[go: nahoru, domu]

1/*
2 * IOMMU for IPMMU/IPMMUI
3 * Copyright (C) 2012  Hideki EIRAKU
4 *
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License as published by
7 * the Free Software Foundation; version 2 of the License.
8 */
9
10#include <linux/dma-mapping.h>
11#include <linux/io.h>
12#include <linux/iommu.h>
13#include <linux/platform_device.h>
14#include <linux/sizes.h>
15#include <linux/slab.h>
16#include <asm/dma-iommu.h>
17#include "shmobile-ipmmu.h"
18
19#define L1_SIZE CONFIG_SHMOBILE_IOMMU_L1SIZE
20#define L1_LEN (L1_SIZE / 4)
21#define L1_ALIGN L1_SIZE
22#define L2_SIZE SZ_1K
23#define L2_LEN (L2_SIZE / 4)
24#define L2_ALIGN L2_SIZE
25
26struct shmobile_iommu_domain_pgtable {
27	uint32_t *pgtable;
28	dma_addr_t handle;
29};
30
31struct shmobile_iommu_archdata {
32	struct list_head attached_list;
33	struct dma_iommu_mapping *iommu_mapping;
34	spinlock_t attach_lock;
35	struct shmobile_iommu_domain *attached;
36	int num_attached_devices;
37	struct shmobile_ipmmu *ipmmu;
38};
39
40struct shmobile_iommu_domain {
41	struct shmobile_iommu_domain_pgtable l1, l2[L1_LEN];
42	spinlock_t map_lock;
43	spinlock_t attached_list_lock;
44	struct list_head attached_list;
45};
46
47static struct shmobile_iommu_archdata *ipmmu_archdata;
48static struct kmem_cache *l1cache, *l2cache;
49
50static int pgtable_alloc(struct shmobile_iommu_domain_pgtable *pgtable,
51			 struct kmem_cache *cache, size_t size)
52{
53	pgtable->pgtable = kmem_cache_zalloc(cache, GFP_ATOMIC);
54	if (!pgtable->pgtable)
55		return -ENOMEM;
56	pgtable->handle = dma_map_single(NULL, pgtable->pgtable, size,
57					 DMA_TO_DEVICE);
58	return 0;
59}
60
61static void pgtable_free(struct shmobile_iommu_domain_pgtable *pgtable,
62			 struct kmem_cache *cache, size_t size)
63{
64	dma_unmap_single(NULL, pgtable->handle, size, DMA_TO_DEVICE);
65	kmem_cache_free(cache, pgtable->pgtable);
66}
67
68static uint32_t pgtable_read(struct shmobile_iommu_domain_pgtable *pgtable,
69			     unsigned int index)
70{
71	return pgtable->pgtable[index];
72}
73
74static void pgtable_write(struct shmobile_iommu_domain_pgtable *pgtable,
75			  unsigned int index, unsigned int count, uint32_t val)
76{
77	unsigned int i;
78
79	for (i = 0; i < count; i++)
80		pgtable->pgtable[index + i] = val;
81	dma_sync_single_for_device(NULL, pgtable->handle + index * sizeof(val),
82				   sizeof(val) * count, DMA_TO_DEVICE);
83}
84
85static int shmobile_iommu_domain_init(struct iommu_domain *domain)
86{
87	struct shmobile_iommu_domain *sh_domain;
88	int i, ret;
89
90	sh_domain = kmalloc(sizeof(*sh_domain), GFP_KERNEL);
91	if (!sh_domain)
92		return -ENOMEM;
93	ret = pgtable_alloc(&sh_domain->l1, l1cache, L1_SIZE);
94	if (ret < 0) {
95		kfree(sh_domain);
96		return ret;
97	}
98	for (i = 0; i < L1_LEN; i++)
99		sh_domain->l2[i].pgtable = NULL;
100	spin_lock_init(&sh_domain->map_lock);
101	spin_lock_init(&sh_domain->attached_list_lock);
102	INIT_LIST_HEAD(&sh_domain->attached_list);
103	domain->priv = sh_domain;
104	return 0;
105}
106
107static void shmobile_iommu_domain_destroy(struct iommu_domain *domain)
108{
109	struct shmobile_iommu_domain *sh_domain = domain->priv;
110	int i;
111
112	for (i = 0; i < L1_LEN; i++) {
113		if (sh_domain->l2[i].pgtable)
114			pgtable_free(&sh_domain->l2[i], l2cache, L2_SIZE);
115	}
116	pgtable_free(&sh_domain->l1, l1cache, L1_SIZE);
117	kfree(sh_domain);
118	domain->priv = NULL;
119}
120
121static int shmobile_iommu_attach_device(struct iommu_domain *domain,
122					struct device *dev)
123{
124	struct shmobile_iommu_archdata *archdata = dev->archdata.iommu;
125	struct shmobile_iommu_domain *sh_domain = domain->priv;
126	int ret = -EBUSY;
127
128	if (!archdata)
129		return -ENODEV;
130	spin_lock(&sh_domain->attached_list_lock);
131	spin_lock(&archdata->attach_lock);
132	if (archdata->attached != sh_domain) {
133		if (archdata->attached)
134			goto err;
135		ipmmu_tlb_set(archdata->ipmmu, sh_domain->l1.handle, L1_SIZE,
136			      0);
137		ipmmu_tlb_flush(archdata->ipmmu);
138		archdata->attached = sh_domain;
139		archdata->num_attached_devices = 0;
140		list_add(&archdata->attached_list, &sh_domain->attached_list);
141	}
142	archdata->num_attached_devices++;
143	ret = 0;
144err:
145	spin_unlock(&archdata->attach_lock);
146	spin_unlock(&sh_domain->attached_list_lock);
147	return ret;
148}
149
150static void shmobile_iommu_detach_device(struct iommu_domain *domain,
151					 struct device *dev)
152{
153	struct shmobile_iommu_archdata *archdata = dev->archdata.iommu;
154	struct shmobile_iommu_domain *sh_domain = domain->priv;
155
156	if (!archdata)
157		return;
158	spin_lock(&sh_domain->attached_list_lock);
159	spin_lock(&archdata->attach_lock);
160	archdata->num_attached_devices--;
161	if (!archdata->num_attached_devices) {
162		ipmmu_tlb_set(archdata->ipmmu, 0, 0, 0);
163		ipmmu_tlb_flush(archdata->ipmmu);
164		archdata->attached = NULL;
165		list_del(&archdata->attached_list);
166	}
167	spin_unlock(&archdata->attach_lock);
168	spin_unlock(&sh_domain->attached_list_lock);
169}
170
171static void domain_tlb_flush(struct shmobile_iommu_domain *sh_domain)
172{
173	struct shmobile_iommu_archdata *archdata;
174
175	spin_lock(&sh_domain->attached_list_lock);
176	list_for_each_entry(archdata, &sh_domain->attached_list, attached_list)
177		ipmmu_tlb_flush(archdata->ipmmu);
178	spin_unlock(&sh_domain->attached_list_lock);
179}
180
181static int l2alloc(struct shmobile_iommu_domain *sh_domain,
182		   unsigned int l1index)
183{
184	int ret;
185
186	if (!sh_domain->l2[l1index].pgtable) {
187		ret = pgtable_alloc(&sh_domain->l2[l1index], l2cache, L2_SIZE);
188		if (ret < 0)
189			return ret;
190	}
191	pgtable_write(&sh_domain->l1, l1index, 1,
192		      sh_domain->l2[l1index].handle | 0x1);
193	return 0;
194}
195
196static void l2realfree(struct shmobile_iommu_domain_pgtable *l2)
197{
198	if (l2->pgtable)
199		pgtable_free(l2, l2cache, L2_SIZE);
200}
201
202static void l2free(struct shmobile_iommu_domain *sh_domain,
203		   unsigned int l1index,
204		   struct shmobile_iommu_domain_pgtable *l2)
205{
206	pgtable_write(&sh_domain->l1, l1index, 1, 0);
207	if (sh_domain->l2[l1index].pgtable) {
208		*l2 = sh_domain->l2[l1index];
209		sh_domain->l2[l1index].pgtable = NULL;
210	}
211}
212
213static int shmobile_iommu_map(struct iommu_domain *domain, unsigned long iova,
214			      phys_addr_t paddr, size_t size, int prot)
215{
216	struct shmobile_iommu_domain_pgtable l2 = { .pgtable = NULL };
217	struct shmobile_iommu_domain *sh_domain = domain->priv;
218	unsigned int l1index, l2index;
219	int ret;
220
221	l1index = iova >> 20;
222	switch (size) {
223	case SZ_4K:
224		l2index = (iova >> 12) & 0xff;
225		spin_lock(&sh_domain->map_lock);
226		ret = l2alloc(sh_domain, l1index);
227		if (!ret)
228			pgtable_write(&sh_domain->l2[l1index], l2index, 1,
229				      paddr | 0xff2);
230		spin_unlock(&sh_domain->map_lock);
231		break;
232	case SZ_64K:
233		l2index = (iova >> 12) & 0xf0;
234		spin_lock(&sh_domain->map_lock);
235		ret = l2alloc(sh_domain, l1index);
236		if (!ret)
237			pgtable_write(&sh_domain->l2[l1index], l2index, 0x10,
238				      paddr | 0xff1);
239		spin_unlock(&sh_domain->map_lock);
240		break;
241	case SZ_1M:
242		spin_lock(&sh_domain->map_lock);
243		l2free(sh_domain, l1index, &l2);
244		pgtable_write(&sh_domain->l1, l1index, 1, paddr | 0xc02);
245		spin_unlock(&sh_domain->map_lock);
246		ret = 0;
247		break;
248	default:
249		ret = -EINVAL;
250	}
251	if (!ret)
252		domain_tlb_flush(sh_domain);
253	l2realfree(&l2);
254	return ret;
255}
256
257static size_t shmobile_iommu_unmap(struct iommu_domain *domain,
258				   unsigned long iova, size_t size)
259{
260	struct shmobile_iommu_domain_pgtable l2 = { .pgtable = NULL };
261	struct shmobile_iommu_domain *sh_domain = domain->priv;
262	unsigned int l1index, l2index;
263	uint32_t l2entry = 0;
264	size_t ret = 0;
265
266	l1index = iova >> 20;
267	if (!(iova & 0xfffff) && size >= SZ_1M) {
268		spin_lock(&sh_domain->map_lock);
269		l2free(sh_domain, l1index, &l2);
270		spin_unlock(&sh_domain->map_lock);
271		ret = SZ_1M;
272		goto done;
273	}
274	l2index = (iova >> 12) & 0xff;
275	spin_lock(&sh_domain->map_lock);
276	if (sh_domain->l2[l1index].pgtable)
277		l2entry = pgtable_read(&sh_domain->l2[l1index], l2index);
278	switch (l2entry & 3) {
279	case 1:
280		if (l2index & 0xf)
281			break;
282		pgtable_write(&sh_domain->l2[l1index], l2index, 0x10, 0);
283		ret = SZ_64K;
284		break;
285	case 2:
286		pgtable_write(&sh_domain->l2[l1index], l2index, 1, 0);
287		ret = SZ_4K;
288		break;
289	}
290	spin_unlock(&sh_domain->map_lock);
291done:
292	if (ret)
293		domain_tlb_flush(sh_domain);
294	l2realfree(&l2);
295	return ret;
296}
297
298static phys_addr_t shmobile_iommu_iova_to_phys(struct iommu_domain *domain,
299					       dma_addr_t iova)
300{
301	struct shmobile_iommu_domain *sh_domain = domain->priv;
302	uint32_t l1entry = 0, l2entry = 0;
303	unsigned int l1index, l2index;
304
305	l1index = iova >> 20;
306	l2index = (iova >> 12) & 0xff;
307	spin_lock(&sh_domain->map_lock);
308	if (sh_domain->l2[l1index].pgtable)
309		l2entry = pgtable_read(&sh_domain->l2[l1index], l2index);
310	else
311		l1entry = pgtable_read(&sh_domain->l1, l1index);
312	spin_unlock(&sh_domain->map_lock);
313	switch (l2entry & 3) {
314	case 1:
315		return (l2entry & ~0xffff) | (iova & 0xffff);
316	case 2:
317		return (l2entry & ~0xfff) | (iova & 0xfff);
318	default:
319		if ((l1entry & 3) == 2)
320			return (l1entry & ~0xfffff) | (iova & 0xfffff);
321		return 0;
322	}
323}
324
325static int find_dev_name(struct shmobile_ipmmu *ipmmu, const char *dev_name)
326{
327	unsigned int i, n = ipmmu->num_dev_names;
328
329	for (i = 0; i < n; i++) {
330		if (strcmp(ipmmu->dev_names[i], dev_name) == 0)
331			return 1;
332	}
333	return 0;
334}
335
336static int shmobile_iommu_add_device(struct device *dev)
337{
338	struct shmobile_iommu_archdata *archdata = ipmmu_archdata;
339	struct dma_iommu_mapping *mapping;
340
341	if (!find_dev_name(archdata->ipmmu, dev_name(dev)))
342		return 0;
343	mapping = archdata->iommu_mapping;
344	if (!mapping) {
345		mapping = arm_iommu_create_mapping(&platform_bus_type, 0,
346						   L1_LEN << 20);
347		if (IS_ERR(mapping))
348			return PTR_ERR(mapping);
349		archdata->iommu_mapping = mapping;
350	}
351	dev->archdata.iommu = archdata;
352	if (arm_iommu_attach_device(dev, mapping))
353		pr_err("arm_iommu_attach_device failed\n");
354	return 0;
355}
356
357static const struct iommu_ops shmobile_iommu_ops = {
358	.domain_init = shmobile_iommu_domain_init,
359	.domain_destroy = shmobile_iommu_domain_destroy,
360	.attach_dev = shmobile_iommu_attach_device,
361	.detach_dev = shmobile_iommu_detach_device,
362	.map = shmobile_iommu_map,
363	.unmap = shmobile_iommu_unmap,
364	.iova_to_phys = shmobile_iommu_iova_to_phys,
365	.add_device = shmobile_iommu_add_device,
366	.pgsize_bitmap = SZ_1M | SZ_64K | SZ_4K,
367};
368
369int ipmmu_iommu_init(struct shmobile_ipmmu *ipmmu)
370{
371	static struct shmobile_iommu_archdata *archdata;
372
373	l1cache = kmem_cache_create("shmobile-iommu-pgtable1", L1_SIZE,
374				    L1_ALIGN, SLAB_HWCACHE_ALIGN, NULL);
375	if (!l1cache)
376		return -ENOMEM;
377	l2cache = kmem_cache_create("shmobile-iommu-pgtable2", L2_SIZE,
378				    L2_ALIGN, SLAB_HWCACHE_ALIGN, NULL);
379	if (!l2cache) {
380		kmem_cache_destroy(l1cache);
381		return -ENOMEM;
382	}
383	archdata = kzalloc(sizeof(*archdata), GFP_KERNEL);
384	if (!archdata) {
385		kmem_cache_destroy(l1cache);
386		kmem_cache_destroy(l2cache);
387		return -ENOMEM;
388	}
389	spin_lock_init(&archdata->attach_lock);
390	archdata->ipmmu = ipmmu;
391	ipmmu_archdata = archdata;
392	bus_set_iommu(&platform_bus_type, &shmobile_iommu_ops);
393	return 0;
394}
395