1 /*
2 * Copyright 2018 Advanced Micro Devices, Inc.
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
13 *
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20 * OTHER DEALINGS IN THE SOFTWARE.
21 */
22
23 #include <linux/printk.h>
24 #include <linux/device.h>
25 #include <linux/slab.h>
26 #include <linux/pci.h>
27 #include <linux/amd-iommu.h>
28 #include "kfd_priv.h"
29 #include "kfd_dbgmgr.h"
30 #include "kfd_topology.h"
31 #include "kfd_iommu.h"
32
33 static const u32 required_iommu_flags = AMD_IOMMU_DEVICE_FLAG_ATS_SUP |
34 AMD_IOMMU_DEVICE_FLAG_PRI_SUP |
35 AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
36
37 /** kfd_iommu_check_device - Check whether IOMMU is available for device
38 */
kfd_iommu_check_device(struct kfd_dev * kfd)39 int kfd_iommu_check_device(struct kfd_dev *kfd)
40 {
41 struct amd_iommu_device_info iommu_info;
42 int err;
43
44 if (!kfd->use_iommu_v2)
45 return -ENODEV;
46
47 iommu_info.flags = 0;
48 err = amd_iommu_device_info(kfd->pdev, &iommu_info);
49 if (err)
50 return err;
51
52 if ((iommu_info.flags & required_iommu_flags) != required_iommu_flags)
53 return -ENODEV;
54
55 return 0;
56 }
57
58 /** kfd_iommu_device_init - Initialize IOMMU for device
59 */
kfd_iommu_device_init(struct kfd_dev * kfd)60 int kfd_iommu_device_init(struct kfd_dev *kfd)
61 {
62 struct amd_iommu_device_info iommu_info;
63 unsigned int pasid_limit;
64 int err;
65
66 if (!kfd->use_iommu_v2)
67 return 0;
68
69 iommu_info.flags = 0;
70 err = amd_iommu_device_info(kfd->pdev, &iommu_info);
71 if (err < 0) {
72 dev_err(kfd_device,
73 "error getting iommu info. is the iommu enabled?\n");
74 return -ENODEV;
75 }
76
77 if ((iommu_info.flags & required_iommu_flags) != required_iommu_flags) {
78 dev_err(kfd_device,
79 "error required iommu flags ats %i, pri %i, pasid %i\n",
80 (iommu_info.flags & AMD_IOMMU_DEVICE_FLAG_ATS_SUP) != 0,
81 (iommu_info.flags & AMD_IOMMU_DEVICE_FLAG_PRI_SUP) != 0,
82 (iommu_info.flags & AMD_IOMMU_DEVICE_FLAG_PASID_SUP)
83 != 0);
84 return -ENODEV;
85 }
86
87 pasid_limit = min_t(unsigned int,
88 (unsigned int)(1 << kfd->device_info->max_pasid_bits),
89 iommu_info.max_pasids);
90
91 if (!kfd_set_pasid_limit(pasid_limit)) {
92 dev_err(kfd_device, "error setting pasid limit\n");
93 return -EBUSY;
94 }
95
96 return 0;
97 }
98
99 /** kfd_iommu_bind_process_to_device - Have the IOMMU bind a process
100 *
101 * Binds the given process to the given device using its PASID. This
102 * enables IOMMUv2 address translation for the process on the device.
103 *
104 * This function assumes that the process mutex is held.
105 */
kfd_iommu_bind_process_to_device(struct kfd_process_device * pdd)106 int kfd_iommu_bind_process_to_device(struct kfd_process_device *pdd)
107 {
108 struct kfd_dev *dev = pdd->dev;
109 struct kfd_process *p = pdd->process;
110 int err;
111
112 if (!dev->use_iommu_v2 || pdd->bound == PDD_BOUND)
113 return 0;
114
115 if (unlikely(pdd->bound == PDD_BOUND_SUSPENDED)) {
116 pr_err("Binding PDD_BOUND_SUSPENDED pdd is unexpected!\n");
117 return -EINVAL;
118 }
119
120 err = amd_iommu_bind_pasid(dev->pdev, p->pasid, p->lead_thread);
121 if (!err)
122 pdd->bound = PDD_BOUND;
123
124 return err;
125 }
126
127 /** kfd_iommu_unbind_process - Unbind process from all devices
128 *
129 * This removes all IOMMU device bindings of the process. To be used
130 * before process termination.
131 */
kfd_iommu_unbind_process(struct kfd_process * p)132 void kfd_iommu_unbind_process(struct kfd_process *p)
133 {
134 struct kfd_process_device *pdd;
135
136 list_for_each_entry(pdd, &p->per_device_data, per_device_list)
137 if (pdd->bound == PDD_BOUND)
138 amd_iommu_unbind_pasid(pdd->dev->pdev, p->pasid);
139 }
140
141 /* Callback for process shutdown invoked by the IOMMU driver */
iommu_pasid_shutdown_callback(struct pci_dev * pdev,u32 pasid)142 static void iommu_pasid_shutdown_callback(struct pci_dev *pdev, u32 pasid)
143 {
144 struct kfd_dev *dev = kfd_device_by_pci_dev(pdev);
145 struct kfd_process *p;
146 struct kfd_process_device *pdd;
147
148 if (!dev)
149 return;
150
151 /*
152 * Look for the process that matches the pasid. If there is no such
153 * process, we either released it in amdkfd's own notifier, or there
154 * is a bug. Unfortunately, there is no way to tell...
155 */
156 p = kfd_lookup_process_by_pasid(pasid);
157 if (!p)
158 return;
159
160 pr_debug("Unbinding process 0x%x from IOMMU\n", pasid);
161
162 mutex_lock(kfd_get_dbgmgr_mutex());
163
164 if (dev->dbgmgr && dev->dbgmgr->pasid == p->pasid) {
165 if (!kfd_dbgmgr_unregister(dev->dbgmgr, p)) {
166 kfd_dbgmgr_destroy(dev->dbgmgr);
167 dev->dbgmgr = NULL;
168 }
169 }
170
171 mutex_unlock(kfd_get_dbgmgr_mutex());
172
173 mutex_lock(&p->mutex);
174
175 pdd = kfd_get_process_device_data(dev, p);
176 if (pdd)
177 /* For GPU relying on IOMMU, we need to dequeue here
178 * when PASID is still bound.
179 */
180 kfd_process_dequeue_from_device(pdd);
181
182 mutex_unlock(&p->mutex);
183
184 kfd_unref_process(p);
185 }
186
187 /* This function called by IOMMU driver on PPR failure */
iommu_invalid_ppr_cb(struct pci_dev * pdev,u32 pasid,unsigned long address,u16 flags)188 static int iommu_invalid_ppr_cb(struct pci_dev *pdev, u32 pasid,
189 unsigned long address, u16 flags)
190 {
191 struct kfd_dev *dev;
192
193 dev_warn_ratelimited(kfd_device,
194 "Invalid PPR device %x:%x.%x pasid 0x%x address 0x%lX flags 0x%X",
195 pdev->bus->number,
196 PCI_SLOT(pdev->devfn),
197 PCI_FUNC(pdev->devfn),
198 pasid,
199 address,
200 flags);
201
202 dev = kfd_device_by_pci_dev(pdev);
203 if (!WARN_ON(!dev))
204 kfd_signal_iommu_event(dev, pasid, address,
205 flags & PPR_FAULT_WRITE, flags & PPR_FAULT_EXEC);
206
207 return AMD_IOMMU_INV_PRI_RSP_INVALID;
208 }
209
210 /*
211 * Bind processes do the device that have been temporarily unbound
212 * (PDD_BOUND_SUSPENDED) in kfd_unbind_processes_from_device.
213 */
kfd_bind_processes_to_device(struct kfd_dev * kfd)214 static int kfd_bind_processes_to_device(struct kfd_dev *kfd)
215 {
216 struct kfd_process_device *pdd;
217 struct kfd_process *p;
218 unsigned int temp;
219 int err = 0;
220
221 int idx = srcu_read_lock(&kfd_processes_srcu);
222
223 hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
224 mutex_lock(&p->mutex);
225 pdd = kfd_get_process_device_data(kfd, p);
226
227 if (WARN_ON(!pdd) || pdd->bound != PDD_BOUND_SUSPENDED) {
228 mutex_unlock(&p->mutex);
229 continue;
230 }
231
232 err = amd_iommu_bind_pasid(kfd->pdev, p->pasid,
233 p->lead_thread);
234 if (err < 0) {
235 pr_err("Unexpected pasid 0x%x binding failure\n",
236 p->pasid);
237 mutex_unlock(&p->mutex);
238 break;
239 }
240
241 pdd->bound = PDD_BOUND;
242 mutex_unlock(&p->mutex);
243 }
244
245 srcu_read_unlock(&kfd_processes_srcu, idx);
246
247 return err;
248 }
249
250 /*
251 * Mark currently bound processes as PDD_BOUND_SUSPENDED. These
252 * processes will be restored to PDD_BOUND state in
253 * kfd_bind_processes_to_device.
254 */
kfd_unbind_processes_from_device(struct kfd_dev * kfd)255 static void kfd_unbind_processes_from_device(struct kfd_dev *kfd)
256 {
257 struct kfd_process_device *pdd;
258 struct kfd_process *p;
259 unsigned int temp;
260
261 int idx = srcu_read_lock(&kfd_processes_srcu);
262
263 hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
264 mutex_lock(&p->mutex);
265 pdd = kfd_get_process_device_data(kfd, p);
266
267 if (WARN_ON(!pdd)) {
268 mutex_unlock(&p->mutex);
269 continue;
270 }
271
272 if (pdd->bound == PDD_BOUND)
273 pdd->bound = PDD_BOUND_SUSPENDED;
274 mutex_unlock(&p->mutex);
275 }
276
277 srcu_read_unlock(&kfd_processes_srcu, idx);
278 }
279
280 /** kfd_iommu_suspend - Prepare IOMMU for suspend
281 *
282 * This unbinds processes from the device and disables the IOMMU for
283 * the device.
284 */
kfd_iommu_suspend(struct kfd_dev * kfd)285 void kfd_iommu_suspend(struct kfd_dev *kfd)
286 {
287 if (!kfd->use_iommu_v2)
288 return;
289
290 kfd_unbind_processes_from_device(kfd);
291
292 amd_iommu_set_invalidate_ctx_cb(kfd->pdev, NULL);
293 amd_iommu_set_invalid_ppr_cb(kfd->pdev, NULL);
294 amd_iommu_free_device(kfd->pdev);
295 }
296
297 /** kfd_iommu_resume - Restore IOMMU after resume
298 *
299 * This reinitializes the IOMMU for the device and re-binds previously
300 * suspended processes to the device.
301 */
kfd_iommu_resume(struct kfd_dev * kfd)302 int kfd_iommu_resume(struct kfd_dev *kfd)
303 {
304 unsigned int pasid_limit;
305 int err;
306
307 if (!kfd->use_iommu_v2)
308 return 0;
309
310 pasid_limit = kfd_get_pasid_limit();
311
312 err = amd_iommu_init_device(kfd->pdev, pasid_limit);
313 if (err)
314 return -ENXIO;
315
316 amd_iommu_set_invalidate_ctx_cb(kfd->pdev,
317 iommu_pasid_shutdown_callback);
318 amd_iommu_set_invalid_ppr_cb(kfd->pdev,
319 iommu_invalid_ppr_cb);
320
321 err = kfd_bind_processes_to_device(kfd);
322 if (err) {
323 amd_iommu_set_invalidate_ctx_cb(kfd->pdev, NULL);
324 amd_iommu_set_invalid_ppr_cb(kfd->pdev, NULL);
325 amd_iommu_free_device(kfd->pdev);
326 return err;
327 }
328
329 return 0;
330 }
331
332 extern bool amd_iommu_pc_supported(void);
333 extern u8 amd_iommu_pc_get_max_banks(u16 devid);
334 extern u8 amd_iommu_pc_get_max_counters(u16 devid);
335
336 /** kfd_iommu_add_perf_counters - Add IOMMU performance counters to topology
337 */
kfd_iommu_add_perf_counters(struct kfd_topology_device * kdev)338 int kfd_iommu_add_perf_counters(struct kfd_topology_device *kdev)
339 {
340 struct kfd_perf_properties *props;
341
342 if (!(kdev->node_props.capability & HSA_CAP_ATS_PRESENT))
343 return 0;
344
345 if (!amd_iommu_pc_supported())
346 return 0;
347
348 props = kfd_alloc_struct(props);
349 if (!props)
350 return -ENOMEM;
351 strcpy(props->block_name, "iommu");
352 props->max_concurrent = amd_iommu_pc_get_max_banks(0) *
353 amd_iommu_pc_get_max_counters(0); /* assume one iommu */
354 list_add_tail(&props->list, &kdev->perf_props);
355
356 return 0;
357 }
358