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/device.h>
13 #include <linux/dma-map-ops.h>
14 #include <linux/dma-iommu.h>
15 #include <linux/gfp.h>
16 #include <linux/huge_mm.h>
17 #include <linux/iommu.h>
18 #include <linux/iova.h>
19 #include <linux/irq.h>
21 #include <linux/mutex.h>
22 #include <linux/pci.h>
23 #include <linux/swiotlb.h>
24 #include <linux/scatterlist.h>
25 #include <linux/vmalloc.h>
26 #include <linux/crash_dump.h>
27 #include <linux/dma-direct.h>
29 struct iommu_dma_msi_page {
30 struct list_head list;
35 enum iommu_dma_cookie_type {
36 IOMMU_DMA_IOVA_COOKIE,
40 struct iommu_dma_cookie {
41 enum iommu_dma_cookie_type type;
43 /* Full allocator for IOMMU_DMA_IOVA_COOKIE */
44 struct iova_domain iovad;
45 /* Trivial linear page allocator for IOMMU_DMA_MSI_COOKIE */
48 struct list_head msi_page_list;
50 /* Domain for flush queue callback; NULL if flush queue not in use */
51 struct iommu_domain *fq_domain;
54 static DEFINE_STATIC_KEY_FALSE(iommu_deferred_attach_enabled);
55 bool iommu_dma_forcedac __read_mostly;
57 static int __init iommu_dma_forcedac_setup(char *str)
59 int ret = kstrtobool(str, &iommu_dma_forcedac);
61 if (!ret && iommu_dma_forcedac)
62 pr_info("Forcing DAC for PCI devices\n");
65 early_param("iommu.forcedac", iommu_dma_forcedac_setup);
67 static void iommu_dma_entry_dtor(unsigned long data)
69 struct page *freelist = (struct page *)data;
72 unsigned long p = (unsigned long)page_address(freelist);
74 freelist = freelist->freelist;
79 static inline size_t cookie_msi_granule(struct iommu_dma_cookie *cookie)
81 if (cookie->type == IOMMU_DMA_IOVA_COOKIE)
82 return cookie->iovad.granule;
86 static struct iommu_dma_cookie *cookie_alloc(enum iommu_dma_cookie_type type)
88 struct iommu_dma_cookie *cookie;
90 cookie = kzalloc(sizeof(*cookie), GFP_KERNEL);
92 INIT_LIST_HEAD(&cookie->msi_page_list);
99 * iommu_get_dma_cookie - Acquire DMA-API resources for a domain
100 * @domain: IOMMU domain to prepare for DMA-API usage
102 * IOMMU drivers should normally call this from their domain_alloc
103 * callback when domain->type == IOMMU_DOMAIN_DMA.
105 int iommu_get_dma_cookie(struct iommu_domain *domain)
107 if (domain->iova_cookie)
110 domain->iova_cookie = cookie_alloc(IOMMU_DMA_IOVA_COOKIE);
111 if (!domain->iova_cookie)
116 EXPORT_SYMBOL(iommu_get_dma_cookie);
119 * iommu_get_msi_cookie - Acquire just MSI remapping resources
120 * @domain: IOMMU domain to prepare
121 * @base: Start address of IOVA region for MSI mappings
123 * Users who manage their own IOVA allocation and do not want DMA API support,
124 * but would still like to take advantage of automatic MSI remapping, can use
125 * this to initialise their own domain appropriately. Users should reserve a
126 * contiguous IOVA region, starting at @base, large enough to accommodate the
127 * number of PAGE_SIZE mappings necessary to cover every MSI doorbell address
128 * used by the devices attached to @domain.
130 int iommu_get_msi_cookie(struct iommu_domain *domain, dma_addr_t base)
132 struct iommu_dma_cookie *cookie;
134 if (domain->type != IOMMU_DOMAIN_UNMANAGED)
137 if (domain->iova_cookie)
140 cookie = cookie_alloc(IOMMU_DMA_MSI_COOKIE);
144 cookie->msi_iova = base;
145 domain->iova_cookie = cookie;
148 EXPORT_SYMBOL(iommu_get_msi_cookie);
151 * iommu_put_dma_cookie - Release a domain's DMA mapping resources
152 * @domain: IOMMU domain previously prepared by iommu_get_dma_cookie() or
153 * iommu_get_msi_cookie()
155 * IOMMU drivers should normally call this from their domain_free callback.
157 void iommu_put_dma_cookie(struct iommu_domain *domain)
159 struct iommu_dma_cookie *cookie = domain->iova_cookie;
160 struct iommu_dma_msi_page *msi, *tmp;
165 if (cookie->type == IOMMU_DMA_IOVA_COOKIE && cookie->iovad.granule)
166 put_iova_domain(&cookie->iovad);
168 list_for_each_entry_safe(msi, tmp, &cookie->msi_page_list, list) {
169 list_del(&msi->list);
173 domain->iova_cookie = NULL;
175 EXPORT_SYMBOL(iommu_put_dma_cookie);
178 * iommu_dma_get_resv_regions - Reserved region driver helper
179 * @dev: Device from iommu_get_resv_regions()
180 * @list: Reserved region list from iommu_get_resv_regions()
182 * IOMMU drivers can use this to implement their .get_resv_regions callback
183 * for general non-IOMMU-specific reservations. Currently, this covers GICv3
184 * ITS region reservation on ACPI based ARM platforms that may require HW MSI
187 void iommu_dma_get_resv_regions(struct device *dev, struct list_head *list)
190 if (!is_of_node(dev_iommu_fwspec_get(dev)->iommu_fwnode))
191 iort_iommu_msi_get_resv_regions(dev, list);
194 EXPORT_SYMBOL(iommu_dma_get_resv_regions);
196 static int cookie_init_hw_msi_region(struct iommu_dma_cookie *cookie,
197 phys_addr_t start, phys_addr_t end)
199 struct iova_domain *iovad = &cookie->iovad;
200 struct iommu_dma_msi_page *msi_page;
203 start -= iova_offset(iovad, start);
204 num_pages = iova_align(iovad, end - start) >> iova_shift(iovad);
206 for (i = 0; i < num_pages; i++) {
207 msi_page = kmalloc(sizeof(*msi_page), GFP_KERNEL);
211 msi_page->phys = start;
212 msi_page->iova = start;
213 INIT_LIST_HEAD(&msi_page->list);
214 list_add(&msi_page->list, &cookie->msi_page_list);
215 start += iovad->granule;
221 static int iova_reserve_pci_windows(struct pci_dev *dev,
222 struct iova_domain *iovad)
224 struct pci_host_bridge *bridge = pci_find_host_bridge(dev->bus);
225 struct resource_entry *window;
226 unsigned long lo, hi;
227 phys_addr_t start = 0, end;
229 resource_list_for_each_entry(window, &bridge->windows) {
230 if (resource_type(window->res) != IORESOURCE_MEM)
233 lo = iova_pfn(iovad, window->res->start - window->offset);
234 hi = iova_pfn(iovad, window->res->end - window->offset);
235 reserve_iova(iovad, lo, hi);
238 /* Get reserved DMA windows from host bridge */
239 resource_list_for_each_entry(window, &bridge->dma_ranges) {
240 end = window->res->start - window->offset;
243 lo = iova_pfn(iovad, start);
244 hi = iova_pfn(iovad, end);
245 reserve_iova(iovad, lo, hi);
246 } else if (end < start) {
247 /* dma_ranges list should be sorted */
249 "Failed to reserve IOVA [%pa-%pa]\n",
254 start = window->res->end - window->offset + 1;
255 /* If window is last entry */
256 if (window->node.next == &bridge->dma_ranges &&
257 end != ~(phys_addr_t)0) {
258 end = ~(phys_addr_t)0;
266 static int iova_reserve_iommu_regions(struct device *dev,
267 struct iommu_domain *domain)
269 struct iommu_dma_cookie *cookie = domain->iova_cookie;
270 struct iova_domain *iovad = &cookie->iovad;
271 struct iommu_resv_region *region;
272 LIST_HEAD(resv_regions);
275 if (dev_is_pci(dev)) {
276 ret = iova_reserve_pci_windows(to_pci_dev(dev), iovad);
281 iommu_get_resv_regions(dev, &resv_regions);
282 list_for_each_entry(region, &resv_regions, list) {
283 unsigned long lo, hi;
285 /* We ARE the software that manages these! */
286 if (region->type == IOMMU_RESV_SW_MSI)
289 lo = iova_pfn(iovad, region->start);
290 hi = iova_pfn(iovad, region->start + region->length - 1);
291 reserve_iova(iovad, lo, hi);
293 if (region->type == IOMMU_RESV_MSI)
294 ret = cookie_init_hw_msi_region(cookie, region->start,
295 region->start + region->length);
299 iommu_put_resv_regions(dev, &resv_regions);
304 static void iommu_dma_flush_iotlb_all(struct iova_domain *iovad)
306 struct iommu_dma_cookie *cookie;
307 struct iommu_domain *domain;
309 cookie = container_of(iovad, struct iommu_dma_cookie, iovad);
310 domain = cookie->fq_domain;
312 domain->ops->flush_iotlb_all(domain);
315 static bool dev_is_untrusted(struct device *dev)
317 return dev_is_pci(dev) && to_pci_dev(dev)->untrusted;
321 * iommu_dma_init_domain - Initialise a DMA mapping domain
322 * @domain: IOMMU domain previously prepared by iommu_get_dma_cookie()
323 * @base: IOVA at which the mappable address space starts
324 * @limit: Last address of the IOVA space
325 * @dev: Device the domain is being initialised for
327 * @base and @limit + 1 should be exact multiples of IOMMU page granularity to
328 * avoid rounding surprises. If necessary, we reserve the page at address 0
329 * to ensure it is an invalid IOVA. It is safe to reinitialise a domain, but
330 * any change which could make prior IOVAs invalid will fail.
332 static int iommu_dma_init_domain(struct iommu_domain *domain, dma_addr_t base,
333 dma_addr_t limit, struct device *dev)
335 struct iommu_dma_cookie *cookie = domain->iova_cookie;
336 unsigned long order, base_pfn;
337 struct iova_domain *iovad;
339 if (!cookie || cookie->type != IOMMU_DMA_IOVA_COOKIE)
342 iovad = &cookie->iovad;
344 /* Use the smallest supported page size for IOVA granularity */
345 order = __ffs(domain->pgsize_bitmap);
346 base_pfn = max_t(unsigned long, 1, base >> order);
348 /* Check the domain allows at least some access to the device... */
349 if (domain->geometry.force_aperture) {
350 if (base > domain->geometry.aperture_end ||
351 limit < domain->geometry.aperture_start) {
352 pr_warn("specified DMA range outside IOMMU capability\n");
355 /* ...then finally give it a kicking to make sure it fits */
356 base_pfn = max_t(unsigned long, base_pfn,
357 domain->geometry.aperture_start >> order);
360 /* start_pfn is always nonzero for an already-initialised domain */
361 if (iovad->start_pfn) {
362 if (1UL << order != iovad->granule ||
363 base_pfn != iovad->start_pfn) {
364 pr_warn("Incompatible range for DMA domain\n");
371 init_iova_domain(iovad, 1UL << order, base_pfn);
373 if (!cookie->fq_domain && (!dev || !dev_is_untrusted(dev)) &&
374 domain->ops->flush_iotlb_all && !iommu_get_dma_strict(domain)) {
375 if (init_iova_flush_queue(iovad, iommu_dma_flush_iotlb_all,
376 iommu_dma_entry_dtor))
377 pr_warn("iova flush queue initialization failed\n");
379 cookie->fq_domain = domain;
385 return iova_reserve_iommu_regions(dev, domain);
389 * dma_info_to_prot - Translate DMA API directions and attributes to IOMMU API
391 * @dir: Direction of DMA transfer
392 * @coherent: Is the DMA master cache-coherent?
393 * @attrs: DMA attributes for the mapping
395 * Return: corresponding IOMMU API page protection flags
397 static int dma_info_to_prot(enum dma_data_direction dir, bool coherent,
400 int prot = coherent ? IOMMU_CACHE : 0;
402 if (attrs & DMA_ATTR_PRIVILEGED)
406 case DMA_BIDIRECTIONAL:
407 return prot | IOMMU_READ | IOMMU_WRITE;
409 return prot | IOMMU_READ;
410 case DMA_FROM_DEVICE:
411 return prot | IOMMU_WRITE;
417 static dma_addr_t iommu_dma_alloc_iova(struct iommu_domain *domain,
418 size_t size, u64 dma_limit, struct device *dev)
420 struct iommu_dma_cookie *cookie = domain->iova_cookie;
421 struct iova_domain *iovad = &cookie->iovad;
422 unsigned long shift, iova_len, iova = 0;
424 if (cookie->type == IOMMU_DMA_MSI_COOKIE) {
425 cookie->msi_iova += size;
426 return cookie->msi_iova - size;
429 shift = iova_shift(iovad);
430 iova_len = size >> shift;
432 * Freeing non-power-of-two-sized allocations back into the IOVA caches
433 * will come back to bite us badly, so we have to waste a bit of space
434 * rounding up anything cacheable to make sure that can't happen. The
435 * order of the unadjusted size will still match upon freeing.
437 if (iova_len < (1 << (IOVA_RANGE_CACHE_MAX_SIZE - 1)))
438 iova_len = roundup_pow_of_two(iova_len);
440 dma_limit = min_not_zero(dma_limit, dev->bus_dma_limit);
442 if (domain->geometry.force_aperture)
443 dma_limit = min(dma_limit, (u64)domain->geometry.aperture_end);
445 /* Try to get PCI devices a SAC address */
446 if (dma_limit > DMA_BIT_MASK(32) && !iommu_dma_forcedac && dev_is_pci(dev))
447 iova = alloc_iova_fast(iovad, iova_len,
448 DMA_BIT_MASK(32) >> shift, false);
451 iova = alloc_iova_fast(iovad, iova_len, dma_limit >> shift,
454 return (dma_addr_t)iova << shift;
457 static void iommu_dma_free_iova(struct iommu_dma_cookie *cookie,
458 dma_addr_t iova, size_t size, struct page *freelist)
460 struct iova_domain *iovad = &cookie->iovad;
462 /* The MSI case is only ever cleaning up its most recent allocation */
463 if (cookie->type == IOMMU_DMA_MSI_COOKIE)
464 cookie->msi_iova -= size;
465 else if (cookie->fq_domain) /* non-strict mode */
466 queue_iova(iovad, iova_pfn(iovad, iova),
467 size >> iova_shift(iovad),
468 (unsigned long)freelist);
470 free_iova_fast(iovad, iova_pfn(iovad, iova),
471 size >> iova_shift(iovad));
474 static void __iommu_dma_unmap(struct device *dev, dma_addr_t dma_addr,
477 struct iommu_domain *domain = iommu_get_dma_domain(dev);
478 struct iommu_dma_cookie *cookie = domain->iova_cookie;
479 struct iova_domain *iovad = &cookie->iovad;
480 size_t iova_off = iova_offset(iovad, dma_addr);
481 struct iommu_iotlb_gather iotlb_gather;
484 dma_addr -= iova_off;
485 size = iova_align(iovad, size + iova_off);
486 iommu_iotlb_gather_init(&iotlb_gather);
488 unmapped = iommu_unmap_fast(domain, dma_addr, size, &iotlb_gather);
489 WARN_ON(unmapped != size);
491 if (!cookie->fq_domain)
492 iommu_iotlb_sync(domain, &iotlb_gather);
493 iommu_dma_free_iova(cookie, dma_addr, size, iotlb_gather.freelist);
496 static void __iommu_dma_unmap_swiotlb(struct device *dev, dma_addr_t dma_addr,
497 size_t size, enum dma_data_direction dir,
500 struct iommu_domain *domain = iommu_get_dma_domain(dev);
503 phys = iommu_iova_to_phys(domain, dma_addr);
507 __iommu_dma_unmap(dev, dma_addr, size);
509 if (unlikely(is_swiotlb_buffer(phys)))
510 swiotlb_tbl_unmap_single(dev, phys, size, dir, attrs);
513 static dma_addr_t __iommu_dma_map(struct device *dev, phys_addr_t phys,
514 size_t size, int prot, u64 dma_mask)
516 struct iommu_domain *domain = iommu_get_dma_domain(dev);
517 struct iommu_dma_cookie *cookie = domain->iova_cookie;
518 struct iova_domain *iovad = &cookie->iovad;
519 size_t iova_off = iova_offset(iovad, phys);
522 if (static_branch_unlikely(&iommu_deferred_attach_enabled) &&
523 iommu_deferred_attach(dev, domain))
524 return DMA_MAPPING_ERROR;
526 size = iova_align(iovad, size + iova_off);
528 iova = iommu_dma_alloc_iova(domain, size, dma_mask, dev);
530 return DMA_MAPPING_ERROR;
532 if (iommu_map_atomic(domain, iova, phys - iova_off, size, prot)) {
533 iommu_dma_free_iova(cookie, iova, size, NULL);
534 return DMA_MAPPING_ERROR;
536 return iova + iova_off;
539 static dma_addr_t __iommu_dma_map_swiotlb(struct device *dev, phys_addr_t phys,
540 size_t org_size, dma_addr_t dma_mask, bool coherent,
541 enum dma_data_direction dir, unsigned long attrs)
543 int prot = dma_info_to_prot(dir, coherent, attrs);
544 struct iommu_domain *domain = iommu_get_dma_domain(dev);
545 struct iommu_dma_cookie *cookie = domain->iova_cookie;
546 struct iova_domain *iovad = &cookie->iovad;
547 size_t aligned_size = org_size;
553 * If both the physical buffer start address and size are
554 * page aligned, we don't need to use a bounce page.
556 if (IS_ENABLED(CONFIG_SWIOTLB) && dev_is_untrusted(dev) &&
557 iova_offset(iovad, phys | org_size)) {
558 aligned_size = iova_align(iovad, org_size);
559 phys = swiotlb_tbl_map_single(dev, phys, org_size,
560 aligned_size, dir, attrs);
562 if (phys == DMA_MAPPING_ERROR)
563 return DMA_MAPPING_ERROR;
565 /* Cleanup the padding area. */
566 padding_start = phys_to_virt(phys);
567 padding_size = aligned_size;
569 if (!(attrs & DMA_ATTR_SKIP_CPU_SYNC) &&
570 (dir == DMA_TO_DEVICE ||
571 dir == DMA_BIDIRECTIONAL)) {
572 padding_start += org_size;
573 padding_size -= org_size;
576 memset(padding_start, 0, padding_size);
579 iova = __iommu_dma_map(dev, phys, aligned_size, prot, dma_mask);
580 if (iova == DMA_MAPPING_ERROR && is_swiotlb_buffer(phys))
581 swiotlb_tbl_unmap_single(dev, phys, org_size, dir, attrs);
585 static void __iommu_dma_free_pages(struct page **pages, int count)
588 __free_page(pages[count]);
592 static struct page **__iommu_dma_alloc_pages(struct device *dev,
593 unsigned int count, unsigned long order_mask, gfp_t gfp)
596 unsigned int i = 0, nid = dev_to_node(dev);
598 order_mask &= (2U << MAX_ORDER) - 1;
602 pages = kvzalloc(count * sizeof(*pages), GFP_KERNEL);
606 /* IOMMU can map any pages, so himem can also be used here */
607 gfp |= __GFP_NOWARN | __GFP_HIGHMEM;
609 /* It makes no sense to muck about with huge pages */
613 struct page *page = NULL;
614 unsigned int order_size;
617 * Higher-order allocations are a convenience rather
618 * than a necessity, hence using __GFP_NORETRY until
619 * falling back to minimum-order allocations.
621 for (order_mask &= (2U << __fls(count)) - 1;
622 order_mask; order_mask &= ~order_size) {
623 unsigned int order = __fls(order_mask);
624 gfp_t alloc_flags = gfp;
626 order_size = 1U << order;
627 if (order_mask > order_size)
628 alloc_flags |= __GFP_NORETRY;
629 page = alloc_pages_node(nid, alloc_flags, order);
633 split_page(page, order);
637 __iommu_dma_free_pages(pages, i);
648 * If size is less than PAGE_SIZE, then a full CPU page will be allocated,
649 * but an IOMMU which supports smaller pages might not map the whole thing.
651 static struct page **__iommu_dma_alloc_noncontiguous(struct device *dev,
652 size_t size, struct sg_table *sgt, gfp_t gfp, pgprot_t prot,
655 struct iommu_domain *domain = iommu_get_dma_domain(dev);
656 struct iommu_dma_cookie *cookie = domain->iova_cookie;
657 struct iova_domain *iovad = &cookie->iovad;
658 bool coherent = dev_is_dma_coherent(dev);
659 int ioprot = dma_info_to_prot(DMA_BIDIRECTIONAL, coherent, attrs);
660 unsigned int count, min_size, alloc_sizes = domain->pgsize_bitmap;
664 if (static_branch_unlikely(&iommu_deferred_attach_enabled) &&
665 iommu_deferred_attach(dev, domain))
668 min_size = alloc_sizes & -alloc_sizes;
669 if (min_size < PAGE_SIZE) {
670 min_size = PAGE_SIZE;
671 alloc_sizes |= PAGE_SIZE;
673 size = ALIGN(size, min_size);
675 if (attrs & DMA_ATTR_ALLOC_SINGLE_PAGES)
676 alloc_sizes = min_size;
678 count = PAGE_ALIGN(size) >> PAGE_SHIFT;
679 pages = __iommu_dma_alloc_pages(dev, count, alloc_sizes >> PAGE_SHIFT,
684 size = iova_align(iovad, size);
685 iova = iommu_dma_alloc_iova(domain, size, dev->coherent_dma_mask, dev);
689 if (sg_alloc_table_from_pages(sgt, pages, count, 0, size, GFP_KERNEL))
692 if (!(ioprot & IOMMU_CACHE)) {
693 struct scatterlist *sg;
696 for_each_sg(sgt->sgl, sg, sgt->orig_nents, i)
697 arch_dma_prep_coherent(sg_page(sg), sg->length);
700 if (iommu_map_sg_atomic(domain, iova, sgt->sgl, sgt->orig_nents, ioprot)
704 sgt->sgl->dma_address = iova;
705 sgt->sgl->dma_length = size;
711 iommu_dma_free_iova(cookie, iova, size, NULL);
713 __iommu_dma_free_pages(pages, count);
717 static void *iommu_dma_alloc_remap(struct device *dev, size_t size,
718 dma_addr_t *dma_handle, gfp_t gfp, pgprot_t prot,
725 pages = __iommu_dma_alloc_noncontiguous(dev, size, &sgt, gfp, prot,
729 *dma_handle = sgt.sgl->dma_address;
731 vaddr = dma_common_pages_remap(pages, size, prot,
732 __builtin_return_address(0));
738 __iommu_dma_unmap(dev, *dma_handle, size);
739 __iommu_dma_free_pages(pages, PAGE_ALIGN(size) >> PAGE_SHIFT);
743 #ifdef CONFIG_DMA_REMAP
744 static struct sg_table *iommu_dma_alloc_noncontiguous(struct device *dev,
745 size_t size, enum dma_data_direction dir, gfp_t gfp,
748 struct dma_sgt_handle *sh;
750 sh = kmalloc(sizeof(*sh), gfp);
754 sh->pages = __iommu_dma_alloc_noncontiguous(dev, size, &sh->sgt, gfp,
763 static void iommu_dma_free_noncontiguous(struct device *dev, size_t size,
764 struct sg_table *sgt, enum dma_data_direction dir)
766 struct dma_sgt_handle *sh = sgt_handle(sgt);
768 __iommu_dma_unmap(dev, sgt->sgl->dma_address, size);
769 __iommu_dma_free_pages(sh->pages, PAGE_ALIGN(size) >> PAGE_SHIFT);
770 sg_free_table(&sh->sgt);
772 #endif /* CONFIG_DMA_REMAP */
774 static void iommu_dma_sync_single_for_cpu(struct device *dev,
775 dma_addr_t dma_handle, size_t size, enum dma_data_direction dir)
779 if (dev_is_dma_coherent(dev) && !dev_is_untrusted(dev))
782 phys = iommu_iova_to_phys(iommu_get_dma_domain(dev), dma_handle);
783 if (!dev_is_dma_coherent(dev))
784 arch_sync_dma_for_cpu(phys, size, dir);
786 if (is_swiotlb_buffer(phys))
787 swiotlb_sync_single_for_cpu(dev, phys, size, dir);
790 static void iommu_dma_sync_single_for_device(struct device *dev,
791 dma_addr_t dma_handle, size_t size, enum dma_data_direction dir)
795 if (dev_is_dma_coherent(dev) && !dev_is_untrusted(dev))
798 phys = iommu_iova_to_phys(iommu_get_dma_domain(dev), dma_handle);
799 if (is_swiotlb_buffer(phys))
800 swiotlb_sync_single_for_device(dev, phys, size, dir);
802 if (!dev_is_dma_coherent(dev))
803 arch_sync_dma_for_device(phys, size, dir);
806 static void iommu_dma_sync_sg_for_cpu(struct device *dev,
807 struct scatterlist *sgl, int nelems,
808 enum dma_data_direction dir)
810 struct scatterlist *sg;
813 if (dev_is_dma_coherent(dev) && !dev_is_untrusted(dev))
816 for_each_sg(sgl, sg, nelems, i) {
817 if (!dev_is_dma_coherent(dev))
818 arch_sync_dma_for_cpu(sg_phys(sg), sg->length, dir);
820 if (is_swiotlb_buffer(sg_phys(sg)))
821 swiotlb_sync_single_for_cpu(dev, sg_phys(sg),
826 static void iommu_dma_sync_sg_for_device(struct device *dev,
827 struct scatterlist *sgl, int nelems,
828 enum dma_data_direction dir)
830 struct scatterlist *sg;
833 if (dev_is_dma_coherent(dev) && !dev_is_untrusted(dev))
836 for_each_sg(sgl, sg, nelems, i) {
837 if (is_swiotlb_buffer(sg_phys(sg)))
838 swiotlb_sync_single_for_device(dev, sg_phys(sg),
841 if (!dev_is_dma_coherent(dev))
842 arch_sync_dma_for_device(sg_phys(sg), sg->length, dir);
846 static dma_addr_t iommu_dma_map_page(struct device *dev, struct page *page,
847 unsigned long offset, size_t size, enum dma_data_direction dir,
850 phys_addr_t phys = page_to_phys(page) + offset;
851 bool coherent = dev_is_dma_coherent(dev);
852 dma_addr_t dma_handle;
854 dma_handle = __iommu_dma_map_swiotlb(dev, phys, size, dma_get_mask(dev),
855 coherent, dir, attrs);
856 if (!coherent && !(attrs & DMA_ATTR_SKIP_CPU_SYNC) &&
857 dma_handle != DMA_MAPPING_ERROR)
858 arch_sync_dma_for_device(phys, size, dir);
862 static void iommu_dma_unmap_page(struct device *dev, dma_addr_t dma_handle,
863 size_t size, enum dma_data_direction dir, unsigned long attrs)
865 if (!(attrs & DMA_ATTR_SKIP_CPU_SYNC))
866 iommu_dma_sync_single_for_cpu(dev, dma_handle, size, dir);
867 __iommu_dma_unmap_swiotlb(dev, dma_handle, size, dir, attrs);
871 * Prepare a successfully-mapped scatterlist to give back to the caller.
873 * At this point the segments are already laid out by iommu_dma_map_sg() to
874 * avoid individually crossing any boundaries, so we merely need to check a
875 * segment's start address to avoid concatenating across one.
877 static int __finalise_sg(struct device *dev, struct scatterlist *sg, int nents,
880 struct scatterlist *s, *cur = sg;
881 unsigned long seg_mask = dma_get_seg_boundary(dev);
882 unsigned int cur_len = 0, max_len = dma_get_max_seg_size(dev);
885 for_each_sg(sg, s, nents, i) {
886 /* Restore this segment's original unaligned fields first */
887 unsigned int s_iova_off = sg_dma_address(s);
888 unsigned int s_length = sg_dma_len(s);
889 unsigned int s_iova_len = s->length;
891 s->offset += s_iova_off;
892 s->length = s_length;
893 sg_dma_address(s) = DMA_MAPPING_ERROR;
897 * Now fill in the real DMA data. If...
898 * - there is a valid output segment to append to
899 * - and this segment starts on an IOVA page boundary
900 * - but doesn't fall at a segment boundary
901 * - and wouldn't make the resulting output segment too long
903 if (cur_len && !s_iova_off && (dma_addr & seg_mask) &&
904 (max_len - cur_len >= s_length)) {
905 /* ...then concatenate it with the previous one */
908 /* Otherwise start the next output segment */
914 sg_dma_address(cur) = dma_addr + s_iova_off;
917 sg_dma_len(cur) = cur_len;
918 dma_addr += s_iova_len;
920 if (s_length + s_iova_off < s_iova_len)
927 * If mapping failed, then just restore the original list,
928 * but making sure the DMA fields are invalidated.
930 static void __invalidate_sg(struct scatterlist *sg, int nents)
932 struct scatterlist *s;
935 for_each_sg(sg, s, nents, i) {
936 if (sg_dma_address(s) != DMA_MAPPING_ERROR)
937 s->offset += sg_dma_address(s);
939 s->length = sg_dma_len(s);
940 sg_dma_address(s) = DMA_MAPPING_ERROR;
945 static void iommu_dma_unmap_sg_swiotlb(struct device *dev, struct scatterlist *sg,
946 int nents, enum dma_data_direction dir, unsigned long attrs)
948 struct scatterlist *s;
951 for_each_sg(sg, s, nents, i)
952 __iommu_dma_unmap_swiotlb(dev, sg_dma_address(s),
953 sg_dma_len(s), dir, attrs);
956 static int iommu_dma_map_sg_swiotlb(struct device *dev, struct scatterlist *sg,
957 int nents, enum dma_data_direction dir, unsigned long attrs)
959 struct scatterlist *s;
962 for_each_sg(sg, s, nents, i) {
963 sg_dma_address(s) = __iommu_dma_map_swiotlb(dev, sg_phys(s),
964 s->length, dma_get_mask(dev),
965 dev_is_dma_coherent(dev), dir, attrs);
966 if (sg_dma_address(s) == DMA_MAPPING_ERROR)
968 sg_dma_len(s) = s->length;
974 iommu_dma_unmap_sg_swiotlb(dev, sg, i, dir, attrs | DMA_ATTR_SKIP_CPU_SYNC);
979 * The DMA API client is passing in a scatterlist which could describe
980 * any old buffer layout, but the IOMMU API requires everything to be
981 * aligned to IOMMU pages. Hence the need for this complicated bit of
982 * impedance-matching, to be able to hand off a suitably-aligned list,
983 * but still preserve the original offsets and sizes for the caller.
985 static int iommu_dma_map_sg(struct device *dev, struct scatterlist *sg,
986 int nents, enum dma_data_direction dir, unsigned long attrs)
988 struct iommu_domain *domain = iommu_get_dma_domain(dev);
989 struct iommu_dma_cookie *cookie = domain->iova_cookie;
990 struct iova_domain *iovad = &cookie->iovad;
991 struct scatterlist *s, *prev = NULL;
992 int prot = dma_info_to_prot(dir, dev_is_dma_coherent(dev), attrs);
995 unsigned long mask = dma_get_seg_boundary(dev);
998 if (static_branch_unlikely(&iommu_deferred_attach_enabled) &&
999 iommu_deferred_attach(dev, domain))
1002 if (!(attrs & DMA_ATTR_SKIP_CPU_SYNC))
1003 iommu_dma_sync_sg_for_device(dev, sg, nents, dir);
1005 if (dev_is_untrusted(dev))
1006 return iommu_dma_map_sg_swiotlb(dev, sg, nents, dir, attrs);
1009 * Work out how much IOVA space we need, and align the segments to
1010 * IOVA granules for the IOMMU driver to handle. With some clever
1011 * trickery we can modify the list in-place, but reversibly, by
1012 * stashing the unaligned parts in the as-yet-unused DMA fields.
1014 for_each_sg(sg, s, nents, i) {
1015 size_t s_iova_off = iova_offset(iovad, s->offset);
1016 size_t s_length = s->length;
1017 size_t pad_len = (mask - iova_len + 1) & mask;
1019 sg_dma_address(s) = s_iova_off;
1020 sg_dma_len(s) = s_length;
1021 s->offset -= s_iova_off;
1022 s_length = iova_align(iovad, s_length + s_iova_off);
1023 s->length = s_length;
1026 * Due to the alignment of our single IOVA allocation, we can
1027 * depend on these assumptions about the segment boundary mask:
1028 * - If mask size >= IOVA size, then the IOVA range cannot
1029 * possibly fall across a boundary, so we don't care.
1030 * - If mask size < IOVA size, then the IOVA range must start
1031 * exactly on a boundary, therefore we can lay things out
1032 * based purely on segment lengths without needing to know
1033 * the actual addresses beforehand.
1034 * - The mask must be a power of 2, so pad_len == 0 if
1035 * iova_len == 0, thus we cannot dereference prev the first
1036 * time through here (i.e. before it has a meaningful value).
1038 if (pad_len && pad_len < s_length - 1) {
1039 prev->length += pad_len;
1040 iova_len += pad_len;
1043 iova_len += s_length;
1047 iova = iommu_dma_alloc_iova(domain, iova_len, dma_get_mask(dev), dev);
1049 goto out_restore_sg;
1052 * We'll leave any physical concatenation to the IOMMU driver's
1053 * implementation - it knows better than we do.
1055 if (iommu_map_sg_atomic(domain, iova, sg, nents, prot) < iova_len)
1058 return __finalise_sg(dev, sg, nents, iova);
1061 iommu_dma_free_iova(cookie, iova, iova_len, NULL);
1063 __invalidate_sg(sg, nents);
1067 static void iommu_dma_unmap_sg(struct device *dev, struct scatterlist *sg,
1068 int nents, enum dma_data_direction dir, unsigned long attrs)
1070 dma_addr_t start, end;
1071 struct scatterlist *tmp;
1074 if (!(attrs & DMA_ATTR_SKIP_CPU_SYNC))
1075 iommu_dma_sync_sg_for_cpu(dev, sg, nents, dir);
1077 if (dev_is_untrusted(dev)) {
1078 iommu_dma_unmap_sg_swiotlb(dev, sg, nents, dir, attrs);
1083 * The scatterlist segments are mapped into a single
1084 * contiguous IOVA allocation, so this is incredibly easy.
1086 start = sg_dma_address(sg);
1087 for_each_sg(sg_next(sg), tmp, nents - 1, i) {
1088 if (sg_dma_len(tmp) == 0)
1092 end = sg_dma_address(sg) + sg_dma_len(sg);
1093 __iommu_dma_unmap(dev, start, end - start);
1096 static dma_addr_t iommu_dma_map_resource(struct device *dev, phys_addr_t phys,
1097 size_t size, enum dma_data_direction dir, unsigned long attrs)
1099 return __iommu_dma_map(dev, phys, size,
1100 dma_info_to_prot(dir, false, attrs) | IOMMU_MMIO,
1104 static void iommu_dma_unmap_resource(struct device *dev, dma_addr_t handle,
1105 size_t size, enum dma_data_direction dir, unsigned long attrs)
1107 __iommu_dma_unmap(dev, handle, size);
1110 static void __iommu_dma_free(struct device *dev, size_t size, void *cpu_addr)
1112 size_t alloc_size = PAGE_ALIGN(size);
1113 int count = alloc_size >> PAGE_SHIFT;
1114 struct page *page = NULL, **pages = NULL;
1116 /* Non-coherent atomic allocation? Easy */
1117 if (IS_ENABLED(CONFIG_DMA_DIRECT_REMAP) &&
1118 dma_free_from_pool(dev, cpu_addr, alloc_size))
1121 if (IS_ENABLED(CONFIG_DMA_REMAP) && is_vmalloc_addr(cpu_addr)) {
1123 * If it the address is remapped, then it's either non-coherent
1124 * or highmem CMA, or an iommu_dma_alloc_remap() construction.
1126 pages = dma_common_find_pages(cpu_addr);
1128 page = vmalloc_to_page(cpu_addr);
1129 dma_common_free_remap(cpu_addr, alloc_size);
1131 /* Lowmem means a coherent atomic or CMA allocation */
1132 page = virt_to_page(cpu_addr);
1136 __iommu_dma_free_pages(pages, count);
1138 dma_free_contiguous(dev, page, alloc_size);
1141 static void iommu_dma_free(struct device *dev, size_t size, void *cpu_addr,
1142 dma_addr_t handle, unsigned long attrs)
1144 __iommu_dma_unmap(dev, handle, size);
1145 __iommu_dma_free(dev, size, cpu_addr);
1148 static void *iommu_dma_alloc_pages(struct device *dev, size_t size,
1149 struct page **pagep, gfp_t gfp, unsigned long attrs)
1151 bool coherent = dev_is_dma_coherent(dev);
1152 size_t alloc_size = PAGE_ALIGN(size);
1153 int node = dev_to_node(dev);
1154 struct page *page = NULL;
1157 page = dma_alloc_contiguous(dev, alloc_size, gfp);
1159 page = alloc_pages_node(node, gfp, get_order(alloc_size));
1163 if (IS_ENABLED(CONFIG_DMA_REMAP) && (!coherent || PageHighMem(page))) {
1164 pgprot_t prot = dma_pgprot(dev, PAGE_KERNEL, attrs);
1166 cpu_addr = dma_common_contiguous_remap(page, alloc_size,
1167 prot, __builtin_return_address(0));
1169 goto out_free_pages;
1172 arch_dma_prep_coherent(page, size);
1174 cpu_addr = page_address(page);
1178 memset(cpu_addr, 0, alloc_size);
1181 dma_free_contiguous(dev, page, alloc_size);
1185 static void *iommu_dma_alloc(struct device *dev, size_t size,
1186 dma_addr_t *handle, gfp_t gfp, unsigned long attrs)
1188 bool coherent = dev_is_dma_coherent(dev);
1189 int ioprot = dma_info_to_prot(DMA_BIDIRECTIONAL, coherent, attrs);
1190 struct page *page = NULL;
1195 if (IS_ENABLED(CONFIG_DMA_REMAP) && gfpflags_allow_blocking(gfp) &&
1196 !(attrs & DMA_ATTR_FORCE_CONTIGUOUS)) {
1197 return iommu_dma_alloc_remap(dev, size, handle, gfp,
1198 dma_pgprot(dev, PAGE_KERNEL, attrs), attrs);
1201 if (IS_ENABLED(CONFIG_DMA_DIRECT_REMAP) &&
1202 !gfpflags_allow_blocking(gfp) && !coherent)
1203 page = dma_alloc_from_pool(dev, PAGE_ALIGN(size), &cpu_addr,
1206 cpu_addr = iommu_dma_alloc_pages(dev, size, &page, gfp, attrs);
1210 *handle = __iommu_dma_map(dev, page_to_phys(page), size, ioprot,
1211 dev->coherent_dma_mask);
1212 if (*handle == DMA_MAPPING_ERROR) {
1213 __iommu_dma_free(dev, size, cpu_addr);
1220 static int iommu_dma_mmap(struct device *dev, struct vm_area_struct *vma,
1221 void *cpu_addr, dma_addr_t dma_addr, size_t size,
1222 unsigned long attrs)
1224 unsigned long nr_pages = PAGE_ALIGN(size) >> PAGE_SHIFT;
1225 unsigned long pfn, off = vma->vm_pgoff;
1228 vma->vm_page_prot = dma_pgprot(dev, vma->vm_page_prot, attrs);
1230 if (dma_mmap_from_dev_coherent(dev, vma, cpu_addr, size, &ret))
1233 if (off >= nr_pages || vma_pages(vma) > nr_pages - off)
1236 if (IS_ENABLED(CONFIG_DMA_REMAP) && is_vmalloc_addr(cpu_addr)) {
1237 struct page **pages = dma_common_find_pages(cpu_addr);
1240 return vm_map_pages(vma, pages, nr_pages);
1241 pfn = vmalloc_to_pfn(cpu_addr);
1243 pfn = page_to_pfn(virt_to_page(cpu_addr));
1246 return remap_pfn_range(vma, vma->vm_start, pfn + off,
1247 vma->vm_end - vma->vm_start,
1251 static int iommu_dma_get_sgtable(struct device *dev, struct sg_table *sgt,
1252 void *cpu_addr, dma_addr_t dma_addr, size_t size,
1253 unsigned long attrs)
1258 if (IS_ENABLED(CONFIG_DMA_REMAP) && is_vmalloc_addr(cpu_addr)) {
1259 struct page **pages = dma_common_find_pages(cpu_addr);
1262 return sg_alloc_table_from_pages(sgt, pages,
1263 PAGE_ALIGN(size) >> PAGE_SHIFT,
1264 0, size, GFP_KERNEL);
1267 page = vmalloc_to_page(cpu_addr);
1269 page = virt_to_page(cpu_addr);
1272 ret = sg_alloc_table(sgt, 1, GFP_KERNEL);
1274 sg_set_page(sgt->sgl, page, PAGE_ALIGN(size), 0);
1278 static unsigned long iommu_dma_get_merge_boundary(struct device *dev)
1280 struct iommu_domain *domain = iommu_get_dma_domain(dev);
1282 return (1UL << __ffs(domain->pgsize_bitmap)) - 1;
1285 static const struct dma_map_ops iommu_dma_ops = {
1286 .alloc = iommu_dma_alloc,
1287 .free = iommu_dma_free,
1288 .alloc_pages = dma_common_alloc_pages,
1289 .free_pages = dma_common_free_pages,
1290 #ifdef CONFIG_DMA_REMAP
1291 .alloc_noncontiguous = iommu_dma_alloc_noncontiguous,
1292 .free_noncontiguous = iommu_dma_free_noncontiguous,
1294 .mmap = iommu_dma_mmap,
1295 .get_sgtable = iommu_dma_get_sgtable,
1296 .map_page = iommu_dma_map_page,
1297 .unmap_page = iommu_dma_unmap_page,
1298 .map_sg = iommu_dma_map_sg,
1299 .unmap_sg = iommu_dma_unmap_sg,
1300 .sync_single_for_cpu = iommu_dma_sync_single_for_cpu,
1301 .sync_single_for_device = iommu_dma_sync_single_for_device,
1302 .sync_sg_for_cpu = iommu_dma_sync_sg_for_cpu,
1303 .sync_sg_for_device = iommu_dma_sync_sg_for_device,
1304 .map_resource = iommu_dma_map_resource,
1305 .unmap_resource = iommu_dma_unmap_resource,
1306 .get_merge_boundary = iommu_dma_get_merge_boundary,
1310 * The IOMMU core code allocates the default DMA domain, which the underlying
1311 * IOMMU driver needs to support via the dma-iommu layer.
1313 void iommu_setup_dma_ops(struct device *dev, u64 dma_base, u64 dma_limit)
1315 struct iommu_domain *domain = iommu_get_domain_for_dev(dev);
1321 * The IOMMU core code allocates the default DMA domain, which the
1322 * underlying IOMMU driver needs to support via the dma-iommu layer.
1324 if (domain->type == IOMMU_DOMAIN_DMA) {
1325 if (iommu_dma_init_domain(domain, dma_base, dma_limit, dev))
1327 dev->dma_ops = &iommu_dma_ops;
1332 pr_warn("Failed to set up IOMMU for device %s; retaining platform DMA ops\n",
1335 EXPORT_SYMBOL_GPL(iommu_setup_dma_ops);
1337 static struct iommu_dma_msi_page *iommu_dma_get_msi_page(struct device *dev,
1338 phys_addr_t msi_addr, struct iommu_domain *domain)
1340 struct iommu_dma_cookie *cookie = domain->iova_cookie;
1341 struct iommu_dma_msi_page *msi_page;
1343 int prot = IOMMU_WRITE | IOMMU_NOEXEC | IOMMU_MMIO;
1344 size_t size = cookie_msi_granule(cookie);
1346 msi_addr &= ~(phys_addr_t)(size - 1);
1347 list_for_each_entry(msi_page, &cookie->msi_page_list, list)
1348 if (msi_page->phys == msi_addr)
1351 msi_page = kzalloc(sizeof(*msi_page), GFP_KERNEL);
1355 iova = iommu_dma_alloc_iova(domain, size, dma_get_mask(dev), dev);
1359 if (iommu_map(domain, iova, msi_addr, size, prot))
1362 INIT_LIST_HEAD(&msi_page->list);
1363 msi_page->phys = msi_addr;
1364 msi_page->iova = iova;
1365 list_add(&msi_page->list, &cookie->msi_page_list);
1369 iommu_dma_free_iova(cookie, iova, size, NULL);
1375 int iommu_dma_prepare_msi(struct msi_desc *desc, phys_addr_t msi_addr)
1377 struct device *dev = msi_desc_to_dev(desc);
1378 struct iommu_domain *domain = iommu_get_domain_for_dev(dev);
1379 struct iommu_dma_msi_page *msi_page;
1380 static DEFINE_MUTEX(msi_prepare_lock); /* see below */
1382 if (!domain || !domain->iova_cookie) {
1383 desc->iommu_cookie = NULL;
1388 * In fact the whole prepare operation should already be serialised by
1389 * irq_domain_mutex further up the callchain, but that's pretty subtle
1390 * on its own, so consider this locking as failsafe documentation...
1392 mutex_lock(&msi_prepare_lock);
1393 msi_page = iommu_dma_get_msi_page(dev, msi_addr, domain);
1394 mutex_unlock(&msi_prepare_lock);
1396 msi_desc_set_iommu_cookie(desc, msi_page);
1403 void iommu_dma_compose_msi_msg(struct msi_desc *desc,
1404 struct msi_msg *msg)
1406 struct device *dev = msi_desc_to_dev(desc);
1407 const struct iommu_domain *domain = iommu_get_domain_for_dev(dev);
1408 const struct iommu_dma_msi_page *msi_page;
1410 msi_page = msi_desc_get_iommu_cookie(desc);
1412 if (!domain || !domain->iova_cookie || WARN_ON(!msi_page))
1415 msg->address_hi = upper_32_bits(msi_page->iova);
1416 msg->address_lo &= cookie_msi_granule(domain->iova_cookie) - 1;
1417 msg->address_lo += lower_32_bits(msi_page->iova);
1420 static int iommu_dma_init(void)
1422 if (is_kdump_kernel())
1423 static_branch_enable(&iommu_deferred_attach_enabled);
1425 return iova_cache_get();
1427 arch_initcall(iommu_dma_init);