xref: /linux/arch/x86/events/amd/iommu.c (revision b7019ac550eb3916f34d79db583e9b7ea2524afa)
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 #define pr_fmt(fmt)	"perf/amd_iommu: " fmt
15 
16 #include <linux/perf_event.h>
17 #include <linux/init.h>
18 #include <linux/cpumask.h>
19 #include <linux/slab.h>
20 
21 #include "../perf_event.h"
22 #include "iommu.h"
23 
24 #define COUNTER_SHIFT		16
25 
26 /* iommu pmu conf masks */
27 #define GET_CSOURCE(x)     ((x)->conf & 0xFFULL)
28 #define GET_DEVID(x)       (((x)->conf >> 8)  & 0xFFFFULL)
29 #define GET_DOMID(x)       (((x)->conf >> 24) & 0xFFFFULL)
30 #define GET_PASID(x)       (((x)->conf >> 40) & 0xFFFFFULL)
31 
32 /* iommu pmu conf1 masks */
33 #define GET_DEVID_MASK(x)  ((x)->conf1  & 0xFFFFULL)
34 #define GET_DOMID_MASK(x)  (((x)->conf1 >> 16) & 0xFFFFULL)
35 #define GET_PASID_MASK(x)  (((x)->conf1 >> 32) & 0xFFFFFULL)
36 
37 #define IOMMU_NAME_SIZE 16
38 
39 struct perf_amd_iommu {
40 	struct list_head list;
41 	struct pmu pmu;
42 	struct amd_iommu *iommu;
43 	char name[IOMMU_NAME_SIZE];
44 	u8 max_banks;
45 	u8 max_counters;
46 	u64 cntr_assign_mask;
47 	raw_spinlock_t lock;
48 };
49 
50 static LIST_HEAD(perf_amd_iommu_list);
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(domid,      "config:24-39");
58 PMU_FORMAT_ATTR(pasid,      "config:40-59");
59 PMU_FORMAT_ATTR(devid_mask, "config1:0-15");
60 PMU_FORMAT_ATTR(domid_mask, "config1:16-31");
61 PMU_FORMAT_ATTR(pasid_mask, "config1:32-51");
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 static struct attribute_group amd_iommu_events_group = {
83 	.name = "events",
84 };
85 
86 struct amd_iommu_event_desc {
87 	struct kobj_attribute attr;
88 	const char *event;
89 };
90 
91 static ssize_t _iommu_event_show(struct kobject *kobj,
92 				struct kobj_attribute *attr, char *buf)
93 {
94 	struct amd_iommu_event_desc *event =
95 		container_of(attr, struct amd_iommu_event_desc, attr);
96 	return sprintf(buf, "%s\n", event->event);
97 }
98 
99 #define AMD_IOMMU_EVENT_DESC(_name, _event)			\
100 {								\
101 	.attr  = __ATTR(_name, 0444, _iommu_event_show, NULL),	\
102 	.event = _event,					\
103 }
104 
105 static struct amd_iommu_event_desc amd_iommu_v2_event_descs[] = {
106 	AMD_IOMMU_EVENT_DESC(mem_pass_untrans,        "csource=0x01"),
107 	AMD_IOMMU_EVENT_DESC(mem_pass_pretrans,       "csource=0x02"),
108 	AMD_IOMMU_EVENT_DESC(mem_pass_excl,           "csource=0x03"),
109 	AMD_IOMMU_EVENT_DESC(mem_target_abort,        "csource=0x04"),
110 	AMD_IOMMU_EVENT_DESC(mem_trans_total,         "csource=0x05"),
111 	AMD_IOMMU_EVENT_DESC(mem_iommu_tlb_pte_hit,   "csource=0x06"),
112 	AMD_IOMMU_EVENT_DESC(mem_iommu_tlb_pte_mis,   "csource=0x07"),
113 	AMD_IOMMU_EVENT_DESC(mem_iommu_tlb_pde_hit,   "csource=0x08"),
114 	AMD_IOMMU_EVENT_DESC(mem_iommu_tlb_pde_mis,   "csource=0x09"),
115 	AMD_IOMMU_EVENT_DESC(mem_dte_hit,             "csource=0x0a"),
116 	AMD_IOMMU_EVENT_DESC(mem_dte_mis,             "csource=0x0b"),
117 	AMD_IOMMU_EVENT_DESC(page_tbl_read_tot,       "csource=0x0c"),
118 	AMD_IOMMU_EVENT_DESC(page_tbl_read_nst,       "csource=0x0d"),
119 	AMD_IOMMU_EVENT_DESC(page_tbl_read_gst,       "csource=0x0e"),
120 	AMD_IOMMU_EVENT_DESC(int_dte_hit,             "csource=0x0f"),
121 	AMD_IOMMU_EVENT_DESC(int_dte_mis,             "csource=0x10"),
122 	AMD_IOMMU_EVENT_DESC(cmd_processed,           "csource=0x11"),
123 	AMD_IOMMU_EVENT_DESC(cmd_processed_inv,       "csource=0x12"),
124 	AMD_IOMMU_EVENT_DESC(tlb_inv,                 "csource=0x13"),
125 	AMD_IOMMU_EVENT_DESC(ign_rd_wr_mmio_1ff8h,    "csource=0x14"),
126 	AMD_IOMMU_EVENT_DESC(vapic_int_non_guest,     "csource=0x15"),
127 	AMD_IOMMU_EVENT_DESC(vapic_int_guest,         "csource=0x16"),
128 	AMD_IOMMU_EVENT_DESC(smi_recv,                "csource=0x17"),
129 	AMD_IOMMU_EVENT_DESC(smi_blk,                 "csource=0x18"),
130 	{ /* end: all zeroes */ },
131 };
132 
133 /*---------------------------------------------
134  * sysfs cpumask attributes
135  *---------------------------------------------*/
136 static cpumask_t iommu_cpumask;
137 
138 static ssize_t _iommu_cpumask_show(struct device *dev,
139 				   struct device_attribute *attr,
140 				   char *buf)
141 {
142 	return cpumap_print_to_pagebuf(true, buf, &iommu_cpumask);
143 }
144 static DEVICE_ATTR(cpumask, S_IRUGO, _iommu_cpumask_show, NULL);
145 
146 static struct attribute *iommu_cpumask_attrs[] = {
147 	&dev_attr_cpumask.attr,
148 	NULL,
149 };
150 
151 static struct attribute_group amd_iommu_cpumask_group = {
152 	.attrs = iommu_cpumask_attrs,
153 };
154 
155 /*---------------------------------------------*/
156 
157 static int get_next_avail_iommu_bnk_cntr(struct perf_event *event)
158 {
159 	struct perf_amd_iommu *piommu = container_of(event->pmu, struct perf_amd_iommu, pmu);
160 	int max_cntrs = piommu->max_counters;
161 	int max_banks = piommu->max_banks;
162 	u32 shift, bank, cntr;
163 	unsigned long flags;
164 	int retval;
165 
166 	raw_spin_lock_irqsave(&piommu->lock, flags);
167 
168 	for (bank = 0, shift = 0; bank < max_banks; bank++) {
169 		for (cntr = 0; cntr < max_cntrs; cntr++) {
170 			shift = bank + (bank*3) + cntr;
171 			if (piommu->cntr_assign_mask & BIT_ULL(shift)) {
172 				continue;
173 			} else {
174 				piommu->cntr_assign_mask |= BIT_ULL(shift);
175 				event->hw.iommu_bank = bank;
176 				event->hw.iommu_cntr = cntr;
177 				retval = 0;
178 				goto out;
179 			}
180 		}
181 	}
182 	retval = -ENOSPC;
183 out:
184 	raw_spin_unlock_irqrestore(&piommu->lock, flags);
185 	return retval;
186 }
187 
188 static int clear_avail_iommu_bnk_cntr(struct perf_amd_iommu *perf_iommu,
189 					u8 bank, u8 cntr)
190 {
191 	unsigned long flags;
192 	int max_banks, max_cntrs;
193 	int shift = 0;
194 
195 	max_banks = perf_iommu->max_banks;
196 	max_cntrs = perf_iommu->max_counters;
197 
198 	if ((bank > max_banks) || (cntr > max_cntrs))
199 		return -EINVAL;
200 
201 	shift = bank + cntr + (bank*3);
202 
203 	raw_spin_lock_irqsave(&perf_iommu->lock, flags);
204 	perf_iommu->cntr_assign_mask &= ~(1ULL<<shift);
205 	raw_spin_unlock_irqrestore(&perf_iommu->lock, flags);
206 
207 	return 0;
208 }
209 
210 static int perf_iommu_event_init(struct perf_event *event)
211 {
212 	struct hw_perf_event *hwc = &event->hw;
213 
214 	/* test the event attr type check for PMU enumeration */
215 	if (event->attr.type != event->pmu->type)
216 		return -ENOENT;
217 
218 	/*
219 	 * IOMMU counters are shared across all cores.
220 	 * Therefore, it does not support per-process mode.
221 	 * Also, it does not support event sampling mode.
222 	 */
223 	if (is_sampling_event(event) || event->attach_state & PERF_ATTACH_TASK)
224 		return -EINVAL;
225 
226 	if (event->cpu < 0)
227 		return -EINVAL;
228 
229 	/* update the hw_perf_event struct with the iommu config data */
230 	hwc->conf  = event->attr.config;
231 	hwc->conf1 = event->attr.config1;
232 
233 	return 0;
234 }
235 
236 static inline struct amd_iommu *perf_event_2_iommu(struct perf_event *ev)
237 {
238 	return (container_of(ev->pmu, struct perf_amd_iommu, pmu))->iommu;
239 }
240 
241 static void perf_iommu_enable_event(struct perf_event *ev)
242 {
243 	struct amd_iommu *iommu = perf_event_2_iommu(ev);
244 	struct hw_perf_event *hwc = &ev->hw;
245 	u8 bank = hwc->iommu_bank;
246 	u8 cntr = hwc->iommu_cntr;
247 	u64 reg = 0ULL;
248 
249 	reg = GET_CSOURCE(hwc);
250 	amd_iommu_pc_set_reg(iommu, bank, cntr, IOMMU_PC_COUNTER_SRC_REG, &reg);
251 
252 	reg = GET_DEVID_MASK(hwc);
253 	reg = GET_DEVID(hwc) | (reg << 32);
254 	if (reg)
255 		reg |= BIT(31);
256 	amd_iommu_pc_set_reg(iommu, bank, cntr, IOMMU_PC_DEVID_MATCH_REG, &reg);
257 
258 	reg = GET_PASID_MASK(hwc);
259 	reg = GET_PASID(hwc) | (reg << 32);
260 	if (reg)
261 		reg |= BIT(31);
262 	amd_iommu_pc_set_reg(iommu, bank, cntr, IOMMU_PC_PASID_MATCH_REG, &reg);
263 
264 	reg = GET_DOMID_MASK(hwc);
265 	reg = GET_DOMID(hwc) | (reg << 32);
266 	if (reg)
267 		reg |= BIT(31);
268 	amd_iommu_pc_set_reg(iommu, bank, cntr, IOMMU_PC_DOMID_MATCH_REG, &reg);
269 }
270 
271 static void perf_iommu_disable_event(struct perf_event *event)
272 {
273 	struct amd_iommu *iommu = perf_event_2_iommu(event);
274 	struct hw_perf_event *hwc = &event->hw;
275 	u64 reg = 0ULL;
276 
277 	amd_iommu_pc_set_reg(iommu, hwc->iommu_bank, hwc->iommu_cntr,
278 			     IOMMU_PC_COUNTER_SRC_REG, &reg);
279 }
280 
281 static void perf_iommu_start(struct perf_event *event, int flags)
282 {
283 	struct hw_perf_event *hwc = &event->hw;
284 
285 	if (WARN_ON_ONCE(!(hwc->state & PERF_HES_STOPPED)))
286 		return;
287 
288 	WARN_ON_ONCE(!(hwc->state & PERF_HES_UPTODATE));
289 	hwc->state = 0;
290 
291 	if (flags & PERF_EF_RELOAD) {
292 		u64 prev_raw_count = local64_read(&hwc->prev_count);
293 		struct amd_iommu *iommu = perf_event_2_iommu(event);
294 
295 		amd_iommu_pc_set_reg(iommu, hwc->iommu_bank, hwc->iommu_cntr,
296 				     IOMMU_PC_COUNTER_REG, &prev_raw_count);
297 	}
298 
299 	perf_iommu_enable_event(event);
300 	perf_event_update_userpage(event);
301 
302 }
303 
304 static void perf_iommu_read(struct perf_event *event)
305 {
306 	u64 count, prev, delta;
307 	struct hw_perf_event *hwc = &event->hw;
308 	struct amd_iommu *iommu = perf_event_2_iommu(event);
309 
310 	if (amd_iommu_pc_get_reg(iommu, hwc->iommu_bank, hwc->iommu_cntr,
311 				 IOMMU_PC_COUNTER_REG, &count))
312 		return;
313 
314 	/* IOMMU pc counter register is only 48 bits */
315 	count &= GENMASK_ULL(47, 0);
316 
317 	prev = local64_read(&hwc->prev_count);
318 	if (local64_cmpxchg(&hwc->prev_count, prev, count) != prev)
319 		return;
320 
321 	/* Handle 48-bit counter overflow */
322 	delta = (count << COUNTER_SHIFT) - (prev << COUNTER_SHIFT);
323 	delta >>= COUNTER_SHIFT;
324 	local64_add(delta, &event->count);
325 }
326 
327 static void perf_iommu_stop(struct perf_event *event, int flags)
328 {
329 	struct hw_perf_event *hwc = &event->hw;
330 
331 	if (hwc->state & PERF_HES_UPTODATE)
332 		return;
333 
334 	perf_iommu_disable_event(event);
335 	WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
336 	hwc->state |= PERF_HES_STOPPED;
337 
338 	if (hwc->state & PERF_HES_UPTODATE)
339 		return;
340 
341 	perf_iommu_read(event);
342 	hwc->state |= PERF_HES_UPTODATE;
343 }
344 
345 static int perf_iommu_add(struct perf_event *event, int flags)
346 {
347 	int retval;
348 
349 	event->hw.state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
350 
351 	/* request an iommu bank/counter */
352 	retval = get_next_avail_iommu_bnk_cntr(event);
353 	if (retval)
354 		return retval;
355 
356 	if (flags & PERF_EF_START)
357 		perf_iommu_start(event, PERF_EF_RELOAD);
358 
359 	return 0;
360 }
361 
362 static void perf_iommu_del(struct perf_event *event, int flags)
363 {
364 	struct hw_perf_event *hwc = &event->hw;
365 	struct perf_amd_iommu *perf_iommu =
366 			container_of(event->pmu, struct perf_amd_iommu, pmu);
367 
368 	perf_iommu_stop(event, PERF_EF_UPDATE);
369 
370 	/* clear the assigned iommu bank/counter */
371 	clear_avail_iommu_bnk_cntr(perf_iommu,
372 				   hwc->iommu_bank, hwc->iommu_cntr);
373 
374 	perf_event_update_userpage(event);
375 }
376 
377 static __init int _init_events_attrs(void)
378 {
379 	int i = 0, j;
380 	struct attribute **attrs;
381 
382 	while (amd_iommu_v2_event_descs[i].attr.attr.name)
383 		i++;
384 
385 	attrs = kcalloc(i + 1, sizeof(struct attribute **), GFP_KERNEL);
386 	if (!attrs)
387 		return -ENOMEM;
388 
389 	for (j = 0; j < i; j++)
390 		attrs[j] = &amd_iommu_v2_event_descs[j].attr.attr;
391 
392 	amd_iommu_events_group.attrs = attrs;
393 	return 0;
394 }
395 
396 static const struct attribute_group *amd_iommu_attr_groups[] = {
397 	&amd_iommu_format_group,
398 	&amd_iommu_cpumask_group,
399 	&amd_iommu_events_group,
400 	NULL,
401 };
402 
403 static const struct pmu iommu_pmu __initconst = {
404 	.event_init	= perf_iommu_event_init,
405 	.add		= perf_iommu_add,
406 	.del		= perf_iommu_del,
407 	.start		= perf_iommu_start,
408 	.stop		= perf_iommu_stop,
409 	.read		= perf_iommu_read,
410 	.task_ctx_nr	= perf_invalid_context,
411 	.attr_groups	= amd_iommu_attr_groups,
412 	.capabilities	= PERF_PMU_CAP_NO_EXCLUDE,
413 };
414 
415 static __init int init_one_iommu(unsigned int idx)
416 {
417 	struct perf_amd_iommu *perf_iommu;
418 	int ret;
419 
420 	perf_iommu = kzalloc(sizeof(struct perf_amd_iommu), GFP_KERNEL);
421 	if (!perf_iommu)
422 		return -ENOMEM;
423 
424 	raw_spin_lock_init(&perf_iommu->lock);
425 
426 	perf_iommu->pmu          = iommu_pmu;
427 	perf_iommu->iommu        = get_amd_iommu(idx);
428 	perf_iommu->max_banks    = amd_iommu_pc_get_max_banks(idx);
429 	perf_iommu->max_counters = amd_iommu_pc_get_max_counters(idx);
430 
431 	if (!perf_iommu->iommu ||
432 	    !perf_iommu->max_banks ||
433 	    !perf_iommu->max_counters) {
434 		kfree(perf_iommu);
435 		return -EINVAL;
436 	}
437 
438 	snprintf(perf_iommu->name, IOMMU_NAME_SIZE, "amd_iommu_%u", idx);
439 
440 	ret = perf_pmu_register(&perf_iommu->pmu, perf_iommu->name, -1);
441 	if (!ret) {
442 		pr_info("Detected AMD IOMMU #%d (%d banks, %d counters/bank).\n",
443 			idx, perf_iommu->max_banks, perf_iommu->max_counters);
444 		list_add_tail(&perf_iommu->list, &perf_amd_iommu_list);
445 	} else {
446 		pr_warn("Error initializing IOMMU %d.\n", idx);
447 		kfree(perf_iommu);
448 	}
449 	return ret;
450 }
451 
452 static __init int amd_iommu_pc_init(void)
453 {
454 	unsigned int i, cnt = 0;
455 	int ret;
456 
457 	/* Make sure the IOMMU PC resource is available */
458 	if (!amd_iommu_pc_supported())
459 		return -ENODEV;
460 
461 	ret = _init_events_attrs();
462 	if (ret)
463 		return ret;
464 
465 	/*
466 	 * An IOMMU PMU is specific to an IOMMU, and can function independently.
467 	 * So we go through all IOMMUs and ignore the one that fails init
468 	 * unless all IOMMU are failing.
469 	 */
470 	for (i = 0; i < amd_iommu_get_num_iommus(); i++) {
471 		ret = init_one_iommu(i);
472 		if (!ret)
473 			cnt++;
474 	}
475 
476 	if (!cnt) {
477 		kfree(amd_iommu_events_group.attrs);
478 		return -ENODEV;
479 	}
480 
481 	/* Init cpumask attributes to only core 0 */
482 	cpumask_set_cpu(0, &iommu_cpumask);
483 	return 0;
484 }
485 
486 device_initcall(amd_iommu_pc_init);
487