blob: a04d491cf431ca47badf462f972a73b94f8c577e [file] [log] [blame]
Gerald Schaefer8128f232015-08-27 15:33:03 +02001/*
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/pci.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
22struct s390_domain {
23 struct iommu_domain domain;
24 struct list_head devices;
25 unsigned long *dma_table;
26 spinlock_t dma_table_lock;
27 spinlock_t list_lock;
28};
29
30struct s390_domain_device {
31 struct list_head list;
32 struct zpci_dev *zdev;
33};
34
35static struct s390_domain *to_s390_domain(struct iommu_domain *dom)
36{
37 return container_of(dom, struct s390_domain, domain);
38}
39
40static bool s390_iommu_capable(enum iommu_cap cap)
41{
42 switch (cap) {
43 case IOMMU_CAP_CACHE_COHERENCY:
44 return true;
45 case IOMMU_CAP_INTR_REMAP:
46 return true;
47 default:
48 return false;
49 }
50}
51
Sebastian Ott7cd75782015-11-06 17:26:24 +010052static struct iommu_domain *s390_domain_alloc(unsigned domain_type)
Gerald Schaefer8128f232015-08-27 15:33:03 +020053{
54 struct s390_domain *s390_domain;
55
56 if (domain_type != IOMMU_DOMAIN_UNMANAGED)
57 return NULL;
58
59 s390_domain = kzalloc(sizeof(*s390_domain), GFP_KERNEL);
60 if (!s390_domain)
61 return NULL;
62
63 s390_domain->dma_table = dma_alloc_cpu_table();
64 if (!s390_domain->dma_table) {
65 kfree(s390_domain);
66 return NULL;
67 }
68
69 spin_lock_init(&s390_domain->dma_table_lock);
70 spin_lock_init(&s390_domain->list_lock);
71 INIT_LIST_HEAD(&s390_domain->devices);
72
73 return &s390_domain->domain;
74}
75
Sebastian Ott7cd75782015-11-06 17:26:24 +010076static void s390_domain_free(struct iommu_domain *domain)
Gerald Schaefer8128f232015-08-27 15:33:03 +020077{
78 struct s390_domain *s390_domain = to_s390_domain(domain);
79
80 dma_cleanup_tables(s390_domain->dma_table);
81 kfree(s390_domain);
82}
83
84static int s390_iommu_attach_device(struct iommu_domain *domain,
85 struct device *dev)
86{
87 struct s390_domain *s390_domain = to_s390_domain(domain);
88 struct zpci_dev *zdev = to_pci_dev(dev)->sysdata;
89 struct s390_domain_device *domain_device;
90 unsigned long flags;
91 int rc;
92
93 if (!zdev)
94 return -ENODEV;
95
96 domain_device = kzalloc(sizeof(*domain_device), GFP_KERNEL);
97 if (!domain_device)
98 return -ENOMEM;
99
100 if (zdev->dma_table)
101 zpci_dma_exit_device(zdev);
102
103 zdev->dma_table = s390_domain->dma_table;
104 rc = zpci_register_ioat(zdev, 0, zdev->start_dma + PAGE_OFFSET,
105 zdev->start_dma + zdev->iommu_size - 1,
106 (u64) zdev->dma_table);
107 if (rc)
108 goto out_restore;
109
110 spin_lock_irqsave(&s390_domain->list_lock, flags);
111 /* First device defines the DMA range limits */
112 if (list_empty(&s390_domain->devices)) {
113 domain->geometry.aperture_start = zdev->start_dma;
114 domain->geometry.aperture_end = zdev->end_dma;
115 domain->geometry.force_aperture = true;
116 /* Allow only devices with identical DMA range limits */
117 } else if (domain->geometry.aperture_start != zdev->start_dma ||
118 domain->geometry.aperture_end != zdev->end_dma) {
119 rc = -EINVAL;
120 spin_unlock_irqrestore(&s390_domain->list_lock, flags);
121 goto out_restore;
122 }
123 domain_device->zdev = zdev;
124 zdev->s390_domain = s390_domain;
125 list_add(&domain_device->list, &s390_domain->devices);
126 spin_unlock_irqrestore(&s390_domain->list_lock, flags);
127
128 return 0;
129
130out_restore:
131 zpci_dma_init_device(zdev);
132 kfree(domain_device);
133
134 return rc;
135}
136
137static void s390_iommu_detach_device(struct iommu_domain *domain,
138 struct device *dev)
139{
140 struct s390_domain *s390_domain = to_s390_domain(domain);
141 struct zpci_dev *zdev = to_pci_dev(dev)->sysdata;
142 struct s390_domain_device *domain_device, *tmp;
143 unsigned long flags;
144 int found = 0;
145
146 if (!zdev)
147 return;
148
149 spin_lock_irqsave(&s390_domain->list_lock, flags);
150 list_for_each_entry_safe(domain_device, tmp, &s390_domain->devices,
151 list) {
152 if (domain_device->zdev == zdev) {
153 list_del(&domain_device->list);
154 kfree(domain_device);
155 found = 1;
156 break;
157 }
158 }
159 spin_unlock_irqrestore(&s390_domain->list_lock, flags);
160
161 if (found) {
162 zdev->s390_domain = NULL;
163 zpci_unregister_ioat(zdev, 0);
164 zpci_dma_init_device(zdev);
165 }
166}
167
168static int s390_iommu_add_device(struct device *dev)
169{
170 struct iommu_group *group;
171 int rc;
172
173 group = iommu_group_get(dev);
174 if (!group) {
175 group = iommu_group_alloc();
176 if (IS_ERR(group))
177 return PTR_ERR(group);
178 }
179
180 rc = iommu_group_add_device(group, dev);
181 iommu_group_put(group);
182
183 return rc;
184}
185
186static void s390_iommu_remove_device(struct device *dev)
187{
188 struct zpci_dev *zdev = to_pci_dev(dev)->sysdata;
189 struct iommu_domain *domain;
190
191 /*
192 * This is a workaround for a scenario where the IOMMU API common code
193 * "forgets" to call the detach_dev callback: After binding a device
194 * to vfio-pci and completing the VFIO_SET_IOMMU ioctl (which triggers
195 * the attach_dev), removing the device via
196 * "echo 1 > /sys/bus/pci/devices/.../remove" won't trigger detach_dev,
197 * only remove_device will be called via the BUS_NOTIFY_REMOVED_DEVICE
198 * notifier.
199 *
200 * So let's call detach_dev from here if it hasn't been called before.
201 */
202 if (zdev && zdev->s390_domain) {
203 domain = iommu_get_domain_for_dev(dev);
204 if (domain)
205 s390_iommu_detach_device(domain, dev);
206 }
207
208 iommu_group_remove_device(dev);
209}
210
211static int s390_iommu_update_trans(struct s390_domain *s390_domain,
212 unsigned long pa, dma_addr_t dma_addr,
213 size_t size, int flags)
214{
215 struct s390_domain_device *domain_device;
216 u8 *page_addr = (u8 *) (pa & PAGE_MASK);
217 dma_addr_t start_dma_addr = dma_addr;
218 unsigned long irq_flags, nr_pages, i;
Sebastian Ott66728ee2015-10-26 11:19:13 +0100219 unsigned long *entry;
Gerald Schaefer8128f232015-08-27 15:33:03 +0200220 int rc = 0;
221
222 if (dma_addr < s390_domain->domain.geometry.aperture_start ||
223 dma_addr + size > s390_domain->domain.geometry.aperture_end)
224 return -EINVAL;
225
226 nr_pages = PAGE_ALIGN(size) >> PAGE_SHIFT;
227 if (!nr_pages)
228 return 0;
229
230 spin_lock_irqsave(&s390_domain->dma_table_lock, irq_flags);
231 for (i = 0; i < nr_pages; i++) {
Sebastian Ott66728ee2015-10-26 11:19:13 +0100232 entry = dma_walk_cpu_trans(s390_domain->dma_table, dma_addr);
233 if (!entry) {
234 rc = -ENOMEM;
235 goto undo_cpu_trans;
236 }
237 dma_update_cpu_trans(entry, page_addr, flags);
Gerald Schaefer8128f232015-08-27 15:33:03 +0200238 page_addr += PAGE_SIZE;
239 dma_addr += PAGE_SIZE;
240 }
241
242 spin_lock(&s390_domain->list_lock);
243 list_for_each_entry(domain_device, &s390_domain->devices, list) {
244 rc = zpci_refresh_trans((u64) domain_device->zdev->fh << 32,
245 start_dma_addr, nr_pages * PAGE_SIZE);
246 if (rc)
247 break;
248 }
249 spin_unlock(&s390_domain->list_lock);
Sebastian Ott66728ee2015-10-26 11:19:13 +0100250
251undo_cpu_trans:
252 if (rc && ((flags & ZPCI_PTE_VALID_MASK) == ZPCI_PTE_VALID)) {
253 flags = ZPCI_PTE_INVALID;
254 while (i-- > 0) {
255 page_addr -= PAGE_SIZE;
256 dma_addr -= PAGE_SIZE;
257 entry = dma_walk_cpu_trans(s390_domain->dma_table,
258 dma_addr);
259 if (!entry)
260 break;
261 dma_update_cpu_trans(entry, page_addr, flags);
262 }
263 }
Gerald Schaefer8128f232015-08-27 15:33:03 +0200264 spin_unlock_irqrestore(&s390_domain->dma_table_lock, irq_flags);
265
266 return rc;
267}
268
269static int s390_iommu_map(struct iommu_domain *domain, unsigned long iova,
270 phys_addr_t paddr, size_t size, int prot)
271{
272 struct s390_domain *s390_domain = to_s390_domain(domain);
273 int flags = ZPCI_PTE_VALID, rc = 0;
274
275 if (!(prot & IOMMU_READ))
276 return -EINVAL;
277
278 if (!(prot & IOMMU_WRITE))
279 flags |= ZPCI_TABLE_PROTECTED;
280
281 rc = s390_iommu_update_trans(s390_domain, (unsigned long) paddr, iova,
282 size, flags);
283
284 return rc;
285}
286
287static phys_addr_t s390_iommu_iova_to_phys(struct iommu_domain *domain,
288 dma_addr_t iova)
289{
290 struct s390_domain *s390_domain = to_s390_domain(domain);
291 unsigned long *sto, *pto, *rto, flags;
292 unsigned int rtx, sx, px;
293 phys_addr_t phys = 0;
294
295 if (iova < domain->geometry.aperture_start ||
296 iova > domain->geometry.aperture_end)
297 return 0;
298
299 rtx = calc_rtx(iova);
300 sx = calc_sx(iova);
301 px = calc_px(iova);
302 rto = s390_domain->dma_table;
303
304 spin_lock_irqsave(&s390_domain->dma_table_lock, flags);
305 if (rto && reg_entry_isvalid(rto[rtx])) {
306 sto = get_rt_sto(rto[rtx]);
307 if (sto && reg_entry_isvalid(sto[sx])) {
308 pto = get_st_pto(sto[sx]);
309 if (pto && pt_entry_isvalid(pto[px]))
310 phys = pto[px] & ZPCI_PTE_ADDR_MASK;
311 }
312 }
313 spin_unlock_irqrestore(&s390_domain->dma_table_lock, flags);
314
315 return phys;
316}
317
318static size_t s390_iommu_unmap(struct iommu_domain *domain,
319 unsigned long iova, size_t size)
320{
321 struct s390_domain *s390_domain = to_s390_domain(domain);
322 int flags = ZPCI_PTE_INVALID;
323 phys_addr_t paddr;
324 int rc;
325
326 paddr = s390_iommu_iova_to_phys(domain, iova);
327 if (!paddr)
328 return 0;
329
330 rc = s390_iommu_update_trans(s390_domain, (unsigned long) paddr, iova,
331 size, flags);
332 if (rc)
333 return 0;
334
335 return size;
336}
337
338static struct iommu_ops s390_iommu_ops = {
339 .capable = s390_iommu_capable,
340 .domain_alloc = s390_domain_alloc,
341 .domain_free = s390_domain_free,
342 .attach_dev = s390_iommu_attach_device,
343 .detach_dev = s390_iommu_detach_device,
344 .map = s390_iommu_map,
345 .unmap = s390_iommu_unmap,
346 .iova_to_phys = s390_iommu_iova_to_phys,
347 .add_device = s390_iommu_add_device,
348 .remove_device = s390_iommu_remove_device,
349 .pgsize_bitmap = S390_IOMMU_PGSIZES,
350};
351
352static int __init s390_iommu_init(void)
353{
354 return bus_set_iommu(&pci_bus_type, &s390_iommu_ops);
355}
356subsys_initcall(s390_iommu_init);