Merge branch 'for-linus-4.6' of git://git.kernel.org/pub/scm/linux/kernel/git/mason...
[deliverable/linux.git] / arch / x86 / events / amd / iommu.c
1 /*
2 * Copyright (C) 2013 Advanced Micro Devices, Inc.
3 *
4 * Author: Steven Kinney <Steven.Kinney@amd.com>
5 * Author: Suravee Suthikulpanit <Suraveee.Suthikulpanit@amd.com>
6 *
7 * Perf: amd_iommu - AMD IOMMU Performance Counter PMU implementation
8 *
9 * This program is free software; you can redistribute it and/or modify
10 * it under the terms of the GNU General Public License version 2 as
11 * published by the Free Software Foundation.
12 */
13
14 #include <linux/perf_event.h>
15 #include <linux/module.h>
16 #include <linux/cpumask.h>
17 #include <linux/slab.h>
18
19 #include "../perf_event.h"
20 #include "iommu.h"
21
22 #define COUNTER_SHIFT 16
23
24 #define _GET_BANK(ev) ((u8)(ev->hw.extra_reg.reg >> 8))
25 #define _GET_CNTR(ev) ((u8)(ev->hw.extra_reg.reg))
26
27 /* iommu pmu config masks */
28 #define _GET_CSOURCE(ev) ((ev->hw.config & 0xFFULL))
29 #define _GET_DEVID(ev) ((ev->hw.config >> 8) & 0xFFFFULL)
30 #define _GET_PASID(ev) ((ev->hw.config >> 24) & 0xFFFFULL)
31 #define _GET_DOMID(ev) ((ev->hw.config >> 40) & 0xFFFFULL)
32 #define _GET_DEVID_MASK(ev) ((ev->hw.extra_reg.config) & 0xFFFFULL)
33 #define _GET_PASID_MASK(ev) ((ev->hw.extra_reg.config >> 16) & 0xFFFFULL)
34 #define _GET_DOMID_MASK(ev) ((ev->hw.extra_reg.config >> 32) & 0xFFFFULL)
35
36 static struct perf_amd_iommu __perf_iommu;
37
38 struct perf_amd_iommu {
39 struct pmu pmu;
40 u8 max_banks;
41 u8 max_counters;
42 u64 cntr_assign_mask;
43 raw_spinlock_t lock;
44 const struct attribute_group *attr_groups[4];
45 };
46
47 #define format_group attr_groups[0]
48 #define cpumask_group attr_groups[1]
49 #define events_group attr_groups[2]
50 #define null_group attr_groups[3]
51
52 /*---------------------------------------------
53 * sysfs format attributes
54 *---------------------------------------------*/
55 PMU_FORMAT_ATTR(csource, "config:0-7");
56 PMU_FORMAT_ATTR(devid, "config:8-23");
57 PMU_FORMAT_ATTR(pasid, "config:24-39");
58 PMU_FORMAT_ATTR(domid, "config:40-55");
59 PMU_FORMAT_ATTR(devid_mask, "config1:0-15");
60 PMU_FORMAT_ATTR(pasid_mask, "config1:16-31");
61 PMU_FORMAT_ATTR(domid_mask, "config1:32-47");
62
63 static struct attribute *iommu_format_attrs[] = {
64 &format_attr_csource.attr,
65 &format_attr_devid.attr,
66 &format_attr_pasid.attr,
67 &format_attr_domid.attr,
68 &format_attr_devid_mask.attr,
69 &format_attr_pasid_mask.attr,
70 &format_attr_domid_mask.attr,
71 NULL,
72 };
73
74 static struct attribute_group amd_iommu_format_group = {
75 .name = "format",
76 .attrs = iommu_format_attrs,
77 };
78
79 /*---------------------------------------------
80 * sysfs events attributes
81 *---------------------------------------------*/
82 struct amd_iommu_event_desc {
83 struct kobj_attribute attr;
84 const char *event;
85 };
86
87 static ssize_t _iommu_event_show(struct kobject *kobj,
88 struct kobj_attribute *attr, char *buf)
89 {
90 struct amd_iommu_event_desc *event =
91 container_of(attr, struct amd_iommu_event_desc, attr);
92 return sprintf(buf, "%s\n", event->event);
93 }
94
95 #define AMD_IOMMU_EVENT_DESC(_name, _event) \
96 { \
97 .attr = __ATTR(_name, 0444, _iommu_event_show, NULL), \
98 .event = _event, \
99 }
100
101 static struct amd_iommu_event_desc amd_iommu_v2_event_descs[] = {
102 AMD_IOMMU_EVENT_DESC(mem_pass_untrans, "csource=0x01"),
103 AMD_IOMMU_EVENT_DESC(mem_pass_pretrans, "csource=0x02"),
104 AMD_IOMMU_EVENT_DESC(mem_pass_excl, "csource=0x03"),
105 AMD_IOMMU_EVENT_DESC(mem_target_abort, "csource=0x04"),
106 AMD_IOMMU_EVENT_DESC(mem_trans_total, "csource=0x05"),
107 AMD_IOMMU_EVENT_DESC(mem_iommu_tlb_pte_hit, "csource=0x06"),
108 AMD_IOMMU_EVENT_DESC(mem_iommu_tlb_pte_mis, "csource=0x07"),
109 AMD_IOMMU_EVENT_DESC(mem_iommu_tlb_pde_hit, "csource=0x08"),
110 AMD_IOMMU_EVENT_DESC(mem_iommu_tlb_pde_mis, "csource=0x09"),
111 AMD_IOMMU_EVENT_DESC(mem_dte_hit, "csource=0x0a"),
112 AMD_IOMMU_EVENT_DESC(mem_dte_mis, "csource=0x0b"),
113 AMD_IOMMU_EVENT_DESC(page_tbl_read_tot, "csource=0x0c"),
114 AMD_IOMMU_EVENT_DESC(page_tbl_read_nst, "csource=0x0d"),
115 AMD_IOMMU_EVENT_DESC(page_tbl_read_gst, "csource=0x0e"),
116 AMD_IOMMU_EVENT_DESC(int_dte_hit, "csource=0x0f"),
117 AMD_IOMMU_EVENT_DESC(int_dte_mis, "csource=0x10"),
118 AMD_IOMMU_EVENT_DESC(cmd_processed, "csource=0x11"),
119 AMD_IOMMU_EVENT_DESC(cmd_processed_inv, "csource=0x12"),
120 AMD_IOMMU_EVENT_DESC(tlb_inv, "csource=0x13"),
121 { /* end: all zeroes */ },
122 };
123
124 /*---------------------------------------------
125 * sysfs cpumask attributes
126 *---------------------------------------------*/
127 static cpumask_t iommu_cpumask;
128
129 static ssize_t _iommu_cpumask_show(struct device *dev,
130 struct device_attribute *attr,
131 char *buf)
132 {
133 return cpumap_print_to_pagebuf(true, buf, &iommu_cpumask);
134 }
135 static DEVICE_ATTR(cpumask, S_IRUGO, _iommu_cpumask_show, NULL);
136
137 static struct attribute *iommu_cpumask_attrs[] = {
138 &dev_attr_cpumask.attr,
139 NULL,
140 };
141
142 static struct attribute_group amd_iommu_cpumask_group = {
143 .attrs = iommu_cpumask_attrs,
144 };
145
146 /*---------------------------------------------*/
147
148 static int get_next_avail_iommu_bnk_cntr(struct perf_amd_iommu *perf_iommu)
149 {
150 unsigned long flags;
151 int shift, bank, cntr, retval;
152 int max_banks = perf_iommu->max_banks;
153 int max_cntrs = perf_iommu->max_counters;
154
155 raw_spin_lock_irqsave(&perf_iommu->lock, flags);
156
157 for (bank = 0, shift = 0; bank < max_banks; bank++) {
158 for (cntr = 0; cntr < max_cntrs; cntr++) {
159 shift = bank + (bank*3) + cntr;
160 if (perf_iommu->cntr_assign_mask & (1ULL<<shift)) {
161 continue;
162 } else {
163 perf_iommu->cntr_assign_mask |= (1ULL<<shift);
164 retval = ((u16)((u16)bank<<8) | (u8)(cntr));
165 goto out;
166 }
167 }
168 }
169 retval = -ENOSPC;
170 out:
171 raw_spin_unlock_irqrestore(&perf_iommu->lock, flags);
172 return retval;
173 }
174
175 static int clear_avail_iommu_bnk_cntr(struct perf_amd_iommu *perf_iommu,
176 u8 bank, u8 cntr)
177 {
178 unsigned long flags;
179 int max_banks, max_cntrs;
180 int shift = 0;
181
182 max_banks = perf_iommu->max_banks;
183 max_cntrs = perf_iommu->max_counters;
184
185 if ((bank > max_banks) || (cntr > max_cntrs))
186 return -EINVAL;
187
188 shift = bank + cntr + (bank*3);
189
190 raw_spin_lock_irqsave(&perf_iommu->lock, flags);
191 perf_iommu->cntr_assign_mask &= ~(1ULL<<shift);
192 raw_spin_unlock_irqrestore(&perf_iommu->lock, flags);
193
194 return 0;
195 }
196
197 static int perf_iommu_event_init(struct perf_event *event)
198 {
199 struct hw_perf_event *hwc = &event->hw;
200 struct perf_amd_iommu *perf_iommu;
201 u64 config, config1;
202
203 /* test the event attr type check for PMU enumeration */
204 if (event->attr.type != event->pmu->type)
205 return -ENOENT;
206
207 /*
208 * IOMMU counters are shared across all cores.
209 * Therefore, it does not support per-process mode.
210 * Also, it does not support event sampling mode.
211 */
212 if (is_sampling_event(event) || event->attach_state & PERF_ATTACH_TASK)
213 return -EINVAL;
214
215 /* IOMMU counters do not have usr/os/guest/host bits */
216 if (event->attr.exclude_user || event->attr.exclude_kernel ||
217 event->attr.exclude_host || event->attr.exclude_guest)
218 return -EINVAL;
219
220 if (event->cpu < 0)
221 return -EINVAL;
222
223 perf_iommu = &__perf_iommu;
224
225 if (event->pmu != &perf_iommu->pmu)
226 return -ENOENT;
227
228 if (perf_iommu) {
229 config = event->attr.config;
230 config1 = event->attr.config1;
231 } else {
232 return -EINVAL;
233 }
234
235 /* integrate with iommu base devid (0000), assume one iommu */
236 perf_iommu->max_banks =
237 amd_iommu_pc_get_max_banks(IOMMU_BASE_DEVID);
238 perf_iommu->max_counters =
239 amd_iommu_pc_get_max_counters(IOMMU_BASE_DEVID);
240 if ((perf_iommu->max_banks == 0) || (perf_iommu->max_counters == 0))
241 return -EINVAL;
242
243 /* update the hw_perf_event struct with the iommu config data */
244 hwc->config = config;
245 hwc->extra_reg.config = config1;
246
247 return 0;
248 }
249
250 static void perf_iommu_enable_event(struct perf_event *ev)
251 {
252 u8 csource = _GET_CSOURCE(ev);
253 u16 devid = _GET_DEVID(ev);
254 u64 reg = 0ULL;
255
256 reg = csource;
257 amd_iommu_pc_get_set_reg_val(devid,
258 _GET_BANK(ev), _GET_CNTR(ev) ,
259 IOMMU_PC_COUNTER_SRC_REG, &reg, true);
260
261 reg = 0ULL | devid | (_GET_DEVID_MASK(ev) << 32);
262 if (reg)
263 reg |= (1UL << 31);
264 amd_iommu_pc_get_set_reg_val(devid,
265 _GET_BANK(ev), _GET_CNTR(ev) ,
266 IOMMU_PC_DEVID_MATCH_REG, &reg, true);
267
268 reg = 0ULL | _GET_PASID(ev) | (_GET_PASID_MASK(ev) << 32);
269 if (reg)
270 reg |= (1UL << 31);
271 amd_iommu_pc_get_set_reg_val(devid,
272 _GET_BANK(ev), _GET_CNTR(ev) ,
273 IOMMU_PC_PASID_MATCH_REG, &reg, true);
274
275 reg = 0ULL | _GET_DOMID(ev) | (_GET_DOMID_MASK(ev) << 32);
276 if (reg)
277 reg |= (1UL << 31);
278 amd_iommu_pc_get_set_reg_val(devid,
279 _GET_BANK(ev), _GET_CNTR(ev) ,
280 IOMMU_PC_DOMID_MATCH_REG, &reg, true);
281 }
282
283 static void perf_iommu_disable_event(struct perf_event *event)
284 {
285 u64 reg = 0ULL;
286
287 amd_iommu_pc_get_set_reg_val(_GET_DEVID(event),
288 _GET_BANK(event), _GET_CNTR(event),
289 IOMMU_PC_COUNTER_SRC_REG, &reg, true);
290 }
291
292 static void perf_iommu_start(struct perf_event *event, int flags)
293 {
294 struct hw_perf_event *hwc = &event->hw;
295
296 pr_debug("perf: amd_iommu:perf_iommu_start\n");
297 if (WARN_ON_ONCE(!(hwc->state & PERF_HES_STOPPED)))
298 return;
299
300 WARN_ON_ONCE(!(hwc->state & PERF_HES_UPTODATE));
301 hwc->state = 0;
302
303 if (flags & PERF_EF_RELOAD) {
304 u64 prev_raw_count = local64_read(&hwc->prev_count);
305 amd_iommu_pc_get_set_reg_val(_GET_DEVID(event),
306 _GET_BANK(event), _GET_CNTR(event),
307 IOMMU_PC_COUNTER_REG, &prev_raw_count, true);
308 }
309
310 perf_iommu_enable_event(event);
311 perf_event_update_userpage(event);
312
313 }
314
315 static void perf_iommu_read(struct perf_event *event)
316 {
317 u64 count = 0ULL;
318 u64 prev_raw_count = 0ULL;
319 u64 delta = 0ULL;
320 struct hw_perf_event *hwc = &event->hw;
321 pr_debug("perf: amd_iommu:perf_iommu_read\n");
322
323 amd_iommu_pc_get_set_reg_val(_GET_DEVID(event),
324 _GET_BANK(event), _GET_CNTR(event),
325 IOMMU_PC_COUNTER_REG, &count, false);
326
327 /* IOMMU pc counter register is only 48 bits */
328 count &= 0xFFFFFFFFFFFFULL;
329
330 prev_raw_count = local64_read(&hwc->prev_count);
331 if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
332 count) != prev_raw_count)
333 return;
334
335 /* Handling 48-bit counter overflowing */
336 delta = (count << COUNTER_SHIFT) - (prev_raw_count << COUNTER_SHIFT);
337 delta >>= COUNTER_SHIFT;
338 local64_add(delta, &event->count);
339
340 }
341
342 static void perf_iommu_stop(struct perf_event *event, int flags)
343 {
344 struct hw_perf_event *hwc = &event->hw;
345 u64 config;
346
347 pr_debug("perf: amd_iommu:perf_iommu_stop\n");
348
349 if (hwc->state & PERF_HES_UPTODATE)
350 return;
351
352 perf_iommu_disable_event(event);
353 WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
354 hwc->state |= PERF_HES_STOPPED;
355
356 if (hwc->state & PERF_HES_UPTODATE)
357 return;
358
359 config = hwc->config;
360 perf_iommu_read(event);
361 hwc->state |= PERF_HES_UPTODATE;
362 }
363
364 static int perf_iommu_add(struct perf_event *event, int flags)
365 {
366 int retval;
367 struct perf_amd_iommu *perf_iommu =
368 container_of(event->pmu, struct perf_amd_iommu, pmu);
369
370 pr_debug("perf: amd_iommu:perf_iommu_add\n");
371 event->hw.state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
372
373 /* request an iommu bank/counter */
374 retval = get_next_avail_iommu_bnk_cntr(perf_iommu);
375 if (retval != -ENOSPC)
376 event->hw.extra_reg.reg = (u16)retval;
377 else
378 return retval;
379
380 if (flags & PERF_EF_START)
381 perf_iommu_start(event, PERF_EF_RELOAD);
382
383 return 0;
384 }
385
386 static void perf_iommu_del(struct perf_event *event, int flags)
387 {
388 struct perf_amd_iommu *perf_iommu =
389 container_of(event->pmu, struct perf_amd_iommu, pmu);
390
391 pr_debug("perf: amd_iommu:perf_iommu_del\n");
392 perf_iommu_stop(event, PERF_EF_UPDATE);
393
394 /* clear the assigned iommu bank/counter */
395 clear_avail_iommu_bnk_cntr(perf_iommu,
396 _GET_BANK(event),
397 _GET_CNTR(event));
398
399 perf_event_update_userpage(event);
400 }
401
402 static __init int _init_events_attrs(struct perf_amd_iommu *perf_iommu)
403 {
404 struct attribute **attrs;
405 struct attribute_group *attr_group;
406 int i = 0, j;
407
408 while (amd_iommu_v2_event_descs[i].attr.attr.name)
409 i++;
410
411 attr_group = kzalloc(sizeof(struct attribute *)
412 * (i + 1) + sizeof(*attr_group), GFP_KERNEL);
413 if (!attr_group)
414 return -ENOMEM;
415
416 attrs = (struct attribute **)(attr_group + 1);
417 for (j = 0; j < i; j++)
418 attrs[j] = &amd_iommu_v2_event_descs[j].attr.attr;
419
420 attr_group->name = "events";
421 attr_group->attrs = attrs;
422 perf_iommu->events_group = attr_group;
423
424 return 0;
425 }
426
427 static __init void amd_iommu_pc_exit(void)
428 {
429 if (__perf_iommu.events_group != NULL) {
430 kfree(__perf_iommu.events_group);
431 __perf_iommu.events_group = NULL;
432 }
433 }
434
435 static __init int _init_perf_amd_iommu(
436 struct perf_amd_iommu *perf_iommu, char *name)
437 {
438 int ret;
439
440 raw_spin_lock_init(&perf_iommu->lock);
441
442 /* Init format attributes */
443 perf_iommu->format_group = &amd_iommu_format_group;
444
445 /* Init cpumask attributes to only core 0 */
446 cpumask_set_cpu(0, &iommu_cpumask);
447 perf_iommu->cpumask_group = &amd_iommu_cpumask_group;
448
449 /* Init events attributes */
450 if (_init_events_attrs(perf_iommu) != 0)
451 pr_err("perf: amd_iommu: Only support raw events.\n");
452
453 /* Init null attributes */
454 perf_iommu->null_group = NULL;
455 perf_iommu->pmu.attr_groups = perf_iommu->attr_groups;
456
457 ret = perf_pmu_register(&perf_iommu->pmu, name, -1);
458 if (ret) {
459 pr_err("perf: amd_iommu: Failed to initialized.\n");
460 amd_iommu_pc_exit();
461 } else {
462 pr_info("perf: amd_iommu: Detected. (%d banks, %d counters/bank)\n",
463 amd_iommu_pc_get_max_banks(IOMMU_BASE_DEVID),
464 amd_iommu_pc_get_max_counters(IOMMU_BASE_DEVID));
465 }
466
467 return ret;
468 }
469
470 static struct perf_amd_iommu __perf_iommu = {
471 .pmu = {
472 .event_init = perf_iommu_event_init,
473 .add = perf_iommu_add,
474 .del = perf_iommu_del,
475 .start = perf_iommu_start,
476 .stop = perf_iommu_stop,
477 .read = perf_iommu_read,
478 },
479 .max_banks = 0x00,
480 .max_counters = 0x00,
481 .cntr_assign_mask = 0ULL,
482 .format_group = NULL,
483 .cpumask_group = NULL,
484 .events_group = NULL,
485 .null_group = NULL,
486 };
487
488 static __init int amd_iommu_pc_init(void)
489 {
490 /* Make sure the IOMMU PC resource is available */
491 if (!amd_iommu_pc_supported())
492 return -ENODEV;
493
494 _init_perf_amd_iommu(&__perf_iommu, "amd_iommu");
495
496 return 0;
497 }
498
499 device_initcall(amd_iommu_pc_init);
This page took 0.041892 seconds and 6 git commands to generate.