blob: a547419907c3a46be4e7a64c36f3f1a776abcfcf [file] [log] [blame]
Jan Glauber828b35f2012-11-29 14:33:30 +01001/*
2 * Copyright IBM Corp. 2012
3 *
4 * Author(s):
5 * Jan Glauber <jang@linux.vnet.ibm.com>
6 */
7
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/pci.h>
14#include <asm/pci_dma.h>
15
Jan Glauber828b35f2012-11-29 14:33:30 +010016static struct kmem_cache *dma_region_table_cache;
17static struct kmem_cache *dma_page_table_cache;
18
19static unsigned long *dma_alloc_cpu_table(void)
20{
21 unsigned long *table, *entry;
22
23 table = kmem_cache_alloc(dma_region_table_cache, GFP_ATOMIC);
24 if (!table)
25 return NULL;
26
27 for (entry = table; entry < table + ZPCI_TABLE_ENTRIES; entry++)
28 *entry = ZPCI_TABLE_INVALID | ZPCI_TABLE_PROTECTED;
29 return table;
30}
31
32static void dma_free_cpu_table(void *table)
33{
34 kmem_cache_free(dma_region_table_cache, table);
35}
36
37static unsigned long *dma_alloc_page_table(void)
38{
39 unsigned long *table, *entry;
40
41 table = kmem_cache_alloc(dma_page_table_cache, GFP_ATOMIC);
42 if (!table)
43 return NULL;
44
45 for (entry = table; entry < table + ZPCI_PT_ENTRIES; entry++)
46 *entry = ZPCI_PTE_INVALID | ZPCI_TABLE_PROTECTED;
47 return table;
48}
49
50static void dma_free_page_table(void *table)
51{
52 kmem_cache_free(dma_page_table_cache, table);
53}
54
55static unsigned long *dma_get_seg_table_origin(unsigned long *entry)
56{
57 unsigned long *sto;
58
59 if (reg_entry_isvalid(*entry))
60 sto = get_rt_sto(*entry);
61 else {
62 sto = dma_alloc_cpu_table();
63 if (!sto)
64 return NULL;
65
66 set_rt_sto(entry, sto);
67 validate_rt_entry(entry);
68 entry_clr_protected(entry);
69 }
70 return sto;
71}
72
73static unsigned long *dma_get_page_table_origin(unsigned long *entry)
74{
75 unsigned long *pto;
76
77 if (reg_entry_isvalid(*entry))
78 pto = get_st_pto(*entry);
79 else {
80 pto = dma_alloc_page_table();
81 if (!pto)
82 return NULL;
83 set_st_pto(entry, pto);
84 validate_st_entry(entry);
85 entry_clr_protected(entry);
86 }
87 return pto;
88}
89
90static unsigned long *dma_walk_cpu_trans(unsigned long *rto, dma_addr_t dma_addr)
91{
92 unsigned long *sto, *pto;
93 unsigned int rtx, sx, px;
94
95 rtx = calc_rtx(dma_addr);
96 sto = dma_get_seg_table_origin(&rto[rtx]);
97 if (!sto)
98 return NULL;
99
100 sx = calc_sx(dma_addr);
101 pto = dma_get_page_table_origin(&sto[sx]);
102 if (!pto)
103 return NULL;
104
105 px = calc_px(dma_addr);
106 return &pto[px];
107}
108
109static void dma_update_cpu_trans(struct zpci_dev *zdev, void *page_addr,
110 dma_addr_t dma_addr, int flags)
111{
112 unsigned long *entry;
113
114 entry = dma_walk_cpu_trans(zdev->dma_table, dma_addr);
115 if (!entry) {
116 WARN_ON_ONCE(1);
117 return;
118 }
119
120 if (flags & ZPCI_PTE_INVALID) {
121 invalidate_pt_entry(entry);
122 return;
123 } else {
124 set_pt_pfaa(entry, page_addr);
125 validate_pt_entry(entry);
126 }
127
128 if (flags & ZPCI_TABLE_PROTECTED)
129 entry_set_protected(entry);
130 else
131 entry_clr_protected(entry);
132}
133
134static int dma_update_trans(struct zpci_dev *zdev, unsigned long pa,
135 dma_addr_t dma_addr, size_t size, int flags)
136{
137 unsigned int nr_pages = PAGE_ALIGN(size) >> PAGE_SHIFT;
138 u8 *page_addr = (u8 *) (pa & PAGE_MASK);
139 dma_addr_t start_dma_addr = dma_addr;
140 unsigned long irq_flags;
141 int i, rc = 0;
142
143 if (!nr_pages)
144 return -EINVAL;
145
146 spin_lock_irqsave(&zdev->dma_table_lock, irq_flags);
147 if (!zdev->dma_table) {
148 dev_err(&zdev->pdev->dev, "Missing DMA table\n");
149 goto no_refresh;
150 }
151
152 for (i = 0; i < nr_pages; i++) {
153 dma_update_cpu_trans(zdev, page_addr, dma_addr, flags);
154 page_addr += PAGE_SIZE;
155 dma_addr += PAGE_SIZE;
156 }
157
158 /*
159 * rpcit is not required to establish new translations when previously
160 * invalid translation-table entries are validated, however it is
161 * required when altering previously valid entries.
162 */
163 if (!zdev->tlb_refresh &&
164 ((flags & ZPCI_PTE_VALID_MASK) == ZPCI_PTE_VALID))
165 /*
166 * TODO: also need to check that the old entry is indeed INVALID
167 * and not only for one page but for the whole range...
168 * -> now we WARN_ON in that case but with lazy unmap that
169 * needs to be redone!
170 */
171 goto no_refresh;
172 rc = rpcit_instr((u64) zdev->fh << 32, start_dma_addr,
173 nr_pages * PAGE_SIZE);
174
175no_refresh:
176 spin_unlock_irqrestore(&zdev->dma_table_lock, irq_flags);
177 return rc;
178}
179
180static void dma_free_seg_table(unsigned long entry)
181{
182 unsigned long *sto = get_rt_sto(entry);
183 int sx;
184
185 for (sx = 0; sx < ZPCI_TABLE_ENTRIES; sx++)
186 if (reg_entry_isvalid(sto[sx]))
187 dma_free_page_table(get_st_pto(sto[sx]));
188
189 dma_free_cpu_table(sto);
190}
191
192static void dma_cleanup_tables(struct zpci_dev *zdev)
193{
Wei Yongjunbafff172012-12-03 16:15:00 +0100194 unsigned long *table;
Jan Glauber828b35f2012-11-29 14:33:30 +0100195 int rtx;
196
197 if (!zdev || !zdev->dma_table)
198 return;
199
Wei Yongjunbafff172012-12-03 16:15:00 +0100200 table = zdev->dma_table;
Jan Glauber828b35f2012-11-29 14:33:30 +0100201 for (rtx = 0; rtx < ZPCI_TABLE_ENTRIES; rtx++)
202 if (reg_entry_isvalid(table[rtx]))
203 dma_free_seg_table(table[rtx]);
204
205 dma_free_cpu_table(table);
206 zdev->dma_table = NULL;
207}
208
209static unsigned long __dma_alloc_iommu(struct zpci_dev *zdev, unsigned long start,
210 int size)
211{
212 unsigned long boundary_size = 0x1000000;
213
214 return iommu_area_alloc(zdev->iommu_bitmap, zdev->iommu_pages,
215 start, size, 0, boundary_size, 0);
216}
217
218static unsigned long dma_alloc_iommu(struct zpci_dev *zdev, int size)
219{
220 unsigned long offset, flags;
221
222 spin_lock_irqsave(&zdev->iommu_bitmap_lock, flags);
223 offset = __dma_alloc_iommu(zdev, zdev->next_bit, size);
224 if (offset == -1)
225 offset = __dma_alloc_iommu(zdev, 0, size);
226
227 if (offset != -1) {
228 zdev->next_bit = offset + size;
229 if (zdev->next_bit >= zdev->iommu_pages)
230 zdev->next_bit = 0;
231 }
232 spin_unlock_irqrestore(&zdev->iommu_bitmap_lock, flags);
233 return offset;
234}
235
236static void dma_free_iommu(struct zpci_dev *zdev, unsigned long offset, int size)
237{
238 unsigned long flags;
239
240 spin_lock_irqsave(&zdev->iommu_bitmap_lock, flags);
241 if (!zdev->iommu_bitmap)
242 goto out;
243 bitmap_clear(zdev->iommu_bitmap, offset, size);
244 if (offset >= zdev->next_bit)
245 zdev->next_bit = offset + size;
246out:
247 spin_unlock_irqrestore(&zdev->iommu_bitmap_lock, flags);
248}
249
250int dma_set_mask(struct device *dev, u64 mask)
251{
252 if (!dev->dma_mask || !dma_supported(dev, mask))
253 return -EIO;
254
255 *dev->dma_mask = mask;
256 return 0;
257}
258EXPORT_SYMBOL_GPL(dma_set_mask);
259
260static dma_addr_t s390_dma_map_pages(struct device *dev, struct page *page,
261 unsigned long offset, size_t size,
262 enum dma_data_direction direction,
263 struct dma_attrs *attrs)
264{
265 struct zpci_dev *zdev = get_zdev(container_of(dev, struct pci_dev, dev));
266 unsigned long nr_pages, iommu_page_index;
267 unsigned long pa = page_to_phys(page) + offset;
268 int flags = ZPCI_PTE_VALID;
269 dma_addr_t dma_addr;
270
271 WARN_ON_ONCE(offset > PAGE_SIZE);
272
273 /* This rounds up number of pages based on size and offset */
274 nr_pages = iommu_num_pages(pa, size, PAGE_SIZE);
275 iommu_page_index = dma_alloc_iommu(zdev, nr_pages);
276 if (iommu_page_index == -1)
277 goto out_err;
278
279 /* Use rounded up size */
280 size = nr_pages * PAGE_SIZE;
281
282 dma_addr = zdev->start_dma + iommu_page_index * PAGE_SIZE;
283 if (dma_addr + size > zdev->end_dma) {
284 dev_err(dev, "(dma_addr: 0x%16.16LX + size: 0x%16.16lx) > end_dma: 0x%16.16Lx\n",
285 dma_addr, size, zdev->end_dma);
286 goto out_free;
287 }
288
289 if (direction == DMA_NONE || direction == DMA_TO_DEVICE)
290 flags |= ZPCI_TABLE_PROTECTED;
291
Jan Glauberd0b08852012-12-11 14:53:35 +0100292 if (!dma_update_trans(zdev, pa, dma_addr, size, flags)) {
293 atomic64_add(nr_pages, (atomic64_t *) &zdev->fmb->mapped_pages);
Jan Glauber828b35f2012-11-29 14:33:30 +0100294 return dma_addr + offset;
Jan Glauberd0b08852012-12-11 14:53:35 +0100295 }
Jan Glauber828b35f2012-11-29 14:33:30 +0100296
297out_free:
298 dma_free_iommu(zdev, iommu_page_index, nr_pages);
299out_err:
300 dev_err(dev, "Failed to map addr: %lx\n", pa);
301 return DMA_ERROR_CODE;
302}
303
304static void s390_dma_unmap_pages(struct device *dev, dma_addr_t dma_addr,
305 size_t size, enum dma_data_direction direction,
306 struct dma_attrs *attrs)
307{
308 struct zpci_dev *zdev = get_zdev(container_of(dev, struct pci_dev, dev));
309 unsigned long iommu_page_index;
310 int npages;
311
312 npages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
313 dma_addr = dma_addr & PAGE_MASK;
314 if (dma_update_trans(zdev, 0, dma_addr, npages * PAGE_SIZE,
315 ZPCI_TABLE_PROTECTED | ZPCI_PTE_INVALID))
316 dev_err(dev, "Failed to unmap addr: %Lx\n", dma_addr);
317
Jan Glauberd0b08852012-12-11 14:53:35 +0100318 atomic64_add(npages, (atomic64_t *) &zdev->fmb->unmapped_pages);
Jan Glauber828b35f2012-11-29 14:33:30 +0100319 iommu_page_index = (dma_addr - zdev->start_dma) >> PAGE_SHIFT;
320 dma_free_iommu(zdev, iommu_page_index, npages);
321}
322
323static void *s390_dma_alloc(struct device *dev, size_t size,
324 dma_addr_t *dma_handle, gfp_t flag,
325 struct dma_attrs *attrs)
326{
Jan Glauberd0b08852012-12-11 14:53:35 +0100327 struct zpci_dev *zdev = get_zdev(container_of(dev, struct pci_dev, dev));
Jan Glauber828b35f2012-11-29 14:33:30 +0100328 struct page *page;
329 unsigned long pa;
330 dma_addr_t map;
331
332 size = PAGE_ALIGN(size);
333 page = alloc_pages(flag, get_order(size));
334 if (!page)
335 return NULL;
Jan Glauberd0b08852012-12-11 14:53:35 +0100336
337 atomic64_add(size / PAGE_SIZE, (atomic64_t *) &zdev->fmb->allocated_pages);
Jan Glauber828b35f2012-11-29 14:33:30 +0100338 pa = page_to_phys(page);
339 memset((void *) pa, 0, size);
340
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));
345 return NULL;
346 }
347
348 if (dma_handle)
349 *dma_handle = map;
350 return (void *) pa;
351}
352
353static void s390_dma_free(struct device *dev, size_t size,
354 void *pa, dma_addr_t dma_handle,
355 struct dma_attrs *attrs)
356{
357 s390_dma_unmap_pages(dev, dma_handle, PAGE_ALIGN(size),
358 DMA_BIDIRECTIONAL, NULL);
359 free_pages((unsigned long) pa, get_order(size));
360}
361
362static int s390_dma_map_sg(struct device *dev, struct scatterlist *sg,
363 int nr_elements, enum dma_data_direction dir,
364 struct dma_attrs *attrs)
365{
366 int mapped_elements = 0;
367 struct scatterlist *s;
368 int i;
369
370 for_each_sg(sg, s, nr_elements, i) {
371 struct page *page = sg_page(s);
372 s->dma_address = s390_dma_map_pages(dev, page, s->offset,
373 s->length, dir, NULL);
374 if (!dma_mapping_error(dev, s->dma_address)) {
375 s->dma_length = s->length;
376 mapped_elements++;
377 } else
378 goto unmap;
379 }
380out:
381 return mapped_elements;
382
383unmap:
384 for_each_sg(sg, s, mapped_elements, i) {
385 if (s->dma_address)
386 s390_dma_unmap_pages(dev, s->dma_address, s->dma_length,
387 dir, NULL);
388 s->dma_address = 0;
389 s->dma_length = 0;
390 }
391 mapped_elements = 0;
392 goto out;
393}
394
395static void s390_dma_unmap_sg(struct device *dev, struct scatterlist *sg,
396 int nr_elements, enum dma_data_direction dir,
397 struct dma_attrs *attrs)
398{
399 struct scatterlist *s;
400 int i;
401
402 for_each_sg(sg, s, nr_elements, i) {
403 s390_dma_unmap_pages(dev, s->dma_address, s->dma_length, dir, NULL);
404 s->dma_address = 0;
405 s->dma_length = 0;
406 }
407}
408
409int zpci_dma_init_device(struct zpci_dev *zdev)
410{
411 unsigned int bitmap_order;
412 int rc;
413
414 spin_lock_init(&zdev->iommu_bitmap_lock);
415 spin_lock_init(&zdev->dma_table_lock);
416
417 zdev->dma_table = dma_alloc_cpu_table();
418 if (!zdev->dma_table) {
419 rc = -ENOMEM;
420 goto out_clean;
421 }
422
423 zdev->iommu_size = (unsigned long) high_memory - PAGE_OFFSET;
424 zdev->iommu_pages = zdev->iommu_size >> PAGE_SHIFT;
425 bitmap_order = get_order(zdev->iommu_pages / 8);
426 pr_info("iommu_size: 0x%lx iommu_pages: 0x%lx bitmap_order: %i\n",
427 zdev->iommu_size, zdev->iommu_pages, bitmap_order);
428
429 zdev->iommu_bitmap = (void *) __get_free_pages(GFP_KERNEL | __GFP_ZERO,
430 bitmap_order);
431 if (!zdev->iommu_bitmap) {
432 rc = -ENOMEM;
433 goto out_reg;
434 }
435
436 rc = zpci_register_ioat(zdev,
437 0,
438 zdev->start_dma + PAGE_OFFSET,
439 zdev->start_dma + zdev->iommu_size - 1,
440 (u64) zdev->dma_table);
441 if (rc)
442 goto out_reg;
443 return 0;
444
445out_reg:
446 dma_free_cpu_table(zdev->dma_table);
447out_clean:
448 return rc;
449}
450
451void zpci_dma_exit_device(struct zpci_dev *zdev)
452{
453 zpci_unregister_ioat(zdev, 0);
454 dma_cleanup_tables(zdev);
455 free_pages((unsigned long) zdev->iommu_bitmap,
456 get_order(zdev->iommu_pages / 8));
457 zdev->iommu_bitmap = NULL;
458 zdev->next_bit = 0;
459}
460
461static int __init dma_alloc_cpu_table_caches(void)
462{
463 dma_region_table_cache = kmem_cache_create("PCI_DMA_region_tables",
464 ZPCI_TABLE_SIZE, ZPCI_TABLE_ALIGN,
465 0, NULL);
466 if (!dma_region_table_cache)
467 return -ENOMEM;
468
469 dma_page_table_cache = kmem_cache_create("PCI_DMA_page_tables",
470 ZPCI_PT_SIZE, ZPCI_PT_ALIGN,
471 0, NULL);
472 if (!dma_page_table_cache) {
473 kmem_cache_destroy(dma_region_table_cache);
474 return -ENOMEM;
475 }
476 return 0;
477}
478
479int __init zpci_dma_init(void)
480{
481 return dma_alloc_cpu_table_caches();
482}
483
484void zpci_dma_exit(void)
485{
486 kmem_cache_destroy(dma_page_table_cache);
487 kmem_cache_destroy(dma_region_table_cache);
488}
489
490#define PREALLOC_DMA_DEBUG_ENTRIES (1 << 16)
491
492static int __init dma_debug_do_init(void)
493{
494 dma_debug_init(PREALLOC_DMA_DEBUG_ENTRIES);
495 return 0;
496}
497fs_initcall(dma_debug_do_init);
498
499struct dma_map_ops s390_dma_ops = {
500 .alloc = s390_dma_alloc,
501 .free = s390_dma_free,
502 .map_sg = s390_dma_map_sg,
503 .unmap_sg = s390_dma_unmap_sg,
504 .map_page = s390_dma_map_pages,
505 .unmap_page = s390_dma_unmap_pages,
506 /* if we support direct DMA this must be conditional */
507 .is_phys = 0,
508 /* dma_supported is unconditionally true without a callback */
509};
510EXPORT_SYMBOL_GPL(s390_dma_ops);