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_zpci_dev(dev);
91 struct s390_domain_device *domain_device;
92 unsigned long flags;
93 int cc, 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 cc = zpci_dma_exit_device(zdev);
104 if (cc) {
105 rc = -EIO;
106 goto out_free;
107 }
108 }
109
110 zdev->dma_table = s390_domain->dma_table;
111 cc = zpci_register_ioat(zdev, 0, zdev->start_dma, zdev->end_dma,
112 (u64) zdev->dma_table);
113 if (cc) {
114 rc = -EIO;
115 goto out_restore;
116 }
117
118 spin_lock_irqsave(&s390_domain->list_lock, flags);
119 /* First device defines the DMA range limits */
120 if (list_empty(&s390_domain->devices)) {
121 domain->geometry.aperture_start = zdev->start_dma;
122 domain->geometry.aperture_end = zdev->end_dma;
123 domain->geometry.force_aperture = true;
124 /* Allow only devices with identical DMA range limits */
125 } else if (domain->geometry.aperture_start != zdev->start_dma ||
126 domain->geometry.aperture_end != zdev->end_dma) {
127 rc = -EINVAL;
128 spin_unlock_irqrestore(&s390_domain->list_lock, flags);
129 goto out_restore;
130 }
131 domain_device->zdev = zdev;
132 zdev->s390_domain = s390_domain;
133 list_add(&domain_device->list, &s390_domain->devices);
134 spin_unlock_irqrestore(&s390_domain->list_lock, flags);
135
136 return 0;
137
138 out_restore:
139 zpci_dma_init_device(zdev);
140 out_free:
141 kfree(domain_device);
142
143 return rc;
144 }
145
s390_iommu_detach_device(struct iommu_domain * domain,struct device * dev)146 static void s390_iommu_detach_device(struct iommu_domain *domain,
147 struct device *dev)
148 {
149 struct s390_domain *s390_domain = to_s390_domain(domain);
150 struct zpci_dev *zdev = to_zpci_dev(dev);
151 struct s390_domain_device *domain_device, *tmp;
152 unsigned long flags;
153 int found = 0;
154
155 if (!zdev)
156 return;
157
158 spin_lock_irqsave(&s390_domain->list_lock, flags);
159 list_for_each_entry_safe(domain_device, tmp, &s390_domain->devices,
160 list) {
161 if (domain_device->zdev == zdev) {
162 list_del(&domain_device->list);
163 kfree(domain_device);
164 found = 1;
165 break;
166 }
167 }
168 spin_unlock_irqrestore(&s390_domain->list_lock, flags);
169
170 if (found) {
171 zdev->s390_domain = NULL;
172 zpci_unregister_ioat(zdev, 0);
173 zpci_dma_init_device(zdev);
174 }
175 }
176
s390_iommu_probe_device(struct device * dev)177 static struct iommu_device *s390_iommu_probe_device(struct device *dev)
178 {
179 struct zpci_dev *zdev = to_zpci_dev(dev);
180
181 return &zdev->iommu_dev;
182 }
183
s390_iommu_release_device(struct device * dev)184 static void s390_iommu_release_device(struct device *dev)
185 {
186 struct zpci_dev *zdev = to_zpci_dev(dev);
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 release_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
s390_iommu_update_trans(struct s390_domain * s390_domain,unsigned long pa,dma_addr_t dma_addr,size_t size,int flags)207 static int s390_iommu_update_trans(struct s390_domain *s390_domain,
208 unsigned long pa, dma_addr_t dma_addr,
209 size_t size, int flags)
210 {
211 struct s390_domain_device *domain_device;
212 u8 *page_addr = (u8 *) (pa & PAGE_MASK);
213 dma_addr_t start_dma_addr = dma_addr;
214 unsigned long irq_flags, nr_pages, i;
215 unsigned long *entry;
216 int rc = 0;
217
218 if (dma_addr < s390_domain->domain.geometry.aperture_start ||
219 dma_addr + size > s390_domain->domain.geometry.aperture_end)
220 return -EINVAL;
221
222 nr_pages = PAGE_ALIGN(size) >> PAGE_SHIFT;
223 if (!nr_pages)
224 return 0;
225
226 spin_lock_irqsave(&s390_domain->dma_table_lock, irq_flags);
227 for (i = 0; i < nr_pages; i++) {
228 entry = dma_walk_cpu_trans(s390_domain->dma_table, dma_addr);
229 if (!entry) {
230 rc = -ENOMEM;
231 goto undo_cpu_trans;
232 }
233 dma_update_cpu_trans(entry, page_addr, flags);
234 page_addr += PAGE_SIZE;
235 dma_addr += PAGE_SIZE;
236 }
237
238 spin_lock(&s390_domain->list_lock);
239 list_for_each_entry(domain_device, &s390_domain->devices, list) {
240 rc = zpci_refresh_trans((u64) domain_device->zdev->fh << 32,
241 start_dma_addr, nr_pages * PAGE_SIZE);
242 if (rc)
243 break;
244 }
245 spin_unlock(&s390_domain->list_lock);
246
247 undo_cpu_trans:
248 if (rc && ((flags & ZPCI_PTE_VALID_MASK) == ZPCI_PTE_VALID)) {
249 flags = ZPCI_PTE_INVALID;
250 while (i-- > 0) {
251 page_addr -= PAGE_SIZE;
252 dma_addr -= PAGE_SIZE;
253 entry = dma_walk_cpu_trans(s390_domain->dma_table,
254 dma_addr);
255 if (!entry)
256 break;
257 dma_update_cpu_trans(entry, page_addr, flags);
258 }
259 }
260 spin_unlock_irqrestore(&s390_domain->dma_table_lock, irq_flags);
261
262 return rc;
263 }
264
s390_iommu_map(struct iommu_domain * domain,unsigned long iova,phys_addr_t paddr,size_t size,int prot,gfp_t gfp)265 static int s390_iommu_map(struct iommu_domain *domain, unsigned long iova,
266 phys_addr_t paddr, size_t size, int prot, gfp_t gfp)
267 {
268 struct s390_domain *s390_domain = to_s390_domain(domain);
269 int flags = ZPCI_PTE_VALID, rc = 0;
270
271 if (!(prot & IOMMU_READ))
272 return -EINVAL;
273
274 if (!(prot & IOMMU_WRITE))
275 flags |= ZPCI_TABLE_PROTECTED;
276
277 rc = s390_iommu_update_trans(s390_domain, (unsigned long) paddr, iova,
278 size, flags);
279
280 return rc;
281 }
282
s390_iommu_iova_to_phys(struct iommu_domain * domain,dma_addr_t iova)283 static phys_addr_t s390_iommu_iova_to_phys(struct iommu_domain *domain,
284 dma_addr_t iova)
285 {
286 struct s390_domain *s390_domain = to_s390_domain(domain);
287 unsigned long *sto, *pto, *rto, flags;
288 unsigned int rtx, sx, px;
289 phys_addr_t phys = 0;
290
291 if (iova < domain->geometry.aperture_start ||
292 iova > domain->geometry.aperture_end)
293 return 0;
294
295 rtx = calc_rtx(iova);
296 sx = calc_sx(iova);
297 px = calc_px(iova);
298 rto = s390_domain->dma_table;
299
300 spin_lock_irqsave(&s390_domain->dma_table_lock, flags);
301 if (rto && reg_entry_isvalid(rto[rtx])) {
302 sto = get_rt_sto(rto[rtx]);
303 if (sto && reg_entry_isvalid(sto[sx])) {
304 pto = get_st_pto(sto[sx]);
305 if (pto && pt_entry_isvalid(pto[px]))
306 phys = pto[px] & ZPCI_PTE_ADDR_MASK;
307 }
308 }
309 spin_unlock_irqrestore(&s390_domain->dma_table_lock, flags);
310
311 return phys;
312 }
313
s390_iommu_unmap(struct iommu_domain * domain,unsigned long iova,size_t size,struct iommu_iotlb_gather * gather)314 static size_t s390_iommu_unmap(struct iommu_domain *domain,
315 unsigned long iova, size_t size,
316 struct iommu_iotlb_gather *gather)
317 {
318 struct s390_domain *s390_domain = to_s390_domain(domain);
319 int flags = ZPCI_PTE_INVALID;
320 phys_addr_t paddr;
321 int rc;
322
323 paddr = s390_iommu_iova_to_phys(domain, iova);
324 if (!paddr)
325 return 0;
326
327 rc = s390_iommu_update_trans(s390_domain, (unsigned long) paddr, iova,
328 size, flags);
329 if (rc)
330 return 0;
331
332 return size;
333 }
334
zpci_init_iommu(struct zpci_dev * zdev)335 int zpci_init_iommu(struct zpci_dev *zdev)
336 {
337 int rc = 0;
338
339 rc = iommu_device_sysfs_add(&zdev->iommu_dev, NULL, NULL,
340 "s390-iommu.%08x", zdev->fid);
341 if (rc)
342 goto out_err;
343
344 rc = iommu_device_register(&zdev->iommu_dev, &s390_iommu_ops, NULL);
345 if (rc)
346 goto out_sysfs;
347
348 return 0;
349
350 out_sysfs:
351 iommu_device_sysfs_remove(&zdev->iommu_dev);
352
353 out_err:
354 return rc;
355 }
356
zpci_destroy_iommu(struct zpci_dev * zdev)357 void zpci_destroy_iommu(struct zpci_dev *zdev)
358 {
359 iommu_device_unregister(&zdev->iommu_dev);
360 iommu_device_sysfs_remove(&zdev->iommu_dev);
361 }
362
363 static const struct iommu_ops s390_iommu_ops = {
364 .capable = s390_iommu_capable,
365 .domain_alloc = s390_domain_alloc,
366 .domain_free = s390_domain_free,
367 .attach_dev = s390_iommu_attach_device,
368 .detach_dev = s390_iommu_detach_device,
369 .map = s390_iommu_map,
370 .unmap = s390_iommu_unmap,
371 .iova_to_phys = s390_iommu_iova_to_phys,
372 .probe_device = s390_iommu_probe_device,
373 .release_device = s390_iommu_release_device,
374 .device_group = generic_device_group,
375 .pgsize_bitmap = S390_IOMMU_PGSIZES,
376 };
377
s390_iommu_init(void)378 static int __init s390_iommu_init(void)
379 {
380 return bus_set_iommu(&pci_bus_type, &s390_iommu_ops);
381 }
382 subsys_initcall(s390_iommu_init);
383
