1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * IOMMU API for s390 PCI devices
4 *
5 * Copyright IBM Corp. 2015
6 * Author(s): Gerald Schaefer <gerald.schaefer@de.ibm.com>
7 */
8
9 #include <linux/pci.h>
10 #include <linux/iommu.h>
11 #include <linux/iommu-helper.h>
12 #include <linux/sizes.h>
13 #include <asm/pci_dma.h>
14
15 /*
16 * Physically contiguous memory regions can be mapped with 4 KiB alignment,
17 * we allow all page sizes that are an order of 4KiB (no special large page
18 * support so far).
19 */
20 #define S390_IOMMU_PGSIZES (~0xFFFUL)
21
22 static const struct iommu_ops s390_iommu_ops;
23
24 struct s390_domain {
25 struct iommu_domain domain;
26 struct list_head devices;
27 unsigned long *dma_table;
28 spinlock_t dma_table_lock;
29 spinlock_t list_lock;
30 };
31
32 struct s390_domain_device {
33 struct list_head list;
34 struct zpci_dev *zdev;
35 };
36
to_s390_domain(struct iommu_domain * dom)37 static struct s390_domain *to_s390_domain(struct iommu_domain *dom)
38 {
39 return container_of(dom, struct s390_domain, domain);
40 }
41
s390_iommu_capable(enum iommu_cap cap)42 static bool s390_iommu_capable(enum iommu_cap cap)
43 {
44 switch (cap) {
45 case IOMMU_CAP_CACHE_COHERENCY:
46 return true;
47 case IOMMU_CAP_INTR_REMAP:
48 return true;
49 default:
50 return false;
51 }
52 }
53
s390_domain_alloc(unsigned domain_type)54 static struct iommu_domain *s390_domain_alloc(unsigned domain_type)
55 {
56 struct s390_domain *s390_domain;
57
58 if (domain_type != IOMMU_DOMAIN_UNMANAGED)
59 return NULL;
60
61 s390_domain = kzalloc(sizeof(*s390_domain), GFP_KERNEL);
62 if (!s390_domain)
63 return NULL;
64
65 s390_domain->dma_table = dma_alloc_cpu_table();
66 if (!s390_domain->dma_table) {
67 kfree(s390_domain);
68 return NULL;
69 }
70
71 spin_lock_init(&s390_domain->dma_table_lock);
72 spin_lock_init(&s390_domain->list_lock);
73 INIT_LIST_HEAD(&s390_domain->devices);
74
75 return &s390_domain->domain;
76 }
77
s390_domain_free(struct iommu_domain * domain)78 static void s390_domain_free(struct iommu_domain *domain)
79 {
80 struct s390_domain *s390_domain = to_s390_domain(domain);
81
82 dma_cleanup_tables(s390_domain->dma_table);
83 kfree(s390_domain);
84 }
85
s390_iommu_attach_device(struct iommu_domain * domain,struct device * dev)86 static int s390_iommu_attach_device(struct iommu_domain *domain,
87 struct device *dev)
88 {
89 struct s390_domain *s390_domain = to_s390_domain(domain);
90 struct zpci_dev *zdev = to_pci_dev(dev)->sysdata;
91 struct s390_domain_device *domain_device;
92 unsigned long flags;
93 int rc;
94
95 if (!zdev)
96 return -ENODEV;
97
98 domain_device = kzalloc(sizeof(*domain_device), GFP_KERNEL);
99 if (!domain_device)
100 return -ENOMEM;
101
102 if (zdev->dma_table)
103 zpci_dma_exit_device(zdev);
104
105 zdev->dma_table = s390_domain->dma_table;
106 rc = zpci_register_ioat(zdev, 0, zdev->start_dma, zdev->end_dma,
107 (u64) zdev->dma_table);
108 if (rc)
109 goto out_restore;
110
111 spin_lock_irqsave(&s390_domain->list_lock, flags);
112 /* First device defines the DMA range limits */
113 if (list_empty(&s390_domain->devices)) {
114 domain->geometry.aperture_start = zdev->start_dma;
115 domain->geometry.aperture_end = zdev->end_dma;
116 domain->geometry.force_aperture = true;
117 /* Allow only devices with identical DMA range limits */
118 } else if (domain->geometry.aperture_start != zdev->start_dma ||
119 domain->geometry.aperture_end != zdev->end_dma) {
120 rc = -EINVAL;
121 spin_unlock_irqrestore(&s390_domain->list_lock, flags);
122 goto out_restore;
123 }
124 domain_device->zdev = zdev;
125 zdev->s390_domain = s390_domain;
126 list_add(&domain_device->list, &s390_domain->devices);
127 spin_unlock_irqrestore(&s390_domain->list_lock, flags);
128
129 return 0;
130
131 out_restore:
132 zpci_dma_init_device(zdev);
133 kfree(domain_device);
134
135 return rc;
136 }
137
s390_iommu_detach_device(struct iommu_domain * domain,struct device * dev)138 static void s390_iommu_detach_device(struct iommu_domain *domain,
139 struct device *dev)
140 {
141 struct s390_domain *s390_domain = to_s390_domain(domain);
142 struct zpci_dev *zdev = to_pci_dev(dev)->sysdata;
143 struct s390_domain_device *domain_device, *tmp;
144 unsigned long flags;
145 int found = 0;
146
147 if (!zdev)
148 return;
149
150 spin_lock_irqsave(&s390_domain->list_lock, flags);
151 list_for_each_entry_safe(domain_device, tmp, &s390_domain->devices,
152 list) {
153 if (domain_device->zdev == zdev) {
154 list_del(&domain_device->list);
155 kfree(domain_device);
156 found = 1;
157 break;
158 }
159 }
160 spin_unlock_irqrestore(&s390_domain->list_lock, flags);
161
162 if (found) {
163 zdev->s390_domain = NULL;
164 zpci_unregister_ioat(zdev, 0);
165 zpci_dma_init_device(zdev);
166 }
167 }
168
s390_iommu_add_device(struct device * dev)169 static int s390_iommu_add_device(struct device *dev)
170 {
171 struct iommu_group *group = iommu_group_get_for_dev(dev);
172 struct zpci_dev *zdev = to_pci_dev(dev)->sysdata;
173
174 if (IS_ERR(group))
175 return PTR_ERR(group);
176
177 iommu_group_put(group);
178 iommu_device_link(&zdev->iommu_dev, dev);
179
180 return 0;
181 }
182
s390_iommu_remove_device(struct device * dev)183 static void s390_iommu_remove_device(struct device *dev)
184 {
185 struct zpci_dev *zdev = to_pci_dev(dev)->sysdata;
186 struct iommu_domain *domain;
187
188 /*
189 * This is a workaround for a scenario where the IOMMU API common code
190 * "forgets" to call the detach_dev callback: After binding a device
191 * to vfio-pci and completing the VFIO_SET_IOMMU ioctl (which triggers
192 * the attach_dev), removing the device via
193 * "echo 1 > /sys/bus/pci/devices/.../remove" won't trigger detach_dev,
194 * only remove_device will be called via the BUS_NOTIFY_REMOVED_DEVICE
195 * notifier.
196 *
197 * So let's call detach_dev from here if it hasn't been called before.
198 */
199 if (zdev && zdev->s390_domain) {
200 domain = iommu_get_domain_for_dev(dev);
201 if (domain)
202 s390_iommu_detach_device(domain, dev);
203 }
204
205 iommu_device_unlink(&zdev->iommu_dev, dev);
206 iommu_group_remove_device(dev);
207 }
208
s390_iommu_update_trans(struct s390_domain * s390_domain,unsigned long pa,dma_addr_t dma_addr,size_t size,int flags)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
s390_iommu_map(struct iommu_domain * domain,unsigned long iova,phys_addr_t paddr,size_t size,int prot)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
s390_iommu_iova_to_phys(struct iommu_domain * domain,dma_addr_t iova)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
s390_iommu_unmap(struct iommu_domain * domain,unsigned long iova,size_t size,struct iommu_iotlb_gather * gather)316 static size_t s390_iommu_unmap(struct iommu_domain *domain,
317 unsigned long iova, size_t size,
318 struct iommu_iotlb_gather *gather)
319 {
320 struct s390_domain *s390_domain = to_s390_domain(domain);
321 int flags = ZPCI_PTE_INVALID;
322 phys_addr_t paddr;
323 int rc;
324
325 paddr = s390_iommu_iova_to_phys(domain, iova);
326 if (!paddr)
327 return 0;
328
329 rc = s390_iommu_update_trans(s390_domain, (unsigned long) paddr, iova,
330 size, flags);
331 if (rc)
332 return 0;
333
334 return size;
335 }
336
zpci_init_iommu(struct zpci_dev * zdev)337 int zpci_init_iommu(struct zpci_dev *zdev)
338 {
339 int rc = 0;
340
341 rc = iommu_device_sysfs_add(&zdev->iommu_dev, NULL, NULL,
342 "s390-iommu.%08x", zdev->fid);
343 if (rc)
344 goto out_err;
345
346 iommu_device_set_ops(&zdev->iommu_dev, &s390_iommu_ops);
347
348 rc = iommu_device_register(&zdev->iommu_dev);
349 if (rc)
350 goto out_sysfs;
351
352 return 0;
353
354 out_sysfs:
355 iommu_device_sysfs_remove(&zdev->iommu_dev);
356
357 out_err:
358 return rc;
359 }
360
zpci_destroy_iommu(struct zpci_dev * zdev)361 void zpci_destroy_iommu(struct zpci_dev *zdev)
362 {
363 iommu_device_unregister(&zdev->iommu_dev);
364 iommu_device_sysfs_remove(&zdev->iommu_dev);
365 }
366
367 static const struct iommu_ops s390_iommu_ops = {
368 .capable = s390_iommu_capable,
369 .domain_alloc = s390_domain_alloc,
370 .domain_free = s390_domain_free,
371 .attach_dev = s390_iommu_attach_device,
372 .detach_dev = s390_iommu_detach_device,
373 .map = s390_iommu_map,
374 .unmap = s390_iommu_unmap,
375 .iova_to_phys = s390_iommu_iova_to_phys,
376 .add_device = s390_iommu_add_device,
377 .remove_device = s390_iommu_remove_device,
378 .device_group = generic_device_group,
379 .pgsize_bitmap = S390_IOMMU_PGSIZES,
380 };
381
s390_iommu_init(void)382 static int __init s390_iommu_init(void)
383 {
384 return bus_set_iommu(&pci_bus_type, &s390_iommu_ops);
385 }
386 subsys_initcall(s390_iommu_init);
387