dir, attrs);
1370 if (!(attrs & DMA_ATTR_SKIP_CPU_SYNC))
1371 iommu_dma_sync_sg_for_device(dev, sg, nents, dir);
1374 * Work out how much IOVA space we need, and align the segments to
1375 * IOVA granules for the IOMMU driver to handle. With some clever
1376 * trickery we can modify the list in-place, but reversibly, by
1377 * stashing the unaligned parts in the as-yet-unused DMA fields.
1379 for_each_sg(sg, s, nents, i) {
1380 size_t s_iova_off = iova_offset(iovad, s->offset);
1381 size_t s_length = s->length;
1382 size_t pad_len = (mask - iova_len + 1) & mask;
1384 if (is_pci_p2pdma_page(sg_page(s))) {
1385 map = pci_p2pdma_map_segment(&p2pdma_state, dev, s);
1387 case PCI_P2PDMA_MAP_BUS_ADDR:
1389 * iommu_map_sg() will skip this segment as
1390 * it is marked as a bus address,
1391 * __finalise_sg() will copy the dma address
1392 * into the output segment.
1395 case PCI_P2PDMA_MAP_THRU_HOST_BRIDGE:
1397 * Mapping through host bridge should be
1398 * mapped with regular IOVAs, thus we
1399 * do nothing here and continue below.
1404 goto out_restore_sg;
1408 sg_dma_address(s) = s_iova_off;
1409 sg_dma_len(s) = s_length;
1410 s->offset -= s_iova_off;
1411 s_length = iova_align(iovad, s_length + s_iova_off);
1412 s->length = s_length;
1415 * Due to the alignment of our single IOVA allocation, we can
1416 * depend on these assumptions about the segment boundary mask:
1417 * - If mask size >= IOVA size, then the IOVA range cannot
1418 * possibly fall across a boundary, so we don't care.
1419 * - If mask size < IOVA size, then the IOVA range must start
1420 * exactly on a boundary, therefore we can lay things out
1421 * based purely on segment lengths without needing to know
1422 * the actual addresses beforehand.
1423 * - The mask must be a power of 2, so pad_len == 0 if
1424 * iova_len == 0, thus we cannot dereference prev the first
1425 * time through here (i.e. before it has a meaningful value).
1427 if (pad_len && pad_len < s_length - 1) {
1428 prev->length += pad_len;
1429 iova_len += pad_len;
1432 iova_len += s_length;
1437 return __finalise_sg(dev, sg, nents, 0);
1439 iova = iommu_dma_alloc_iova(domain, iova_len, dma_get_mask(dev), dev);
1442 goto out_restore_sg;
1446 * We'll leave any physical concatenation to the IOMMU driver's
1447 * implementation - it knows better than we do.
1449 ret = iommu_map_sg(domain, iova, sg, nents, prot, GFP_ATOMIC);
1450 if (ret < 0 || ret < iova_len)
1453 return __finalise_sg(dev, sg, nents, iova);
1456 iommu_dma_free_iova(cookie, iova, iova_len, NULL);
1458 __invalidate_sg(sg, nents);
1460 if (ret != -ENOMEM && ret != -EREMOTEIO)
1465 static void iommu_dma_unmap_sg(struct device *dev, struct scatterlist *sg,
1466 int nents, enum dma_data_direction dir, unsigned long attrs)
1468 dma_addr_t end = 0, start;
1469 struct scatterlist *tmp;
1472 if (sg_dma_is_swiotlb(sg)) {
1473 iommu_dma_unmap_sg_swiotlb(dev, sg, nents, dir, attrs);
1477 if (!(attrs & DMA_ATTR_SKIP_CPU_SYNC))
1478 iommu_dma_sync_sg_for_cpu(dev, sg, nents, dir);
1481 * The scatterlist segments are mapped into a single
1482 * contiguous IOVA allocation, the start and end points
1483 * just have to be determined.
1485 for_each_sg(sg, tmp, nents, i) {
1486 if (sg_dma_is_bus_address(tmp)) {
1487 sg_dma_unmark_bus_address(tmp);
1491 if (sg_dma_len(tmp) == 0)
1494 start = sg_dma_address(tmp);
1499 for_each_sg(tmp, tmp, nents, i) {
1500 if (sg_dma_is_bus_address(tmp)) {
1501 sg_dma_unmark_bus_address(tmp);
1505 if (sg_dma_len(tmp) == 0)
1508 end = sg_dma_address(tmp) + sg_dma_len(tmp);
1512 __iommu_dma_unmap(dev, start, end - start);
1515 static dma_addr_t iommu_dma_map_resource(struct device *dev, phys_addr_t phys,
1516 size_t size, enum dma_data_direction dir, unsigned long attrs)
1518 return __iommu_dma_map(dev, phys, size,
1519 dma_info_to_prot(dir, false, attrs) | IOMMU_MMIO,
1523 static void iommu_dma_unmap_resource(struct device *dev, dma_addr_t handle,
1524 size_t size, enum dma_data_direction dir, unsigned long attrs)
1526 __iommu_dma_unmap(dev, handle, size);
1529 static void __iommu_dma_free(struct device *dev, size_t size, void *cpu_addr)
1531 size_t alloc_size = PAGE_ALIGN(size);
1532 int count = alloc_size >> PAGE_SHIFT;
1533 struct page *page = NULL, **pages = NULL;
1535 /* Non-coherent atomic allocation? Easy */
1536 if (IS_ENABLED(CONFIG_DMA_DIRECT_REMAP) &&
1537 dma_free_from_pool(dev, cpu_addr, alloc_size))
1540 if (is_vmalloc_addr(cpu_addr)) {
1542 * If it the address is remapped, then it's either non-coherent
1543 * or highmem CMA, or an iommu_dma_alloc_remap() construction.
1545 pages = dma_common_find_pages(cpu_addr);
1547 page = vmalloc_to_page(cpu_addr);
1548 dma_common_free_remap(cpu_addr, alloc_size);
1550 /* Lowmem means a coherent atomic or CMA allocation */
1551 page = virt_to_page(cpu_addr);
1555 __iommu_dma_free_pages(pages, count);
1557 dma_free_contiguous(dev, page, alloc_size);
1560 static void iommu_dma_free(struct device *dev, size_t size, void *cpu_addr,
1561 dma_addr_t handle, unsigned long attrs)
1563 __iommu_dma_unmap(dev, handle, size);
1564 __iommu_dma_free(dev, size, cpu_addr);
1567 static void *iommu_dma_alloc_pages(struct device *dev, size_t size,
1568 struct page **pagep, gfp_t gfp, unsigned long attrs)
1570 bool coherent = dev_is_dma_coherent(dev);
1571 size_t alloc_size = PAGE_ALIGN(size);
1572 int node = dev_to_node(dev);
1573 struct page *page = NULL;
1576 page = dma_alloc_contiguous(dev, alloc_size, gfp);
1578 page = alloc_pages_node(node, gfp, get_order(alloc_size));
1582 if (!coherent || PageHighMem(page)) {
1583 pgprot_t prot = dma_pgprot(dev, PAGE_KERNEL, attrs);
1585 cpu_addr = dma_common_contiguous_remap(page, alloc_size,
1586 prot, __builtin_return_address(0));
1588 goto out_free_pages;
1591 arch_dma_prep_coherent(page, size);
1593 cpu_addr = page_address(page);
1597 memset(cpu_addr, 0, alloc_size);
1600 dma_free_contiguous(dev, page, alloc_size);
1604 static void *iommu_dma_alloc(struct device *dev, size_t size,
1605 dma_addr_t *handle, gfp_t gfp, unsigned long attrs)
1607 bool coherent = dev_is_dma_coherent(dev);
1608 int ioprot = dma_info_to_prot(DMA_BIDIRECTIONAL, coherent, attrs);
1609 struct page *page = NULL;
1614 if (gfpflags_allow_blocking(gfp) &&
1615 !(attrs & DMA_ATTR_FORCE_CONTIGUOUS)) {
1616 return iommu_dma_alloc_remap(dev, size, handle, gfp,
1617 dma_pgprot(dev, PAGE_KERNEL, attrs), attrs);
1620 if (IS_ENABLED(CONFIG_DMA_DIRECT_REMAP) &&
1621 !gfpflags_allow_blocking(gfp) && !coherent)
1622 page = dma_alloc_from_pool(dev, PAGE_ALIGN(size), &cpu_addr,
1625 cpu_addr = iommu_dma_alloc_pages(dev, size, &page, gfp, attrs);
1629 *handle = __iommu_dma_map(dev, page_to_phys(page), size, ioprot,
1630 dev->coherent_dma_mask);
1631 if (*handle == DMA_MAPPING_ERROR) {
1632 __iommu_dma_free(dev, size, cpu_addr);
1639 static int iommu_dma_mmap(struct device *dev, struct vm_area_struct *vma,
1640 void *cpu_addr, dma_addr_t dma_addr, size_t size,
1641 unsigned long attrs)
1643 unsigned long nr_pages = PAGE_ALIGN(size) >> PAGE_SHIFT;
1644 unsigned long pfn, off = vma->vm_pgoff;
1647 vma->vm_page_prot = dma_pgprot(dev, vma->vm_page_prot, attrs);
1649 if (dma_mmap_from_dev_coherent(dev, vma, cpu_addr, size, &ret))
1652 if (off >= nr_pages || vma_pages(vma) > nr_pages - off)
1655 if (is_vmalloc_addr(cpu_addr)) {
1656 struct page **pages = dma_common_find_pages(cpu_addr);
1659 return vm_map_pages(vma, pages, nr_pages);
1660 pfn = vmalloc_to_pfn(cpu_addr);
1662 pfn = page_to_pfn(virt_to_page(cpu_addr));
1665 return remap_pfn_range(vma, vma->vm_start, pfn + off,
1666 vma->vm_end - vma->vm_start,
1670 static int iommu_dma_get_sgtable(struct device *dev, struct sg_table *sgt,
1671 void *cpu_addr, dma_addr_t dma_addr, size_t size,
1672 unsigned long attrs)
1677 if (is_vmalloc_addr(cpu_addr)) {
1678 struct page **pages = dma_common_find_pages(cpu_addr);
1681 return sg_alloc_table_from_pages(sgt, pages,
1682 PAGE_ALIGN(size) >> PAGE_SHIFT,
1683 0, size, GFP_KERNEL);
1686 page = vmalloc_to_page(cpu_addr);
1688 page = virt_to_page(cpu_addr);
1691 ret = sg_alloc_table(sgt, 1, GFP_KERNEL);
1693 sg_set_page(sgt->sgl, page, PAGE_ALIGN(size), 0);
1697 static unsigned long iommu_dma_get_merge_boundary(struct device *dev)
1699 struct iommu_domain *domain = iommu_get_dma_domain(dev);
1701 return (1UL << __ffs(domain->pgsize_bitmap)) - 1;
1704 static size_t iommu_dma_opt_mapping_size(void)
1706 return iova_rcache_range();
1709 static const struct dma_map_ops iommu_dma_ops = {
1710 .flags = DMA_F_PCI_P2PDMA_SUPPORTED,
1711 .alloc = iommu_dma_alloc,
1712 .free = iommu_dma_free,
1713 .alloc_pages = dma_common_alloc_pages,
1714 .free_pages = dma_common_free_pages,
1715 .alloc_noncontiguous = iommu_dma_alloc_noncontiguous,
1716 .free_noncontiguous = iommu_dma_free_noncontiguous,
1717 .mmap = iommu_dma_mmap,
1718 .get_sgtable = iommu_dma_get_sgtable,
1719 .map_page = iommu_dma_map_page,
1720 .unmap_page = iommu_dma_unmap_page,
1721 .map_sg = iommu_dma_map_sg,
1722 .unmap_sg = iommu_dma_unmap_sg,
1723 .sync_single_for_cpu = iommu_dma_sync_single_for_cpu,
1724 .sync_single_for_device = iommu_dma_sync_single_for_device,
1725 .sync_sg_for_cpu = iommu_dma_sync_sg_for_cpu,
1726 .sync_sg_for_device = iommu_dma_sync_sg_for_device,
1727 .map_resource = iommu_dma_map_resource,
1728 .unmap_resource = iommu_dma_unmap_resource,
1729 .get_merge_boundary = iommu_dma_get_merge_boundary,
1730 .opt_mapping_size = iommu_dma_opt_mapping_size,
1734 * The IOMMU core code allocates the default DMA domain, which the underlying
1735 * IOMMU driver needs to support via the dma-iommu layer.
1737 void iommu_setup_dma_ops(struct device *dev, u64 dma_base, u64 dma_limit)
1739 struct iommu_domain *domain = iommu_get_domain_for_dev(dev);
1745 * The IOMMU core code allocates the default DMA domain, which the
1746 * underlying IOMMU driver needs to support via the dma-iommu layer.
1748 if (iommu_is_dma_domain(domain)) {
1749 if (iommu_dma_init_domain(domain, dma_base, dma_limit, dev))
1751 dev->dma_ops = &iommu_dma_ops;
1756 pr_warn("Failed to set up IOMMU for device %s; retaining platform DMA ops\n",
1759 EXPORT_SYMBOL_GPL(iommu_setup_dma_ops);
1761 static struct iommu_dma_msi_page *iommu_dma_get_msi_page(struct device *dev,
1762 phys_addr_t msi_addr, struct iommu_domain *domain)
1764 struct iommu_dma_cookie *cookie = domain->iova_cookie;
1765 struct iommu_dma_msi_page *msi_page;
1767 int prot = IOMMU_WRITE | IOMMU_NOEXEC | IOMMU_MMIO;
1768 size_t size = cookie_msi_granule(cookie);
1770 msi_addr &= ~(phys_addr_t)(size - 1);
1771 list_for_each_entry(msi_page, &cookie->msi_page_list, list)
1772 if (msi_page->phys == msi_addr)
1775 msi_page = kzalloc(sizeof(*msi_page), GFP_KERNEL);
1779 iova = iommu_dma_alloc_iova(domain, size, dma_get_mask(dev), dev);
1783 if (iommu_map(domain, iova, msi_addr, size, prot, GFP_KERNEL))
1786 INIT_LIST_HEAD(&msi_page->list);
1787 msi_page->phys = msi_addr;
1788 msi_page->iova = iova;
1789 list_add(&msi_page->list, &cookie->msi_page_list);
1793 iommu_dma_free_iova(cookie, iova, size, NULL);
1800 * iommu_dma_prepare_msi() - Map the MSI page in the IOMMU domain
1801 * @desc: MSI descriptor, will store the MSI page
1802 * @msi_addr: MSI target address to be mapped
1804 * Return: 0 on success or negative error code if the mapping failed.
1806 int iommu_dma_prepare_msi(struct msi_desc *desc, phys_addr_t msi_addr)
1808 struct device *dev = msi_desc_to_dev(desc);
1809 struct iommu_domain *domain = iommu_get_domain_for_dev(dev);
1810 struct iommu_dma_msi_page *msi_page;
1811 static DEFINE_MUTEX(msi_prepare_lock); /* see below */
1813 if (!domain || !domain->iova_cookie) {
1814 desc->iommu_cookie = NULL;
1819 * In fact the whole prepare operation should already be serialised by
1820 * irq_domain_mutex further up the callchain, but that's pretty subtle
1821 * on its own, so consider this locking as failsafe documentation...
1823 mutex_lock(&msi_prepare_lock);
1824 msi_page = iommu_dma_get_msi_page(dev, msi_addr, domain);
1825 mutex_unlock(&msi_prepare_lock);
1827 msi_desc_set_iommu_cookie(desc, msi_page);
1835 * iommu_dma_compose_msi_msg() - Apply translation to an MSI message
1836 * @desc: MSI descriptor prepared by iommu_dma_prepare_msi()
1837 * @msg: MSI message containing target physical address
1839 void iommu_dma_compose_msi_msg(struct msi_desc *desc, struct msi_msg *msg)
1841 struct device *dev = msi_desc_to_dev(desc);
1842 const struct iommu_domain *domain = iommu_get_domain_for_dev(dev);
1843 const struct iommu_dma_msi_page *msi_page;
1845 msi_page = msi_desc_get_iommu_cookie(desc);
1847 if (!domain || !domain->iova_cookie || WARN_ON(!msi_page))
1850 msg->address_hi = upper_32_bits(msi_page->iova);
1851 msg->address_lo &= cookie_msi_granule(domain->iova_cookie) - 1;
1852 msg->address_lo += lower_32_bits(msi_page->iova);
1855 static int iommu_dma_init(void)
1857 if (is_kdump_kernel())
1858 static_branch_enable(&iommu_deferred_attach_enabled);
1860 return iova_cache_get();
1862 arch_initcall(iommu_dma_init);