2 * Copyright IBM Corp. 2012
5 * Jan Glauber <jang@linux.vnet.ibm.com>
8 #include <linux/kernel.h>
9 #include <linux/slab.h>
10 #include <linux/export.h>
11 #include <linux/iommu-helper.h>
12 #include <linux/dma-mapping.h>
13 #include <linux/vmalloc.h>
14 #include <linux/pci.h>
15 #include <asm/pci_dma.h>
17 static struct kmem_cache *dma_region_table_cache;
18 static struct kmem_cache *dma_page_table_cache;
19 static int s390_iommu_strict;
21 static int zpci_refresh_global(struct zpci_dev *zdev)
23 return zpci_refresh_trans((u64) zdev->fh << 32, zdev->start_dma,
24 zdev->iommu_pages * PAGE_SIZE);
27 unsigned long *dma_alloc_cpu_table(void)
29 unsigned long *table, *entry;
31 table = kmem_cache_alloc(dma_region_table_cache, GFP_ATOMIC);
35 for (entry = table; entry < table + ZPCI_TABLE_ENTRIES; entry++)
36 *entry = ZPCI_TABLE_INVALID | ZPCI_TABLE_PROTECTED;
40 static void dma_free_cpu_table(void *table)
42 kmem_cache_free(dma_region_table_cache, table);
45 static unsigned long *dma_alloc_page_table(void)
47 unsigned long *table, *entry;
49 table = kmem_cache_alloc(dma_page_table_cache, GFP_ATOMIC);
53 for (entry = table; entry < table + ZPCI_PT_ENTRIES; entry++)
54 *entry = ZPCI_PTE_INVALID | ZPCI_TABLE_PROTECTED;
58 static void dma_free_page_table(void *table)
60 kmem_cache_free(dma_page_table_cache, table);
63 static unsigned long *dma_get_seg_table_origin(unsigned long *entry)
67 if (reg_entry_isvalid(*entry))
68 sto = get_rt_sto(*entry);
70 sto = dma_alloc_cpu_table();
74 set_rt_sto(entry, sto);
75 validate_rt_entry(entry);
76 entry_clr_protected(entry);
81 static unsigned long *dma_get_page_table_origin(unsigned long *entry)
85 if (reg_entry_isvalid(*entry))
86 pto = get_st_pto(*entry);
88 pto = dma_alloc_page_table();
91 set_st_pto(entry, pto);
92 validate_st_entry(entry);
93 entry_clr_protected(entry);
98 static unsigned long *dma_walk_cpu_trans(unsigned long *rto, dma_addr_t dma_addr)
100 unsigned long *sto, *pto;
101 unsigned int rtx, sx, px;
103 rtx = calc_rtx(dma_addr);
104 sto = dma_get_seg_table_origin(&rto[rtx]);
108 sx = calc_sx(dma_addr);
109 pto = dma_get_page_table_origin(&sto[sx]);
113 px = calc_px(dma_addr);
117 void dma_update_cpu_trans(unsigned long *dma_table, void *page_addr,
118 dma_addr_t dma_addr, int flags)
120 unsigned long *entry;
122 entry = dma_walk_cpu_trans(dma_table, dma_addr);
128 if (flags & ZPCI_PTE_INVALID) {
129 invalidate_pt_entry(entry);
132 set_pt_pfaa(entry, page_addr);
133 validate_pt_entry(entry);
136 if (flags & ZPCI_TABLE_PROTECTED)
137 entry_set_protected(entry);
139 entry_clr_protected(entry);
142 static int dma_update_trans(struct zpci_dev *zdev, unsigned long pa,
143 dma_addr_t dma_addr, size_t size, int flags)
145 unsigned int nr_pages = PAGE_ALIGN(size) >> PAGE_SHIFT;
146 u8 *page_addr = (u8 *) (pa & PAGE_MASK);
147 dma_addr_t start_dma_addr = dma_addr;
148 unsigned long irq_flags;
154 spin_lock_irqsave(&zdev->dma_table_lock, irq_flags);
155 if (!zdev->dma_table)
158 for (i = 0; i < nr_pages; i++) {
159 dma_update_cpu_trans(zdev->dma_table, page_addr, dma_addr,
161 page_addr += PAGE_SIZE;
162 dma_addr += PAGE_SIZE;
166 * With zdev->tlb_refresh == 0, rpcit is not required to establish new
167 * translations when previously invalid translation-table entries are
168 * validated. With lazy unmap, it also is skipped for previously valid
169 * entries, but a global rpcit is then required before any address can
170 * be re-used, i.e. after each iommu bitmap wrap-around.
172 if (!zdev->tlb_refresh &&
173 (!s390_iommu_strict ||
174 ((flags & ZPCI_PTE_VALID_MASK) == ZPCI_PTE_VALID)))
177 rc = zpci_refresh_trans((u64) zdev->fh << 32, start_dma_addr,
178 nr_pages * PAGE_SIZE);
181 spin_unlock_irqrestore(&zdev->dma_table_lock, irq_flags);
185 void dma_free_seg_table(unsigned long entry)
187 unsigned long *sto = get_rt_sto(entry);
190 for (sx = 0; sx < ZPCI_TABLE_ENTRIES; sx++)
191 if (reg_entry_isvalid(sto[sx]))
192 dma_free_page_table(get_st_pto(sto[sx]));
194 dma_free_cpu_table(sto);
197 void dma_cleanup_tables(unsigned long *table)
204 for (rtx = 0; rtx < ZPCI_TABLE_ENTRIES; rtx++)
205 if (reg_entry_isvalid(table[rtx]))
206 dma_free_seg_table(table[rtx]);
208 dma_free_cpu_table(table);
211 static unsigned long __dma_alloc_iommu(struct zpci_dev *zdev,
212 unsigned long start, int size)
214 unsigned long boundary_size;
216 boundary_size = ALIGN(dma_get_seg_boundary(&zdev->pdev->dev) + 1,
217 PAGE_SIZE) >> PAGE_SHIFT;
218 return iommu_area_alloc(zdev->iommu_bitmap, zdev->iommu_pages,
219 start, size, 0, boundary_size, 0);
222 static unsigned long dma_alloc_iommu(struct zpci_dev *zdev, int size)
224 unsigned long offset, flags;
227 spin_lock_irqsave(&zdev->iommu_bitmap_lock, flags);
228 offset = __dma_alloc_iommu(zdev, zdev->next_bit, size);
231 offset = __dma_alloc_iommu(zdev, 0, size);
236 zdev->next_bit = offset + size;
237 if (!zdev->tlb_refresh && !s390_iommu_strict && wrap)
238 /* global flush after wrap-around with lazy unmap */
239 zpci_refresh_global(zdev);
241 spin_unlock_irqrestore(&zdev->iommu_bitmap_lock, flags);
245 static void dma_free_iommu(struct zpci_dev *zdev, unsigned long offset, int size)
249 spin_lock_irqsave(&zdev->iommu_bitmap_lock, flags);
250 if (!zdev->iommu_bitmap)
252 bitmap_clear(zdev->iommu_bitmap, offset, size);
254 * Lazy flush for unmap: need to move next_bit to avoid address re-use
257 if (!s390_iommu_strict && offset >= zdev->next_bit)
258 zdev->next_bit = offset + size;
260 spin_unlock_irqrestore(&zdev->iommu_bitmap_lock, flags);
263 static dma_addr_t s390_dma_map_pages(struct device *dev, struct page *page,
264 unsigned long offset, size_t size,
265 enum dma_data_direction direction,
266 struct dma_attrs *attrs)
268 struct zpci_dev *zdev = to_zpci(to_pci_dev(dev));
269 unsigned long nr_pages, iommu_page_index;
270 unsigned long pa = page_to_phys(page) + offset;
271 int flags = ZPCI_PTE_VALID;
274 /* This rounds up number of pages based on size and offset */
275 nr_pages = iommu_num_pages(pa, size, PAGE_SIZE);
276 iommu_page_index = dma_alloc_iommu(zdev, nr_pages);
277 if (iommu_page_index == -1)
280 /* Use rounded up size */
281 size = nr_pages * PAGE_SIZE;
283 dma_addr = zdev->start_dma + iommu_page_index * PAGE_SIZE;
284 if (dma_addr + size > zdev->end_dma)
287 if (direction == DMA_NONE || direction == DMA_TO_DEVICE)
288 flags |= ZPCI_TABLE_PROTECTED;
290 if (!dma_update_trans(zdev, pa, dma_addr, size, flags)) {
291 atomic64_add(nr_pages, &zdev->mapped_pages);
292 return dma_addr + (offset & ~PAGE_MASK);
296 dma_free_iommu(zdev, iommu_page_index, nr_pages);
298 zpci_err("map error:\n");
299 zpci_err_hex(&pa, sizeof(pa));
300 return DMA_ERROR_CODE;
303 static void s390_dma_unmap_pages(struct device *dev, dma_addr_t dma_addr,
304 size_t size, enum dma_data_direction direction,
305 struct dma_attrs *attrs)
307 struct zpci_dev *zdev = to_zpci(to_pci_dev(dev));
308 unsigned long iommu_page_index;
311 npages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
312 dma_addr = dma_addr & PAGE_MASK;
313 if (dma_update_trans(zdev, 0, dma_addr, npages * PAGE_SIZE,
314 ZPCI_TABLE_PROTECTED | ZPCI_PTE_INVALID)) {
315 zpci_err("unmap error:\n");
316 zpci_err_hex(&dma_addr, sizeof(dma_addr));
319 atomic64_add(npages, &zdev->unmapped_pages);
320 iommu_page_index = (dma_addr - zdev->start_dma) >> PAGE_SHIFT;
321 dma_free_iommu(zdev, iommu_page_index, npages);
324 static void *s390_dma_alloc(struct device *dev, size_t size,
325 dma_addr_t *dma_handle, gfp_t flag,
326 struct dma_attrs *attrs)
328 struct zpci_dev *zdev = to_zpci(to_pci_dev(dev));
333 size = PAGE_ALIGN(size);
334 page = alloc_pages(flag, get_order(size));
338 pa = page_to_phys(page);
339 memset((void *) pa, 0, size);
341 map = s390_dma_map_pages(dev, page, pa % PAGE_SIZE,
342 size, DMA_BIDIRECTIONAL, NULL);
343 if (dma_mapping_error(dev, map)) {
344 free_pages(pa, get_order(size));
348 atomic64_add(size / PAGE_SIZE, &zdev->allocated_pages);
354 static void s390_dma_free(struct device *dev, size_t size,
355 void *pa, dma_addr_t dma_handle,
356 struct dma_attrs *attrs)
358 struct zpci_dev *zdev = to_zpci(to_pci_dev(dev));
360 size = PAGE_ALIGN(size);
361 atomic64_sub(size / PAGE_SIZE, &zdev->allocated_pages);
362 s390_dma_unmap_pages(dev, dma_handle, size, DMA_BIDIRECTIONAL, NULL);
363 free_pages((unsigned long) pa, get_order(size));
366 static int s390_dma_map_sg(struct device *dev, struct scatterlist *sg,
367 int nr_elements, enum dma_data_direction dir,
368 struct dma_attrs *attrs)
370 int mapped_elements = 0;
371 struct scatterlist *s;
374 for_each_sg(sg, s, nr_elements, i) {
375 struct page *page = sg_page(s);
376 s->dma_address = s390_dma_map_pages(dev, page, s->offset,
377 s->length, dir, NULL);
378 if (!dma_mapping_error(dev, s->dma_address)) {
379 s->dma_length = s->length;
385 return mapped_elements;
388 for_each_sg(sg, s, mapped_elements, i) {
390 s390_dma_unmap_pages(dev, s->dma_address, s->dma_length,
399 static void s390_dma_unmap_sg(struct device *dev, struct scatterlist *sg,
400 int nr_elements, enum dma_data_direction dir,
401 struct dma_attrs *attrs)
403 struct scatterlist *s;
406 for_each_sg(sg, s, nr_elements, i) {
407 s390_dma_unmap_pages(dev, s->dma_address, s->dma_length, dir, NULL);
413 int zpci_dma_init_device(struct zpci_dev *zdev)
418 * At this point, if the device is part of an IOMMU domain, this would
419 * be a strong hint towards a bug in the IOMMU API (common) code and/or
420 * simultaneous access via IOMMU and DMA API. So let's issue a warning.
422 WARN_ON(zdev->s390_domain);
424 spin_lock_init(&zdev->iommu_bitmap_lock);
425 spin_lock_init(&zdev->dma_table_lock);
427 zdev->dma_table = dma_alloc_cpu_table();
428 if (!zdev->dma_table) {
433 zdev->iommu_size = (unsigned long) high_memory - PAGE_OFFSET;
434 zdev->iommu_pages = zdev->iommu_size >> PAGE_SHIFT;
435 zdev->iommu_bitmap = vzalloc(zdev->iommu_pages / 8);
436 if (!zdev->iommu_bitmap) {
441 rc = zpci_register_ioat(zdev,
443 zdev->start_dma + PAGE_OFFSET,
444 zdev->start_dma + zdev->iommu_size - 1,
445 (u64) zdev->dma_table);
451 dma_free_cpu_table(zdev->dma_table);
456 void zpci_dma_exit_device(struct zpci_dev *zdev)
459 * At this point, if the device is part of an IOMMU domain, this would
460 * be a strong hint towards a bug in the IOMMU API (common) code and/or
461 * simultaneous access via IOMMU and DMA API. So let's issue a warning.
463 WARN_ON(zdev->s390_domain);
465 zpci_unregister_ioat(zdev, 0);
466 dma_cleanup_tables(zdev->dma_table);
467 zdev->dma_table = NULL;
468 vfree(zdev->iommu_bitmap);
469 zdev->iommu_bitmap = NULL;
473 static int __init dma_alloc_cpu_table_caches(void)
475 dma_region_table_cache = kmem_cache_create("PCI_DMA_region_tables",
476 ZPCI_TABLE_SIZE, ZPCI_TABLE_ALIGN,
478 if (!dma_region_table_cache)
481 dma_page_table_cache = kmem_cache_create("PCI_DMA_page_tables",
482 ZPCI_PT_SIZE, ZPCI_PT_ALIGN,
484 if (!dma_page_table_cache) {
485 kmem_cache_destroy(dma_region_table_cache);
491 int __init zpci_dma_init(void)
493 return dma_alloc_cpu_table_caches();
496 void zpci_dma_exit(void)
498 kmem_cache_destroy(dma_page_table_cache);
499 kmem_cache_destroy(dma_region_table_cache);
502 #define PREALLOC_DMA_DEBUG_ENTRIES (1 << 16)
504 static int __init dma_debug_do_init(void)
506 dma_debug_init(PREALLOC_DMA_DEBUG_ENTRIES);
509 fs_initcall(dma_debug_do_init);
511 struct dma_map_ops s390_dma_ops = {
512 .alloc = s390_dma_alloc,
513 .free = s390_dma_free,
514 .map_sg = s390_dma_map_sg,
515 .unmap_sg = s390_dma_unmap_sg,
516 .map_page = s390_dma_map_pages,
517 .unmap_page = s390_dma_unmap_pages,
518 /* if we support direct DMA this must be conditional */
520 /* dma_supported is unconditionally true without a callback */
522 EXPORT_SYMBOL_GPL(s390_dma_ops);
524 static int __init s390_iommu_setup(char *str)
526 if (!strncmp(str, "strict", 6))
527 s390_iommu_strict = 1;
531 __setup("s390_iommu=", s390_iommu_setup);