1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright IBM Corp. 2012
9 #include <linux/kernel.h>
10 #include <linux/slab.h>
11 #include <linux/export.h>
12 #include <linux/iommu-helper.h>
13 #include <linux/dma-map-ops.h>
14 #include <linux/vmalloc.h>
15 #include <linux/pci.h>
16 #include <asm/pci_dma.h>
18 static struct kmem_cache *dma_region_table_cache;
19 static struct kmem_cache *dma_page_table_cache;
20 static int s390_iommu_strict;
21 static u64 s390_iommu_aperture;
22 static u32 s390_iommu_aperture_factor = 1;
24 static int zpci_refresh_global(struct zpci_dev *zdev)
26 return zpci_refresh_trans((u64) zdev->fh << 32, zdev->start_dma,
27 zdev->iommu_pages * PAGE_SIZE);
30 unsigned long *dma_alloc_cpu_table(void)
32 unsigned long *table, *entry;
34 table = kmem_cache_alloc(dma_region_table_cache, GFP_ATOMIC);
38 for (entry = table; entry < table + ZPCI_TABLE_ENTRIES; entry++)
39 *entry = ZPCI_TABLE_INVALID;
43 static void dma_free_cpu_table(void *table)
45 kmem_cache_free(dma_region_table_cache, table);
48 static unsigned long *dma_alloc_page_table(void)
50 unsigned long *table, *entry;
52 table = kmem_cache_alloc(dma_page_table_cache, GFP_ATOMIC);
56 for (entry = table; entry < table + ZPCI_PT_ENTRIES; entry++)
57 *entry = ZPCI_PTE_INVALID;
61 static void dma_free_page_table(void *table)
63 kmem_cache_free(dma_page_table_cache, table);
66 static unsigned long *dma_get_seg_table_origin(unsigned long *rtep)
68 unsigned long old_rte, rte;
71 rte = READ_ONCE(*rtep);
72 if (reg_entry_isvalid(rte)) {
73 sto = get_rt_sto(rte);
75 sto = dma_alloc_cpu_table();
79 set_rt_sto(&rte, virt_to_phys(sto));
80 validate_rt_entry(&rte);
81 entry_clr_protected(&rte);
83 old_rte = cmpxchg(rtep, ZPCI_TABLE_INVALID, rte);
84 if (old_rte != ZPCI_TABLE_INVALID) {
85 /* Somone else was faster, use theirs */
86 dma_free_cpu_table(sto);
87 sto = get_rt_sto(old_rte);
93 static unsigned long *dma_get_page_table_origin(unsigned long *step)
95 unsigned long old_ste, ste;
98 ste = READ_ONCE(*step);
99 if (reg_entry_isvalid(ste)) {
100 pto = get_st_pto(ste);
102 pto = dma_alloc_page_table();
105 set_st_pto(&ste, virt_to_phys(pto));
106 validate_st_entry(&ste);
107 entry_clr_protected(&ste);
109 old_ste = cmpxchg(step, ZPCI_TABLE_INVALID, ste);
110 if (old_ste != ZPCI_TABLE_INVALID) {
111 /* Somone else was faster, use theirs */
112 dma_free_page_table(pto);
113 pto = get_st_pto(old_ste);
119 unsigned long *dma_walk_cpu_trans(unsigned long *rto, dma_addr_t dma_addr)
121 unsigned long *sto, *pto;
122 unsigned int rtx, sx, px;
124 rtx = calc_rtx(dma_addr);
125 sto = dma_get_seg_table_origin(&rto[rtx]);
129 sx = calc_sx(dma_addr);
130 pto = dma_get_page_table_origin(&sto[sx]);
134 px = calc_px(dma_addr);
138 void dma_update_cpu_trans(unsigned long *ptep, phys_addr_t page_addr, int flags)
142 pte = READ_ONCE(*ptep);
143 if (flags & ZPCI_PTE_INVALID) {
144 invalidate_pt_entry(&pte);
146 set_pt_pfaa(&pte, page_addr);
147 validate_pt_entry(&pte);
150 if (flags & ZPCI_TABLE_PROTECTED)
151 entry_set_protected(&pte);
153 entry_clr_protected(&pte);
158 static int __dma_update_trans(struct zpci_dev *zdev, phys_addr_t pa,
159 dma_addr_t dma_addr, size_t size, int flags)
161 unsigned int nr_pages = PAGE_ALIGN(size) >> PAGE_SHIFT;
162 phys_addr_t page_addr = (pa & PAGE_MASK);
163 unsigned long *entry;
169 if (!zdev->dma_table)
172 for (i = 0; i < nr_pages; i++) {
173 entry = dma_walk_cpu_trans(zdev->dma_table, dma_addr);
178 dma_update_cpu_trans(entry, page_addr, flags);
179 page_addr += PAGE_SIZE;
180 dma_addr += PAGE_SIZE;
184 if (rc && ((flags & ZPCI_PTE_VALID_MASK) == ZPCI_PTE_VALID)) {
185 flags = ZPCI_PTE_INVALID;
187 page_addr -= PAGE_SIZE;
188 dma_addr -= PAGE_SIZE;
189 entry = dma_walk_cpu_trans(zdev->dma_table, dma_addr);
192 dma_update_cpu_trans(entry, page_addr, flags);
198 static int __dma_purge_tlb(struct zpci_dev *zdev, dma_addr_t dma_addr,
199 size_t size, int flags)
201 unsigned long irqflags;
205 * With zdev->tlb_refresh == 0, rpcit is not required to establish new
206 * translations when previously invalid translation-table entries are
207 * validated. With lazy unmap, rpcit is skipped for previously valid
208 * entries, but a global rpcit is then required before any address can
209 * be re-used, i.e. after each iommu bitmap wrap-around.
211 if ((flags & ZPCI_PTE_VALID_MASK) == ZPCI_PTE_VALID) {
212 if (!zdev->tlb_refresh)
215 if (!s390_iommu_strict)
219 ret = zpci_refresh_trans((u64) zdev->fh << 32, dma_addr,
221 if (ret == -ENOMEM && !s390_iommu_strict) {
222 /* enable the hypervisor to free some resources */
223 if (zpci_refresh_global(zdev))
226 spin_lock_irqsave(&zdev->iommu_bitmap_lock, irqflags);
227 bitmap_andnot(zdev->iommu_bitmap, zdev->iommu_bitmap,
228 zdev->lazy_bitmap, zdev->iommu_pages);
229 bitmap_zero(zdev->lazy_bitmap, zdev->iommu_pages);
230 spin_unlock_irqrestore(&zdev->iommu_bitmap_lock, irqflags);
237 static int dma_update_trans(struct zpci_dev *zdev, phys_addr_t pa,
238 dma_addr_t dma_addr, size_t size, int flags)
242 rc = __dma_update_trans(zdev, pa, dma_addr, size, flags);
246 rc = __dma_purge_tlb(zdev, dma_addr, size, flags);
247 if (rc && ((flags & ZPCI_PTE_VALID_MASK) == ZPCI_PTE_VALID))
248 __dma_update_trans(zdev, pa, dma_addr, size, ZPCI_PTE_INVALID);
253 void dma_free_seg_table(unsigned long entry)
255 unsigned long *sto = get_rt_sto(entry);
258 for (sx = 0; sx < ZPCI_TABLE_ENTRIES; sx++)
259 if (reg_entry_isvalid(sto[sx]))
260 dma_free_page_table(get_st_pto(sto[sx]));
262 dma_free_cpu_table(sto);
265 void dma_cleanup_tables(unsigned long *table)
272 for (rtx = 0; rtx < ZPCI_TABLE_ENTRIES; rtx++)
273 if (reg_entry_isvalid(table[rtx]))
274 dma_free_seg_table(table[rtx]);
276 dma_free_cpu_table(table);
279 static unsigned long __dma_alloc_iommu(struct device *dev,
280 unsigned long start, int size)
282 struct zpci_dev *zdev = to_zpci(to_pci_dev(dev));
284 return iommu_area_alloc(zdev->iommu_bitmap, zdev->iommu_pages,
285 start, size, zdev->start_dma >> PAGE_SHIFT,
286 dma_get_seg_boundary_nr_pages(dev, PAGE_SHIFT),
290 static dma_addr_t dma_alloc_address(struct device *dev, int size)
292 struct zpci_dev *zdev = to_zpci(to_pci_dev(dev));
293 unsigned long offset, flags;
295 spin_lock_irqsave(&zdev->iommu_bitmap_lock, flags);
296 offset = __dma_alloc_iommu(dev, zdev->next_bit, size);
298 if (!s390_iommu_strict) {
299 /* global flush before DMA addresses are reused */
300 if (zpci_refresh_global(zdev))
303 bitmap_andnot(zdev->iommu_bitmap, zdev->iommu_bitmap,
304 zdev->lazy_bitmap, zdev->iommu_pages);
305 bitmap_zero(zdev->lazy_bitmap, zdev->iommu_pages);
308 offset = __dma_alloc_iommu(dev, 0, size);
312 zdev->next_bit = offset + size;
313 spin_unlock_irqrestore(&zdev->iommu_bitmap_lock, flags);
315 return zdev->start_dma + offset * PAGE_SIZE;
318 spin_unlock_irqrestore(&zdev->iommu_bitmap_lock, flags);
319 return DMA_MAPPING_ERROR;
322 static void dma_free_address(struct device *dev, dma_addr_t dma_addr, int size)
324 struct zpci_dev *zdev = to_zpci(to_pci_dev(dev));
325 unsigned long flags, offset;
327 offset = (dma_addr - zdev->start_dma) >> PAGE_SHIFT;
329 spin_lock_irqsave(&zdev->iommu_bitmap_lock, flags);
330 if (!zdev->iommu_bitmap)
333 if (s390_iommu_strict)
334 bitmap_clear(zdev->iommu_bitmap, offset, size);
336 bitmap_set(zdev->lazy_bitmap, offset, size);
339 spin_unlock_irqrestore(&zdev->iommu_bitmap_lock, flags);
342 static inline void zpci_err_dma(unsigned long rc, unsigned long addr)
347 } __packed data = {rc, addr};
349 zpci_err_hex(&data, sizeof(data));
352 static dma_addr_t s390_dma_map_pages(struct device *dev, struct page *page,
353 unsigned long offset, size_t size,
354 enum dma_data_direction direction,
357 struct zpci_dev *zdev = to_zpci(to_pci_dev(dev));
358 unsigned long pa = page_to_phys(page) + offset;
359 int flags = ZPCI_PTE_VALID;
360 unsigned long nr_pages;
364 /* This rounds up number of pages based on size and offset */
365 nr_pages = iommu_num_pages(pa, size, PAGE_SIZE);
366 dma_addr = dma_alloc_address(dev, nr_pages);
367 if (dma_addr == DMA_MAPPING_ERROR) {
372 /* Use rounded up size */
373 size = nr_pages * PAGE_SIZE;
375 if (direction == DMA_NONE || direction == DMA_TO_DEVICE)
376 flags |= ZPCI_TABLE_PROTECTED;
378 ret = dma_update_trans(zdev, pa, dma_addr, size, flags);
382 atomic64_add(nr_pages, &zdev->mapped_pages);
383 return dma_addr + (offset & ~PAGE_MASK);
386 dma_free_address(dev, dma_addr, nr_pages);
388 zpci_err("map error:\n");
389 zpci_err_dma(ret, pa);
390 return DMA_MAPPING_ERROR;
393 static void s390_dma_unmap_pages(struct device *dev, dma_addr_t dma_addr,
394 size_t size, enum dma_data_direction direction,
397 struct zpci_dev *zdev = to_zpci(to_pci_dev(dev));
400 npages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
401 dma_addr = dma_addr & PAGE_MASK;
402 ret = dma_update_trans(zdev, 0, dma_addr, npages * PAGE_SIZE,
405 zpci_err("unmap error:\n");
406 zpci_err_dma(ret, dma_addr);
410 atomic64_add(npages, &zdev->unmapped_pages);
411 dma_free_address(dev, dma_addr, npages);
414 static void *s390_dma_alloc(struct device *dev, size_t size,
415 dma_addr_t *dma_handle, gfp_t flag,
418 struct zpci_dev *zdev = to_zpci(to_pci_dev(dev));
423 size = PAGE_ALIGN(size);
424 page = alloc_pages(flag | __GFP_ZERO, get_order(size));
428 pa = page_to_phys(page);
429 map = s390_dma_map_pages(dev, page, 0, size, DMA_BIDIRECTIONAL, 0);
430 if (dma_mapping_error(dev, map)) {
431 __free_pages(page, get_order(size));
435 atomic64_add(size / PAGE_SIZE, &zdev->allocated_pages);
438 return phys_to_virt(pa);
441 static void s390_dma_free(struct device *dev, size_t size,
442 void *vaddr, dma_addr_t dma_handle,
445 struct zpci_dev *zdev = to_zpci(to_pci_dev(dev));
447 size = PAGE_ALIGN(size);
448 atomic64_sub(size / PAGE_SIZE, &zdev->allocated_pages);
449 s390_dma_unmap_pages(dev, dma_handle, size, DMA_BIDIRECTIONAL, 0);
450 free_pages((unsigned long)vaddr, get_order(size));
453 /* Map a segment into a contiguous dma address area */
454 static int __s390_dma_map_sg(struct device *dev, struct scatterlist *sg,
455 size_t size, dma_addr_t *handle,
456 enum dma_data_direction dir)
458 unsigned long nr_pages = PAGE_ALIGN(size) >> PAGE_SHIFT;
459 struct zpci_dev *zdev = to_zpci(to_pci_dev(dev));
460 dma_addr_t dma_addr_base, dma_addr;
461 int flags = ZPCI_PTE_VALID;
462 struct scatterlist *s;
466 dma_addr_base = dma_alloc_address(dev, nr_pages);
467 if (dma_addr_base == DMA_MAPPING_ERROR)
470 dma_addr = dma_addr_base;
471 if (dir == DMA_NONE || dir == DMA_TO_DEVICE)
472 flags |= ZPCI_TABLE_PROTECTED;
474 for (s = sg; dma_addr < dma_addr_base + size; s = sg_next(s)) {
475 pa = page_to_phys(sg_page(s));
476 ret = __dma_update_trans(zdev, pa, dma_addr,
477 s->offset + s->length, flags);
481 dma_addr += s->offset + s->length;
483 ret = __dma_purge_tlb(zdev, dma_addr_base, size, flags);
487 *handle = dma_addr_base;
488 atomic64_add(nr_pages, &zdev->mapped_pages);
493 dma_update_trans(zdev, 0, dma_addr_base, dma_addr - dma_addr_base,
495 dma_free_address(dev, dma_addr_base, nr_pages);
496 zpci_err("map error:\n");
497 zpci_err_dma(ret, pa);
501 static int s390_dma_map_sg(struct device *dev, struct scatterlist *sg,
502 int nr_elements, enum dma_data_direction dir,
505 struct scatterlist *s = sg, *start = sg, *dma = sg;
506 unsigned int max = dma_get_max_seg_size(dev);
507 unsigned int size = s->offset + s->length;
508 unsigned int offset = s->offset;
509 int count = 0, i, ret;
511 for (i = 1; i < nr_elements; i++) {
516 if (s->offset || (size & ~PAGE_MASK) ||
517 size + s->length > max) {
518 ret = __s390_dma_map_sg(dev, start, size,
519 &dma->dma_address, dir);
523 dma->dma_address += offset;
524 dma->dma_length = size - offset;
526 size = offset = s->offset;
533 ret = __s390_dma_map_sg(dev, start, size, &dma->dma_address, dir);
537 dma->dma_address += offset;
538 dma->dma_length = size - offset;
542 for_each_sg(sg, s, count, i)
543 s390_dma_unmap_pages(dev, sg_dma_address(s), sg_dma_len(s),
549 static void s390_dma_unmap_sg(struct device *dev, struct scatterlist *sg,
550 int nr_elements, enum dma_data_direction dir,
553 struct scatterlist *s;
556 for_each_sg(sg, s, nr_elements, i) {
558 s390_dma_unmap_pages(dev, s->dma_address, s->dma_length,
565 int zpci_dma_init_device(struct zpci_dev *zdev)
571 * At this point, if the device is part of an IOMMU domain, this would
572 * be a strong hint towards a bug in the IOMMU API (common) code and/or
573 * simultaneous access via IOMMU and DMA API. So let's issue a warning.
575 WARN_ON(zdev->s390_domain);
577 spin_lock_init(&zdev->iommu_bitmap_lock);
579 zdev->dma_table = dma_alloc_cpu_table();
580 if (!zdev->dma_table) {
586 * Restrict the iommu bitmap size to the minimum of the following:
587 * - s390_iommu_aperture which defaults to high_memory
588 * - 3-level pagetable address limit minus start_dma offset
589 * - DMA address range allowed by the hardware (clp query pci fn)
591 * Also set zdev->end_dma to the actual end address of the usable
592 * range, instead of the theoretical maximum as reported by hardware.
594 * This limits the number of concurrently usable DMA mappings since
595 * for each DMA mapped memory address we need a DMA address including
596 * extra DMA addresses for multiple mappings of the same memory address.
598 zdev->start_dma = PAGE_ALIGN(zdev->start_dma);
599 zdev->iommu_size = min3(s390_iommu_aperture,
600 ZPCI_TABLE_SIZE_RT - zdev->start_dma,
601 zdev->end_dma - zdev->start_dma + 1);
602 zdev->end_dma = zdev->start_dma + zdev->iommu_size - 1;
603 zdev->iommu_pages = zdev->iommu_size >> PAGE_SHIFT;
604 zdev->iommu_bitmap = vzalloc(zdev->iommu_pages / 8);
605 if (!zdev->iommu_bitmap) {
609 if (!s390_iommu_strict) {
610 zdev->lazy_bitmap = vzalloc(zdev->iommu_pages / 8);
611 if (!zdev->lazy_bitmap) {
617 if (zpci_register_ioat(zdev, 0, zdev->start_dma, zdev->end_dma,
618 virt_to_phys(zdev->dma_table), &status)) {
625 vfree(zdev->iommu_bitmap);
626 zdev->iommu_bitmap = NULL;
627 vfree(zdev->lazy_bitmap);
628 zdev->lazy_bitmap = NULL;
630 dma_free_cpu_table(zdev->dma_table);
631 zdev->dma_table = NULL;
636 int zpci_dma_exit_device(struct zpci_dev *zdev)
641 * At this point, if the device is part of an IOMMU domain, this would
642 * be a strong hint towards a bug in the IOMMU API (common) code and/or
643 * simultaneous access via IOMMU and DMA API. So let's issue a warning.
645 WARN_ON(zdev->s390_domain);
646 if (zdev_enabled(zdev))
647 cc = zpci_unregister_ioat(zdev, 0);
649 * cc == 3 indicates the function is gone already. This can happen
650 * if the function was deconfigured/disabled suddenly and we have not
651 * received a new handle yet.
656 dma_cleanup_tables(zdev->dma_table);
657 zdev->dma_table = NULL;
658 vfree(zdev->iommu_bitmap);
659 zdev->iommu_bitmap = NULL;
660 vfree(zdev->lazy_bitmap);
661 zdev->lazy_bitmap = NULL;
666 static int __init dma_alloc_cpu_table_caches(void)
668 dma_region_table_cache = kmem_cache_create("PCI_DMA_region_tables",
669 ZPCI_TABLE_SIZE, ZPCI_TABLE_ALIGN,
671 if (!dma_region_table_cache)
674 dma_page_table_cache = kmem_cache_create("PCI_DMA_page_tables",
675 ZPCI_PT_SIZE, ZPCI_PT_ALIGN,
677 if (!dma_page_table_cache) {
678 kmem_cache_destroy(dma_region_table_cache);
684 int __init zpci_dma_init(void)
686 s390_iommu_aperture = (u64)virt_to_phys(high_memory);
687 if (!s390_iommu_aperture_factor)
688 s390_iommu_aperture = ULONG_MAX;
690 s390_iommu_aperture *= s390_iommu_aperture_factor;
692 return dma_alloc_cpu_table_caches();
695 void zpci_dma_exit(void)
697 kmem_cache_destroy(dma_page_table_cache);
698 kmem_cache_destroy(dma_region_table_cache);
701 const struct dma_map_ops s390_pci_dma_ops = {
702 .alloc = s390_dma_alloc,
703 .free = s390_dma_free,
704 .map_sg = s390_dma_map_sg,
705 .unmap_sg = s390_dma_unmap_sg,
706 .map_page = s390_dma_map_pages,
707 .unmap_page = s390_dma_unmap_pages,
708 .mmap = dma_common_mmap,
709 .get_sgtable = dma_common_get_sgtable,
710 .alloc_pages = dma_common_alloc_pages,
711 .free_pages = dma_common_free_pages,
712 /* dma_supported is unconditionally true without a callback */
714 EXPORT_SYMBOL_GPL(s390_pci_dma_ops);
716 static int __init s390_iommu_setup(char *str)
718 if (!strcmp(str, "strict"))
719 s390_iommu_strict = 1;
723 __setup("s390_iommu=", s390_iommu_setup);
725 static int __init s390_iommu_aperture_setup(char *str)
727 if (kstrtou32(str, 10, &s390_iommu_aperture_factor))
728 s390_iommu_aperture_factor = 1;
732 __setup("s390_iommu_aperture=", s390_iommu_aperture_setup);