xref: /openbmc/linux/drivers/iommu/s390-iommu.c (revision bb0eb050)
1 /*
2  * IOMMU API for s390 PCI devices
3  *
4  * Copyright IBM Corp. 2015
5  * Author(s): Gerald Schaefer <gerald.schaefer@de.ibm.com>
6  */
7 
8 #include <linux/pci.h>
9 #include <linux/iommu.h>
10 #include <linux/iommu-helper.h>
11 #include <linux/sizes.h>
12 #include <asm/pci_dma.h>
13 
14 /*
15  * Physically contiguous memory regions can be mapped with 4 KiB alignment,
16  * we allow all page sizes that are an order of 4KiB (no special large page
17  * support so far).
18  */
19 #define S390_IOMMU_PGSIZES	(~0xFFFUL)
20 
21 struct s390_domain {
22 	struct iommu_domain	domain;
23 	struct list_head	devices;
24 	unsigned long		*dma_table;
25 	spinlock_t		dma_table_lock;
26 	spinlock_t		list_lock;
27 };
28 
29 struct s390_domain_device {
30 	struct list_head	list;
31 	struct zpci_dev		*zdev;
32 };
33 
34 static struct s390_domain *to_s390_domain(struct iommu_domain *dom)
35 {
36 	return container_of(dom, struct s390_domain, domain);
37 }
38 
39 static bool s390_iommu_capable(enum iommu_cap cap)
40 {
41 	switch (cap) {
42 	case IOMMU_CAP_CACHE_COHERENCY:
43 		return true;
44 	case IOMMU_CAP_INTR_REMAP:
45 		return true;
46 	default:
47 		return false;
48 	}
49 }
50 
51 static struct iommu_domain *s390_domain_alloc(unsigned domain_type)
52 {
53 	struct s390_domain *s390_domain;
54 
55 	if (domain_type != IOMMU_DOMAIN_UNMANAGED)
56 		return NULL;
57 
58 	s390_domain = kzalloc(sizeof(*s390_domain), GFP_KERNEL);
59 	if (!s390_domain)
60 		return NULL;
61 
62 	s390_domain->dma_table = dma_alloc_cpu_table();
63 	if (!s390_domain->dma_table) {
64 		kfree(s390_domain);
65 		return NULL;
66 	}
67 
68 	spin_lock_init(&s390_domain->dma_table_lock);
69 	spin_lock_init(&s390_domain->list_lock);
70 	INIT_LIST_HEAD(&s390_domain->devices);
71 
72 	return &s390_domain->domain;
73 }
74 
75 static void s390_domain_free(struct iommu_domain *domain)
76 {
77 	struct s390_domain *s390_domain = to_s390_domain(domain);
78 
79 	dma_cleanup_tables(s390_domain->dma_table);
80 	kfree(s390_domain);
81 }
82 
83 static int s390_iommu_attach_device(struct iommu_domain *domain,
84 				    struct device *dev)
85 {
86 	struct s390_domain *s390_domain = to_s390_domain(domain);
87 	struct zpci_dev *zdev = to_pci_dev(dev)->sysdata;
88 	struct s390_domain_device *domain_device;
89 	unsigned long flags;
90 	int rc;
91 
92 	if (!zdev)
93 		return -ENODEV;
94 
95 	domain_device = kzalloc(sizeof(*domain_device), GFP_KERNEL);
96 	if (!domain_device)
97 		return -ENOMEM;
98 
99 	if (zdev->dma_table)
100 		zpci_dma_exit_device(zdev);
101 
102 	zdev->dma_table = s390_domain->dma_table;
103 	rc = zpci_register_ioat(zdev, 0, zdev->start_dma, zdev->end_dma,
104 				(u64) zdev->dma_table);
105 	if (rc)
106 		goto out_restore;
107 
108 	spin_lock_irqsave(&s390_domain->list_lock, flags);
109 	/* First device defines the DMA range limits */
110 	if (list_empty(&s390_domain->devices)) {
111 		domain->geometry.aperture_start = zdev->start_dma;
112 		domain->geometry.aperture_end = zdev->end_dma;
113 		domain->geometry.force_aperture = true;
114 	/* Allow only devices with identical DMA range limits */
115 	} else if (domain->geometry.aperture_start != zdev->start_dma ||
116 		   domain->geometry.aperture_end != zdev->end_dma) {
117 		rc = -EINVAL;
118 		spin_unlock_irqrestore(&s390_domain->list_lock, flags);
119 		goto out_restore;
120 	}
121 	domain_device->zdev = zdev;
122 	zdev->s390_domain = s390_domain;
123 	list_add(&domain_device->list, &s390_domain->devices);
124 	spin_unlock_irqrestore(&s390_domain->list_lock, flags);
125 
126 	return 0;
127 
128 out_restore:
129 	zpci_dma_init_device(zdev);
130 	kfree(domain_device);
131 
132 	return rc;
133 }
134 
135 static void s390_iommu_detach_device(struct iommu_domain *domain,
136 				     struct device *dev)
137 {
138 	struct s390_domain *s390_domain = to_s390_domain(domain);
139 	struct zpci_dev *zdev = to_pci_dev(dev)->sysdata;
140 	struct s390_domain_device *domain_device, *tmp;
141 	unsigned long flags;
142 	int found = 0;
143 
144 	if (!zdev)
145 		return;
146 
147 	spin_lock_irqsave(&s390_domain->list_lock, flags);
148 	list_for_each_entry_safe(domain_device, tmp, &s390_domain->devices,
149 				 list) {
150 		if (domain_device->zdev == zdev) {
151 			list_del(&domain_device->list);
152 			kfree(domain_device);
153 			found = 1;
154 			break;
155 		}
156 	}
157 	spin_unlock_irqrestore(&s390_domain->list_lock, flags);
158 
159 	if (found) {
160 		zdev->s390_domain = NULL;
161 		zpci_unregister_ioat(zdev, 0);
162 		zpci_dma_init_device(zdev);
163 	}
164 }
165 
166 static int s390_iommu_add_device(struct device *dev)
167 {
168 	struct iommu_group *group;
169 	int rc;
170 
171 	group = iommu_group_get(dev);
172 	if (!group) {
173 		group = iommu_group_alloc();
174 		if (IS_ERR(group))
175 			return PTR_ERR(group);
176 	}
177 
178 	rc = iommu_group_add_device(group, dev);
179 	iommu_group_put(group);
180 
181 	return rc;
182 }
183 
184 static void s390_iommu_remove_device(struct device *dev)
185 {
186 	struct zpci_dev *zdev = to_pci_dev(dev)->sysdata;
187 	struct iommu_domain *domain;
188 
189 	/*
190 	 * This is a workaround for a scenario where the IOMMU API common code
191 	 * "forgets" to call the detach_dev callback: After binding a device
192 	 * to vfio-pci and completing the VFIO_SET_IOMMU ioctl (which triggers
193 	 * the attach_dev), removing the device via
194 	 * "echo 1 > /sys/bus/pci/devices/.../remove" won't trigger detach_dev,
195 	 * only remove_device will be called via the BUS_NOTIFY_REMOVED_DEVICE
196 	 * notifier.
197 	 *
198 	 * So let's call detach_dev from here if it hasn't been called before.
199 	 */
200 	if (zdev && zdev->s390_domain) {
201 		domain = iommu_get_domain_for_dev(dev);
202 		if (domain)
203 			s390_iommu_detach_device(domain, dev);
204 	}
205 
206 	iommu_group_remove_device(dev);
207 }
208 
209 static int s390_iommu_update_trans(struct s390_domain *s390_domain,
210 				   unsigned long pa, dma_addr_t dma_addr,
211 				   size_t size, int flags)
212 {
213 	struct s390_domain_device *domain_device;
214 	u8 *page_addr = (u8 *) (pa & PAGE_MASK);
215 	dma_addr_t start_dma_addr = dma_addr;
216 	unsigned long irq_flags, nr_pages, i;
217 	unsigned long *entry;
218 	int rc = 0;
219 
220 	if (dma_addr < s390_domain->domain.geometry.aperture_start ||
221 	    dma_addr + size > s390_domain->domain.geometry.aperture_end)
222 		return -EINVAL;
223 
224 	nr_pages = PAGE_ALIGN(size) >> PAGE_SHIFT;
225 	if (!nr_pages)
226 		return 0;
227 
228 	spin_lock_irqsave(&s390_domain->dma_table_lock, irq_flags);
229 	for (i = 0; i < nr_pages; i++) {
230 		entry = dma_walk_cpu_trans(s390_domain->dma_table, dma_addr);
231 		if (!entry) {
232 			rc = -ENOMEM;
233 			goto undo_cpu_trans;
234 		}
235 		dma_update_cpu_trans(entry, page_addr, flags);
236 		page_addr += PAGE_SIZE;
237 		dma_addr += PAGE_SIZE;
238 	}
239 
240 	spin_lock(&s390_domain->list_lock);
241 	list_for_each_entry(domain_device, &s390_domain->devices, list) {
242 		rc = zpci_refresh_trans((u64) domain_device->zdev->fh << 32,
243 					start_dma_addr, nr_pages * PAGE_SIZE);
244 		if (rc)
245 			break;
246 	}
247 	spin_unlock(&s390_domain->list_lock);
248 
249 undo_cpu_trans:
250 	if (rc && ((flags & ZPCI_PTE_VALID_MASK) == ZPCI_PTE_VALID)) {
251 		flags = ZPCI_PTE_INVALID;
252 		while (i-- > 0) {
253 			page_addr -= PAGE_SIZE;
254 			dma_addr -= PAGE_SIZE;
255 			entry = dma_walk_cpu_trans(s390_domain->dma_table,
256 						   dma_addr);
257 			if (!entry)
258 				break;
259 			dma_update_cpu_trans(entry, page_addr, flags);
260 		}
261 	}
262 	spin_unlock_irqrestore(&s390_domain->dma_table_lock, irq_flags);
263 
264 	return rc;
265 }
266 
267 static int s390_iommu_map(struct iommu_domain *domain, unsigned long iova,
268 			  phys_addr_t paddr, size_t size, int prot)
269 {
270 	struct s390_domain *s390_domain = to_s390_domain(domain);
271 	int flags = ZPCI_PTE_VALID, rc = 0;
272 
273 	if (!(prot & IOMMU_READ))
274 		return -EINVAL;
275 
276 	if (!(prot & IOMMU_WRITE))
277 		flags |= ZPCI_TABLE_PROTECTED;
278 
279 	rc = s390_iommu_update_trans(s390_domain, (unsigned long) paddr, iova,
280 				     size, flags);
281 
282 	return rc;
283 }
284 
285 static phys_addr_t s390_iommu_iova_to_phys(struct iommu_domain *domain,
286 					   dma_addr_t iova)
287 {
288 	struct s390_domain *s390_domain = to_s390_domain(domain);
289 	unsigned long *sto, *pto, *rto, flags;
290 	unsigned int rtx, sx, px;
291 	phys_addr_t phys = 0;
292 
293 	if (iova < domain->geometry.aperture_start ||
294 	    iova > domain->geometry.aperture_end)
295 		return 0;
296 
297 	rtx = calc_rtx(iova);
298 	sx = calc_sx(iova);
299 	px = calc_px(iova);
300 	rto = s390_domain->dma_table;
301 
302 	spin_lock_irqsave(&s390_domain->dma_table_lock, flags);
303 	if (rto && reg_entry_isvalid(rto[rtx])) {
304 		sto = get_rt_sto(rto[rtx]);
305 		if (sto && reg_entry_isvalid(sto[sx])) {
306 			pto = get_st_pto(sto[sx]);
307 			if (pto && pt_entry_isvalid(pto[px]))
308 				phys = pto[px] & ZPCI_PTE_ADDR_MASK;
309 		}
310 	}
311 	spin_unlock_irqrestore(&s390_domain->dma_table_lock, flags);
312 
313 	return phys;
314 }
315 
316 static size_t s390_iommu_unmap(struct iommu_domain *domain,
317 			       unsigned long iova, size_t size)
318 {
319 	struct s390_domain *s390_domain = to_s390_domain(domain);
320 	int flags = ZPCI_PTE_INVALID;
321 	phys_addr_t paddr;
322 	int rc;
323 
324 	paddr = s390_iommu_iova_to_phys(domain, iova);
325 	if (!paddr)
326 		return 0;
327 
328 	rc = s390_iommu_update_trans(s390_domain, (unsigned long) paddr, iova,
329 				     size, flags);
330 	if (rc)
331 		return 0;
332 
333 	return size;
334 }
335 
336 static struct iommu_ops s390_iommu_ops = {
337 	.capable = s390_iommu_capable,
338 	.domain_alloc = s390_domain_alloc,
339 	.domain_free = s390_domain_free,
340 	.attach_dev = s390_iommu_attach_device,
341 	.detach_dev = s390_iommu_detach_device,
342 	.map = s390_iommu_map,
343 	.unmap = s390_iommu_unmap,
344 	.iova_to_phys = s390_iommu_iova_to_phys,
345 	.add_device = s390_iommu_add_device,
346 	.remove_device = s390_iommu_remove_device,
347 	.pgsize_bitmap = S390_IOMMU_PGSIZES,
348 };
349 
350 static int __init s390_iommu_init(void)
351 {
352 	return bus_set_iommu(&pci_bus_type, &s390_iommu_ops);
353 }
354 subsys_initcall(s390_iommu_init);
355