xref: /openbmc/linux/arch/s390/pci/pci_dma.c (revision 79f08d9e)
1 /*
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/vmalloc.h>
14 #include <linux/pci.h>
15 #include <asm/pci_dma.h>
16 
17 static struct kmem_cache *dma_region_table_cache;
18 static struct kmem_cache *dma_page_table_cache;
19 
20 static unsigned long *dma_alloc_cpu_table(void)
21 {
22 	unsigned long *table, *entry;
23 
24 	table = kmem_cache_alloc(dma_region_table_cache, GFP_ATOMIC);
25 	if (!table)
26 		return NULL;
27 
28 	for (entry = table; entry < table + ZPCI_TABLE_ENTRIES; entry++)
29 		*entry = ZPCI_TABLE_INVALID | ZPCI_TABLE_PROTECTED;
30 	return table;
31 }
32 
33 static void dma_free_cpu_table(void *table)
34 {
35 	kmem_cache_free(dma_region_table_cache, table);
36 }
37 
38 static unsigned long *dma_alloc_page_table(void)
39 {
40 	unsigned long *table, *entry;
41 
42 	table = kmem_cache_alloc(dma_page_table_cache, GFP_ATOMIC);
43 	if (!table)
44 		return NULL;
45 
46 	for (entry = table; entry < table + ZPCI_PT_ENTRIES; entry++)
47 		*entry = ZPCI_PTE_INVALID | ZPCI_TABLE_PROTECTED;
48 	return table;
49 }
50 
51 static void dma_free_page_table(void *table)
52 {
53 	kmem_cache_free(dma_page_table_cache, table);
54 }
55 
56 static unsigned long *dma_get_seg_table_origin(unsigned long *entry)
57 {
58 	unsigned long *sto;
59 
60 	if (reg_entry_isvalid(*entry))
61 		sto = get_rt_sto(*entry);
62 	else {
63 		sto = dma_alloc_cpu_table();
64 		if (!sto)
65 			return NULL;
66 
67 		set_rt_sto(entry, sto);
68 		validate_rt_entry(entry);
69 		entry_clr_protected(entry);
70 	}
71 	return sto;
72 }
73 
74 static unsigned long *dma_get_page_table_origin(unsigned long *entry)
75 {
76 	unsigned long *pto;
77 
78 	if (reg_entry_isvalid(*entry))
79 		pto = get_st_pto(*entry);
80 	else {
81 		pto = dma_alloc_page_table();
82 		if (!pto)
83 			return NULL;
84 		set_st_pto(entry, pto);
85 		validate_st_entry(entry);
86 		entry_clr_protected(entry);
87 	}
88 	return pto;
89 }
90 
91 static unsigned long *dma_walk_cpu_trans(unsigned long *rto, dma_addr_t dma_addr)
92 {
93 	unsigned long *sto, *pto;
94 	unsigned int rtx, sx, px;
95 
96 	rtx = calc_rtx(dma_addr);
97 	sto = dma_get_seg_table_origin(&rto[rtx]);
98 	if (!sto)
99 		return NULL;
100 
101 	sx = calc_sx(dma_addr);
102 	pto = dma_get_page_table_origin(&sto[sx]);
103 	if (!pto)
104 		return NULL;
105 
106 	px = calc_px(dma_addr);
107 	return &pto[px];
108 }
109 
110 static void dma_update_cpu_trans(struct zpci_dev *zdev, void *page_addr,
111 				 dma_addr_t dma_addr, int flags)
112 {
113 	unsigned long *entry;
114 
115 	entry = dma_walk_cpu_trans(zdev->dma_table, dma_addr);
116 	if (!entry) {
117 		WARN_ON_ONCE(1);
118 		return;
119 	}
120 
121 	if (flags & ZPCI_PTE_INVALID) {
122 		invalidate_pt_entry(entry);
123 		return;
124 	} else {
125 		set_pt_pfaa(entry, page_addr);
126 		validate_pt_entry(entry);
127 	}
128 
129 	if (flags & ZPCI_TABLE_PROTECTED)
130 		entry_set_protected(entry);
131 	else
132 		entry_clr_protected(entry);
133 }
134 
135 static int dma_update_trans(struct zpci_dev *zdev, unsigned long pa,
136 			    dma_addr_t dma_addr, size_t size, int flags)
137 {
138 	unsigned int nr_pages = PAGE_ALIGN(size) >> PAGE_SHIFT;
139 	u8 *page_addr = (u8 *) (pa & PAGE_MASK);
140 	dma_addr_t start_dma_addr = dma_addr;
141 	unsigned long irq_flags;
142 	int i, rc = 0;
143 
144 	if (!nr_pages)
145 		return -EINVAL;
146 
147 	spin_lock_irqsave(&zdev->dma_table_lock, irq_flags);
148 	if (!zdev->dma_table)
149 		goto no_refresh;
150 
151 	for (i = 0; i < nr_pages; i++) {
152 		dma_update_cpu_trans(zdev, page_addr, dma_addr, flags);
153 		page_addr += PAGE_SIZE;
154 		dma_addr += PAGE_SIZE;
155 	}
156 
157 	/*
158 	 * rpcit is not required to establish new translations when previously
159 	 * invalid translation-table entries are validated, however it is
160 	 * required when altering previously valid entries.
161 	 */
162 	if (!zdev->tlb_refresh &&
163 	    ((flags & ZPCI_PTE_VALID_MASK) == ZPCI_PTE_VALID))
164 		/*
165 		 * TODO: also need to check that the old entry is indeed INVALID
166 		 * and not only for one page but for the whole range...
167 		 * -> now we WARN_ON in that case but with lazy unmap that
168 		 * needs to be redone!
169 		 */
170 		goto no_refresh;
171 
172 	rc = zpci_refresh_trans((u64) zdev->fh << 32, start_dma_addr,
173 				nr_pages * PAGE_SIZE);
174 
175 no_refresh:
176 	spin_unlock_irqrestore(&zdev->dma_table_lock, irq_flags);
177 	return rc;
178 }
179 
180 static 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 
192 static void dma_cleanup_tables(struct zpci_dev *zdev)
193 {
194 	unsigned long *table;
195 	int rtx;
196 
197 	if (!zdev || !zdev->dma_table)
198 		return;
199 
200 	table = zdev->dma_table;
201 	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 
209 static 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 
218 static 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 
236 static 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;
246 out:
247 	spin_unlock_irqrestore(&zdev->iommu_bitmap_lock, flags);
248 }
249 
250 int 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 }
258 EXPORT_SYMBOL_GPL(dma_set_mask);
259 
260 static 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(to_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 	/* This rounds up number of pages based on size and offset */
272 	nr_pages = iommu_num_pages(pa, size, PAGE_SIZE);
273 	iommu_page_index = dma_alloc_iommu(zdev, nr_pages);
274 	if (iommu_page_index == -1)
275 		goto out_err;
276 
277 	/* Use rounded up size */
278 	size = nr_pages * PAGE_SIZE;
279 
280 	dma_addr = zdev->start_dma + iommu_page_index * PAGE_SIZE;
281 	if (dma_addr + size > zdev->end_dma)
282 		goto out_free;
283 
284 	if (direction == DMA_NONE || direction == DMA_TO_DEVICE)
285 		flags |= ZPCI_TABLE_PROTECTED;
286 
287 	if (!dma_update_trans(zdev, pa, dma_addr, size, flags)) {
288 		atomic64_add(nr_pages, (atomic64_t *) &zdev->fmb->mapped_pages);
289 		return dma_addr + (offset & ~PAGE_MASK);
290 	}
291 
292 out_free:
293 	dma_free_iommu(zdev, iommu_page_index, nr_pages);
294 out_err:
295 	zpci_err("map error:\n");
296 	zpci_err_hex(&pa, sizeof(pa));
297 	return DMA_ERROR_CODE;
298 }
299 
300 static void s390_dma_unmap_pages(struct device *dev, dma_addr_t dma_addr,
301 				 size_t size, enum dma_data_direction direction,
302 				 struct dma_attrs *attrs)
303 {
304 	struct zpci_dev *zdev = get_zdev(to_pci_dev(dev));
305 	unsigned long iommu_page_index;
306 	int npages;
307 
308 	npages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
309 	dma_addr = dma_addr & PAGE_MASK;
310 	if (dma_update_trans(zdev, 0, dma_addr, npages * PAGE_SIZE,
311 			     ZPCI_TABLE_PROTECTED | ZPCI_PTE_INVALID)) {
312 		zpci_err("unmap error:\n");
313 		zpci_err_hex(&dma_addr, sizeof(dma_addr));
314 	}
315 
316 	atomic64_add(npages, (atomic64_t *) &zdev->fmb->unmapped_pages);
317 	iommu_page_index = (dma_addr - zdev->start_dma) >> PAGE_SHIFT;
318 	dma_free_iommu(zdev, iommu_page_index, npages);
319 }
320 
321 static void *s390_dma_alloc(struct device *dev, size_t size,
322 			    dma_addr_t *dma_handle, gfp_t flag,
323 			    struct dma_attrs *attrs)
324 {
325 	struct zpci_dev *zdev = get_zdev(to_pci_dev(dev));
326 	struct page *page;
327 	unsigned long pa;
328 	dma_addr_t map;
329 
330 	size = PAGE_ALIGN(size);
331 	page = alloc_pages(flag, get_order(size));
332 	if (!page)
333 		return NULL;
334 
335 	atomic64_add(size / PAGE_SIZE, (atomic64_t *) &zdev->fmb->allocated_pages);
336 	pa = page_to_phys(page);
337 	memset((void *) pa, 0, size);
338 
339 	map = s390_dma_map_pages(dev, page, pa % PAGE_SIZE,
340 				 size, DMA_BIDIRECTIONAL, NULL);
341 	if (dma_mapping_error(dev, map)) {
342 		free_pages(pa, get_order(size));
343 		return NULL;
344 	}
345 
346 	if (dma_handle)
347 		*dma_handle = map;
348 	return (void *) pa;
349 }
350 
351 static void s390_dma_free(struct device *dev, size_t size,
352 			  void *pa, dma_addr_t dma_handle,
353 			  struct dma_attrs *attrs)
354 {
355 	s390_dma_unmap_pages(dev, dma_handle, PAGE_ALIGN(size),
356 			     DMA_BIDIRECTIONAL, NULL);
357 	free_pages((unsigned long) pa, get_order(size));
358 }
359 
360 static int s390_dma_map_sg(struct device *dev, struct scatterlist *sg,
361 			   int nr_elements, enum dma_data_direction dir,
362 			   struct dma_attrs *attrs)
363 {
364 	int mapped_elements = 0;
365 	struct scatterlist *s;
366 	int i;
367 
368 	for_each_sg(sg, s, nr_elements, i) {
369 		struct page *page = sg_page(s);
370 		s->dma_address = s390_dma_map_pages(dev, page, s->offset,
371 						    s->length, dir, NULL);
372 		if (!dma_mapping_error(dev, s->dma_address)) {
373 			s->dma_length = s->length;
374 			mapped_elements++;
375 		} else
376 			goto unmap;
377 	}
378 out:
379 	return mapped_elements;
380 
381 unmap:
382 	for_each_sg(sg, s, mapped_elements, i) {
383 		if (s->dma_address)
384 			s390_dma_unmap_pages(dev, s->dma_address, s->dma_length,
385 					     dir, NULL);
386 		s->dma_address = 0;
387 		s->dma_length = 0;
388 	}
389 	mapped_elements = 0;
390 	goto out;
391 }
392 
393 static void s390_dma_unmap_sg(struct device *dev, struct scatterlist *sg,
394 			      int nr_elements, enum dma_data_direction dir,
395 			      struct dma_attrs *attrs)
396 {
397 	struct scatterlist *s;
398 	int i;
399 
400 	for_each_sg(sg, s, nr_elements, i) {
401 		s390_dma_unmap_pages(dev, s->dma_address, s->dma_length, dir, NULL);
402 		s->dma_address = 0;
403 		s->dma_length = 0;
404 	}
405 }
406 
407 int zpci_dma_init_device(struct zpci_dev *zdev)
408 {
409 	int rc;
410 
411 	spin_lock_init(&zdev->iommu_bitmap_lock);
412 	spin_lock_init(&zdev->dma_table_lock);
413 
414 	zdev->dma_table = dma_alloc_cpu_table();
415 	if (!zdev->dma_table) {
416 		rc = -ENOMEM;
417 		goto out_clean;
418 	}
419 
420 	zdev->iommu_size = (unsigned long) high_memory - PAGE_OFFSET;
421 	zdev->iommu_pages = zdev->iommu_size >> PAGE_SHIFT;
422 	zdev->iommu_bitmap = vzalloc(zdev->iommu_pages / 8);
423 	if (!zdev->iommu_bitmap) {
424 		rc = -ENOMEM;
425 		goto out_reg;
426 	}
427 
428 	rc = zpci_register_ioat(zdev,
429 				0,
430 				zdev->start_dma + PAGE_OFFSET,
431 				zdev->start_dma + zdev->iommu_size - 1,
432 				(u64) zdev->dma_table);
433 	if (rc)
434 		goto out_reg;
435 	return 0;
436 
437 out_reg:
438 	dma_free_cpu_table(zdev->dma_table);
439 out_clean:
440 	return rc;
441 }
442 
443 void zpci_dma_exit_device(struct zpci_dev *zdev)
444 {
445 	zpci_unregister_ioat(zdev, 0);
446 	dma_cleanup_tables(zdev);
447 	vfree(zdev->iommu_bitmap);
448 	zdev->iommu_bitmap = NULL;
449 	zdev->next_bit = 0;
450 }
451 
452 static int __init dma_alloc_cpu_table_caches(void)
453 {
454 	dma_region_table_cache = kmem_cache_create("PCI_DMA_region_tables",
455 					ZPCI_TABLE_SIZE, ZPCI_TABLE_ALIGN,
456 					0, NULL);
457 	if (!dma_region_table_cache)
458 		return -ENOMEM;
459 
460 	dma_page_table_cache = kmem_cache_create("PCI_DMA_page_tables",
461 					ZPCI_PT_SIZE, ZPCI_PT_ALIGN,
462 					0, NULL);
463 	if (!dma_page_table_cache) {
464 		kmem_cache_destroy(dma_region_table_cache);
465 		return -ENOMEM;
466 	}
467 	return 0;
468 }
469 
470 int __init zpci_dma_init(void)
471 {
472 	return dma_alloc_cpu_table_caches();
473 }
474 
475 void zpci_dma_exit(void)
476 {
477 	kmem_cache_destroy(dma_page_table_cache);
478 	kmem_cache_destroy(dma_region_table_cache);
479 }
480 
481 #define PREALLOC_DMA_DEBUG_ENTRIES	(1 << 16)
482 
483 static int __init dma_debug_do_init(void)
484 {
485 	dma_debug_init(PREALLOC_DMA_DEBUG_ENTRIES);
486 	return 0;
487 }
488 fs_initcall(dma_debug_do_init);
489 
490 struct dma_map_ops s390_dma_ops = {
491 	.alloc		= s390_dma_alloc,
492 	.free		= s390_dma_free,
493 	.map_sg		= s390_dma_map_sg,
494 	.unmap_sg	= s390_dma_unmap_sg,
495 	.map_page	= s390_dma_map_pages,
496 	.unmap_page	= s390_dma_unmap_pages,
497 	/* if we support direct DMA this must be conditional */
498 	.is_phys	= 0,
499 	/* dma_supported is unconditionally true without a callback */
500 };
501 EXPORT_SYMBOL_GPL(s390_dma_ops);
502