1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright IBM Corp. 2012
6 * Jan Glauber <jang@linux.vnet.ibm.com>
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-mapping.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;
22 static int zpci_refresh_global(struct zpci_dev *zdev)
24 return zpci_refresh_trans((u64) zdev->fh << 32, zdev->start_dma,
25 zdev->iommu_pages * PAGE_SIZE);
28 unsigned long *dma_alloc_cpu_table(void)
30 unsigned long *table, *entry;
32 table = kmem_cache_alloc(dma_region_table_cache, GFP_ATOMIC);
36 for (entry = table; entry < table + ZPCI_TABLE_ENTRIES; entry++)
37 *entry = ZPCI_TABLE_INVALID;
41 static void dma_free_cpu_table(void *table)
43 kmem_cache_free(dma_region_table_cache, table);
46 static unsigned long *dma_alloc_page_table(void)
48 unsigned long *table, *entry;
50 table = kmem_cache_alloc(dma_page_table_cache, GFP_ATOMIC);
54 for (entry = table; entry < table + ZPCI_PT_ENTRIES; entry++)
55 *entry = ZPCI_PTE_INVALID;
59 static void dma_free_page_table(void *table)
61 kmem_cache_free(dma_page_table_cache, table);
64 static unsigned long *dma_get_seg_table_origin(unsigned long *entry)
68 if (reg_entry_isvalid(*entry))
69 sto = get_rt_sto(*entry);
71 sto = dma_alloc_cpu_table();
75 set_rt_sto(entry, sto);
76 validate_rt_entry(entry);
77 entry_clr_protected(entry);
82 static unsigned long *dma_get_page_table_origin(unsigned long *entry)
86 if (reg_entry_isvalid(*entry))
87 pto = get_st_pto(*entry);
89 pto = dma_alloc_page_table();
92 set_st_pto(entry, pto);
93 validate_st_entry(entry);
94 entry_clr_protected(entry);
99 unsigned long *dma_walk_cpu_trans(unsigned long *rto, dma_addr_t dma_addr)
101 unsigned long *sto, *pto;
102 unsigned int rtx, sx, px;
104 rtx = calc_rtx(dma_addr);
105 sto = dma_get_seg_table_origin(&rto[rtx]);
109 sx = calc_sx(dma_addr);
110 pto = dma_get_page_table_origin(&sto[sx]);
114 px = calc_px(dma_addr);
118 void dma_update_cpu_trans(unsigned long *entry, void *page_addr, int flags)
120 if (flags & ZPCI_PTE_INVALID) {
121 invalidate_pt_entry(entry);
123 set_pt_pfaa(entry, page_addr);
124 validate_pt_entry(entry);
127 if (flags & ZPCI_TABLE_PROTECTED)
128 entry_set_protected(entry);
130 entry_clr_protected(entry);
133 static int __dma_update_trans(struct zpci_dev *zdev, unsigned long pa,
134 dma_addr_t dma_addr, size_t size, int flags)
136 unsigned int nr_pages = PAGE_ALIGN(size) >> PAGE_SHIFT;
137 u8 *page_addr = (u8 *) (pa & PAGE_MASK);
138 unsigned long irq_flags;
139 unsigned long *entry;
145 spin_lock_irqsave(&zdev->dma_table_lock, irq_flags);
146 if (!zdev->dma_table) {
151 for (i = 0; i < nr_pages; i++) {
152 entry = dma_walk_cpu_trans(zdev->dma_table, dma_addr);
157 dma_update_cpu_trans(entry, page_addr, flags);
158 page_addr += PAGE_SIZE;
159 dma_addr += PAGE_SIZE;
163 if (rc && ((flags & ZPCI_PTE_VALID_MASK) == ZPCI_PTE_VALID)) {
164 flags = ZPCI_PTE_INVALID;
166 page_addr -= PAGE_SIZE;
167 dma_addr -= PAGE_SIZE;
168 entry = dma_walk_cpu_trans(zdev->dma_table, dma_addr);
171 dma_update_cpu_trans(entry, page_addr, flags);
175 spin_unlock_irqrestore(&zdev->dma_table_lock, irq_flags);
179 static int __dma_purge_tlb(struct zpci_dev *zdev, dma_addr_t dma_addr,
180 size_t size, int flags)
182 unsigned long irqflags;
186 * With zdev->tlb_refresh == 0, rpcit is not required to establish new
187 * translations when previously invalid translation-table entries are
188 * validated. With lazy unmap, rpcit is skipped for previously valid
189 * entries, but a global rpcit is then required before any address can
190 * be re-used, i.e. after each iommu bitmap wrap-around.
192 if ((flags & ZPCI_PTE_VALID_MASK) == ZPCI_PTE_VALID) {
193 if (!zdev->tlb_refresh)
196 if (!s390_iommu_strict)
200 ret = zpci_refresh_trans((u64) zdev->fh << 32, dma_addr,
202 if (ret == -ENOMEM && !s390_iommu_strict) {
203 /* enable the hypervisor to free some resources */
204 if (zpci_refresh_global(zdev))
207 spin_lock_irqsave(&zdev->iommu_bitmap_lock, irqflags);
208 bitmap_andnot(zdev->iommu_bitmap, zdev->iommu_bitmap,
209 zdev->lazy_bitmap, zdev->iommu_pages);
210 bitmap_zero(zdev->lazy_bitmap, zdev->iommu_pages);
211 spin_unlock_irqrestore(&zdev->iommu_bitmap_lock, irqflags);
218 static int dma_update_trans(struct zpci_dev *zdev, unsigned long pa,
219 dma_addr_t dma_addr, size_t size, int flags)
223 rc = __dma_update_trans(zdev, pa, dma_addr, size, flags);
227 rc = __dma_purge_tlb(zdev, dma_addr, size, flags);
228 if (rc && ((flags & ZPCI_PTE_VALID_MASK) == ZPCI_PTE_VALID))
229 __dma_update_trans(zdev, pa, dma_addr, size, ZPCI_PTE_INVALID);
234 void dma_free_seg_table(unsigned long entry)
236 unsigned long *sto = get_rt_sto(entry);
239 for (sx = 0; sx < ZPCI_TABLE_ENTRIES; sx++)
240 if (reg_entry_isvalid(sto[sx]))
241 dma_free_page_table(get_st_pto(sto[sx]));
243 dma_free_cpu_table(sto);
246 void dma_cleanup_tables(unsigned long *table)
253 for (rtx = 0; rtx < ZPCI_TABLE_ENTRIES; rtx++)
254 if (reg_entry_isvalid(table[rtx]))
255 dma_free_seg_table(table[rtx]);
257 dma_free_cpu_table(table);
260 static unsigned long __dma_alloc_iommu(struct device *dev,
261 unsigned long start, int size)
263 struct zpci_dev *zdev = to_zpci(to_pci_dev(dev));
264 unsigned long boundary_size;
266 boundary_size = ALIGN(dma_get_seg_boundary(dev) + 1,
267 PAGE_SIZE) >> PAGE_SHIFT;
268 return iommu_area_alloc(zdev->iommu_bitmap, zdev->iommu_pages,
269 start, size, zdev->start_dma >> PAGE_SHIFT,
273 static dma_addr_t dma_alloc_address(struct device *dev, int size)
275 struct zpci_dev *zdev = to_zpci(to_pci_dev(dev));
276 unsigned long offset, flags;
278 spin_lock_irqsave(&zdev->iommu_bitmap_lock, flags);
279 offset = __dma_alloc_iommu(dev, zdev->next_bit, size);
281 if (!s390_iommu_strict) {
282 /* global flush before DMA addresses are reused */
283 if (zpci_refresh_global(zdev))
286 bitmap_andnot(zdev->iommu_bitmap, zdev->iommu_bitmap,
287 zdev->lazy_bitmap, zdev->iommu_pages);
288 bitmap_zero(zdev->lazy_bitmap, zdev->iommu_pages);
291 offset = __dma_alloc_iommu(dev, 0, size);
295 zdev->next_bit = offset + size;
296 spin_unlock_irqrestore(&zdev->iommu_bitmap_lock, flags);
298 return zdev->start_dma + offset * PAGE_SIZE;
301 spin_unlock_irqrestore(&zdev->iommu_bitmap_lock, flags);
302 return DMA_MAPPING_ERROR;
305 static void dma_free_address(struct device *dev, dma_addr_t dma_addr, int size)
307 struct zpci_dev *zdev = to_zpci(to_pci_dev(dev));
308 unsigned long flags, offset;
310 offset = (dma_addr - zdev->start_dma) >> PAGE_SHIFT;
312 spin_lock_irqsave(&zdev->iommu_bitmap_lock, flags);
313 if (!zdev->iommu_bitmap)
316 if (s390_iommu_strict)
317 bitmap_clear(zdev->iommu_bitmap, offset, size);
319 bitmap_set(zdev->lazy_bitmap, offset, size);
322 spin_unlock_irqrestore(&zdev->iommu_bitmap_lock, flags);
325 static inline void zpci_err_dma(unsigned long rc, unsigned long addr)
330 } __packed data = {rc, addr};
332 zpci_err_hex(&data, sizeof(data));
335 static dma_addr_t s390_dma_map_pages(struct device *dev, struct page *page,
336 unsigned long offset, size_t size,
337 enum dma_data_direction direction,
340 struct zpci_dev *zdev = to_zpci(to_pci_dev(dev));
341 unsigned long pa = page_to_phys(page) + offset;
342 int flags = ZPCI_PTE_VALID;
343 unsigned long nr_pages;
347 /* This rounds up number of pages based on size and offset */
348 nr_pages = iommu_num_pages(pa, size, PAGE_SIZE);
349 dma_addr = dma_alloc_address(dev, nr_pages);
350 if (dma_addr == DMA_MAPPING_ERROR) {
355 /* Use rounded up size */
356 size = nr_pages * PAGE_SIZE;
358 if (direction == DMA_NONE || direction == DMA_TO_DEVICE)
359 flags |= ZPCI_TABLE_PROTECTED;
361 ret = dma_update_trans(zdev, pa, dma_addr, size, flags);
365 atomic64_add(nr_pages, &zdev->mapped_pages);
366 return dma_addr + (offset & ~PAGE_MASK);
369 dma_free_address(dev, dma_addr, nr_pages);
371 zpci_err("map error:\n");
372 zpci_err_dma(ret, pa);
373 return DMA_MAPPING_ERROR;
376 static void s390_dma_unmap_pages(struct device *dev, dma_addr_t dma_addr,
377 size_t size, enum dma_data_direction direction,
380 struct zpci_dev *zdev = to_zpci(to_pci_dev(dev));
383 npages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
384 dma_addr = dma_addr & PAGE_MASK;
385 ret = dma_update_trans(zdev, 0, dma_addr, npages * PAGE_SIZE,
388 zpci_err("unmap error:\n");
389 zpci_err_dma(ret, dma_addr);
393 atomic64_add(npages, &zdev->unmapped_pages);
394 dma_free_address(dev, dma_addr, npages);
397 static void *s390_dma_alloc(struct device *dev, size_t size,
398 dma_addr_t *dma_handle, gfp_t flag,
401 struct zpci_dev *zdev = to_zpci(to_pci_dev(dev));
406 size = PAGE_ALIGN(size);
407 page = alloc_pages(flag | __GFP_ZERO, get_order(size));
411 pa = page_to_phys(page);
412 map = s390_dma_map_pages(dev, page, 0, size, DMA_BIDIRECTIONAL, 0);
413 if (dma_mapping_error(dev, map)) {
414 free_pages(pa, get_order(size));
418 atomic64_add(size / PAGE_SIZE, &zdev->allocated_pages);
424 static void s390_dma_free(struct device *dev, size_t size,
425 void *pa, dma_addr_t dma_handle,
428 struct zpci_dev *zdev = to_zpci(to_pci_dev(dev));
430 size = PAGE_ALIGN(size);
431 atomic64_sub(size / PAGE_SIZE, &zdev->allocated_pages);
432 s390_dma_unmap_pages(dev, dma_handle, size, DMA_BIDIRECTIONAL, 0);
433 free_pages((unsigned long) pa, get_order(size));
436 /* Map a segment into a contiguous dma address area */
437 static int __s390_dma_map_sg(struct device *dev, struct scatterlist *sg,
438 size_t size, dma_addr_t *handle,
439 enum dma_data_direction dir)
441 unsigned long nr_pages = PAGE_ALIGN(size) >> PAGE_SHIFT;
442 struct zpci_dev *zdev = to_zpci(to_pci_dev(dev));
443 dma_addr_t dma_addr_base, dma_addr;
444 int flags = ZPCI_PTE_VALID;
445 struct scatterlist *s;
446 unsigned long pa = 0;
449 dma_addr_base = dma_alloc_address(dev, nr_pages);
450 if (dma_addr_base == DMA_MAPPING_ERROR)
453 dma_addr = dma_addr_base;
454 if (dir == DMA_NONE || dir == DMA_TO_DEVICE)
455 flags |= ZPCI_TABLE_PROTECTED;
457 for (s = sg; dma_addr < dma_addr_base + size; s = sg_next(s)) {
458 pa = page_to_phys(sg_page(s));
459 ret = __dma_update_trans(zdev, pa, dma_addr,
460 s->offset + s->length, flags);
464 dma_addr += s->offset + s->length;
466 ret = __dma_purge_tlb(zdev, dma_addr_base, size, flags);
470 *handle = dma_addr_base;
471 atomic64_add(nr_pages, &zdev->mapped_pages);
476 dma_update_trans(zdev, 0, dma_addr_base, dma_addr - dma_addr_base,
478 dma_free_address(dev, dma_addr_base, nr_pages);
479 zpci_err("map error:\n");
480 zpci_err_dma(ret, pa);
484 static int s390_dma_map_sg(struct device *dev, struct scatterlist *sg,
485 int nr_elements, enum dma_data_direction dir,
488 struct scatterlist *s = sg, *start = sg, *dma = sg;
489 unsigned int max = dma_get_max_seg_size(dev);
490 unsigned int size = s->offset + s->length;
491 unsigned int offset = s->offset;
494 for (i = 1; i < nr_elements; i++) {
497 s->dma_address = DMA_MAPPING_ERROR;
500 if (s->offset || (size & ~PAGE_MASK) ||
501 size + s->length > max) {
502 if (__s390_dma_map_sg(dev, start, size,
503 &dma->dma_address, dir))
506 dma->dma_address += offset;
507 dma->dma_length = size - offset;
509 size = offset = s->offset;
516 if (__s390_dma_map_sg(dev, start, size, &dma->dma_address, dir))
519 dma->dma_address += offset;
520 dma->dma_length = size - offset;
524 for_each_sg(sg, s, count, i)
525 s390_dma_unmap_pages(dev, sg_dma_address(s), sg_dma_len(s),
531 static void s390_dma_unmap_sg(struct device *dev, struct scatterlist *sg,
532 int nr_elements, enum dma_data_direction dir,
535 struct scatterlist *s;
538 for_each_sg(sg, s, nr_elements, i) {
540 s390_dma_unmap_pages(dev, s->dma_address, s->dma_length,
547 static unsigned long *bitmap_vzalloc(size_t bits, gfp_t flags)
549 size_t n = BITS_TO_LONGS(bits);
552 if (unlikely(check_mul_overflow(n, sizeof(unsigned long), &bytes)))
555 return vzalloc(bytes);
558 int zpci_dma_init_device(struct zpci_dev *zdev)
563 * At this point, if the device is part of an IOMMU domain, this would
564 * be a strong hint towards a bug in the IOMMU API (common) code and/or
565 * simultaneous access via IOMMU and DMA API. So let's issue a warning.
567 WARN_ON(zdev->s390_domain);
569 spin_lock_init(&zdev->iommu_bitmap_lock);
570 spin_lock_init(&zdev->dma_table_lock);
572 zdev->dma_table = dma_alloc_cpu_table();
573 if (!zdev->dma_table) {
579 * Restrict the iommu bitmap size to the minimum of the following:
581 * - 3-level pagetable address limit minus start_dma offset
582 * - DMA address range allowed by the hardware (clp query pci fn)
584 * Also set zdev->end_dma to the actual end address of the usable
585 * range, instead of the theoretical maximum as reported by hardware.
587 zdev->start_dma = PAGE_ALIGN(zdev->start_dma);
588 zdev->iommu_size = min3((u64) high_memory,
589 ZPCI_TABLE_SIZE_RT - zdev->start_dma,
590 zdev->end_dma - zdev->start_dma + 1);
591 zdev->end_dma = zdev->start_dma + zdev->iommu_size - 1;
592 zdev->iommu_pages = zdev->iommu_size >> PAGE_SHIFT;
593 zdev->iommu_bitmap = bitmap_vzalloc(zdev->iommu_pages, GFP_KERNEL);
594 if (!zdev->iommu_bitmap) {
598 if (!s390_iommu_strict) {
599 zdev->lazy_bitmap = bitmap_vzalloc(zdev->iommu_pages, GFP_KERNEL);
600 if (!zdev->lazy_bitmap) {
606 rc = zpci_register_ioat(zdev, 0, zdev->start_dma, zdev->end_dma,
607 (u64) zdev->dma_table);
613 vfree(zdev->iommu_bitmap);
614 zdev->iommu_bitmap = NULL;
615 vfree(zdev->lazy_bitmap);
616 zdev->lazy_bitmap = NULL;
618 dma_free_cpu_table(zdev->dma_table);
619 zdev->dma_table = NULL;
624 void zpci_dma_exit_device(struct zpci_dev *zdev)
627 * At this point, if the device is part of an IOMMU domain, this would
628 * be a strong hint towards a bug in the IOMMU API (common) code and/or
629 * simultaneous access via IOMMU and DMA API. So let's issue a warning.
631 WARN_ON(zdev->s390_domain);
633 if (zpci_unregister_ioat(zdev, 0))
636 dma_cleanup_tables(zdev->dma_table);
637 zdev->dma_table = NULL;
638 vfree(zdev->iommu_bitmap);
639 zdev->iommu_bitmap = NULL;
640 vfree(zdev->lazy_bitmap);
641 zdev->lazy_bitmap = NULL;
646 static int __init dma_alloc_cpu_table_caches(void)
648 dma_region_table_cache = kmem_cache_create("PCI_DMA_region_tables",
649 ZPCI_TABLE_SIZE, ZPCI_TABLE_ALIGN,
651 if (!dma_region_table_cache)
654 dma_page_table_cache = kmem_cache_create("PCI_DMA_page_tables",
655 ZPCI_PT_SIZE, ZPCI_PT_ALIGN,
657 if (!dma_page_table_cache) {
658 kmem_cache_destroy(dma_region_table_cache);
664 int __init zpci_dma_init(void)
666 return dma_alloc_cpu_table_caches();
669 void zpci_dma_exit(void)
671 kmem_cache_destroy(dma_page_table_cache);
672 kmem_cache_destroy(dma_region_table_cache);
675 const struct dma_map_ops s390_pci_dma_ops = {
676 .alloc = s390_dma_alloc,
677 .free = s390_dma_free,
678 .map_sg = s390_dma_map_sg,
679 .unmap_sg = s390_dma_unmap_sg,
680 .map_page = s390_dma_map_pages,
681 .unmap_page = s390_dma_unmap_pages,
682 .mmap = dma_common_mmap,
683 .get_sgtable = dma_common_get_sgtable,
684 /* dma_supported is unconditionally true without a callback */
686 EXPORT_SYMBOL_GPL(s390_pci_dma_ops);
688 static int __init s390_iommu_setup(char *str)
690 if (!strcmp(str, "strict"))
691 s390_iommu_strict = 1;
695 __setup("s390_iommu=", s390_iommu_setup);