xref: /linux/drivers/perf/cxl_pmu.c (revision ae814200e8393fa504dd246e98fcba8f5493de28)
1 // SPDX-License-Identifier: GPL-2.0-only
2 
3 /*
4  * Copyright(c) 2023 Huawei
5  *
6  * The CXL 3.0 specification includes a standard Performance Monitoring Unit,
7  * called the CXL PMU, or CPMU. In order to allow a high degree of
8  * implementation flexibility the specification provides a wide range of
9  * options all of which are self describing.
10  *
11  * Details in CXL rev 3.0 section 8.2.7 CPMU Register Interface
12  */
13 
14 #include <linux/io-64-nonatomic-lo-hi.h>
15 #include <linux/perf_event.h>
16 #include <linux/bitops.h>
17 #include <linux/device.h>
18 #include <linux/bits.h>
19 #include <linux/list.h>
20 #include <linux/bug.h>
21 #include <linux/pci.h>
22 
23 #include "../cxl/cxlpci.h"
24 #include "../cxl/cxl.h"
25 #include "../cxl/pmu.h"
26 
27 #define CXL_PMU_CAP_REG			0x0
28 #define   CXL_PMU_CAP_NUM_COUNTERS_MSK			GENMASK_ULL(5, 0)
29 #define   CXL_PMU_CAP_COUNTER_WIDTH_MSK			GENMASK_ULL(15, 8)
30 #define   CXL_PMU_CAP_NUM_EVN_CAP_REG_SUP_MSK		GENMASK_ULL(24, 20)
31 #define   CXL_PMU_CAP_FILTERS_SUP_MSK			GENMASK_ULL(39, 32)
32 #define     CXL_PMU_FILTER_HDM				BIT(0)
33 #define     CXL_PMU_FILTER_CHAN_RANK_BANK		BIT(1)
34 #define   CXL_PMU_CAP_MSI_N_MSK				GENMASK_ULL(47, 44)
35 #define   CXL_PMU_CAP_WRITEABLE_WHEN_FROZEN		BIT_ULL(48)
36 #define   CXL_PMU_CAP_FREEZE				BIT_ULL(49)
37 #define   CXL_PMU_CAP_INT				BIT_ULL(50)
38 #define   CXL_PMU_CAP_VERSION_MSK			GENMASK_ULL(63, 60)
39 
40 #define CXL_PMU_OVERFLOW_REG		0x10
41 #define CXL_PMU_FREEZE_REG		0x18
42 #define CXL_PMU_EVENT_CAP_REG(n)	(0x100 + 8 * (n))
43 #define   CXL_PMU_EVENT_CAP_SUPPORTED_EVENTS_MSK	GENMASK_ULL(31, 0)
44 #define   CXL_PMU_EVENT_CAP_GROUP_ID_MSK		GENMASK_ULL(47, 32)
45 #define   CXL_PMU_EVENT_CAP_VENDOR_ID_MSK		GENMASK_ULL(63, 48)
46 
47 #define CXL_PMU_COUNTER_CFG_REG(n)	(0x200 + 8 * (n))
48 #define   CXL_PMU_COUNTER_CFG_TYPE_MSK			GENMASK_ULL(1, 0)
49 #define     CXL_PMU_COUNTER_CFG_TYPE_FREE_RUN		0
50 #define     CXL_PMU_COUNTER_CFG_TYPE_FIXED_FUN		1
51 #define     CXL_PMU_COUNTER_CFG_TYPE_CONFIGURABLE	2
52 #define   CXL_PMU_COUNTER_CFG_ENABLE			BIT_ULL(8)
53 #define   CXL_PMU_COUNTER_CFG_INT_ON_OVRFLW		BIT_ULL(9)
54 #define   CXL_PMU_COUNTER_CFG_FREEZE_ON_OVRFLW		BIT_ULL(10)
55 #define   CXL_PMU_COUNTER_CFG_EDGE			BIT_ULL(11)
56 #define   CXL_PMU_COUNTER_CFG_INVERT			BIT_ULL(12)
57 #define   CXL_PMU_COUNTER_CFG_THRESHOLD_MSK		GENMASK_ULL(23, 16)
58 #define   CXL_PMU_COUNTER_CFG_EVENTS_MSK		GENMASK_ULL(55, 24)
59 #define   CXL_PMU_COUNTER_CFG_EVENT_GRP_ID_IDX_MSK	GENMASK_ULL(63, 59)
60 
61 #define CXL_PMU_FILTER_CFG_REG(n, f)	(0x400 + 4 * ((f) + (n) * 8))
62 #define   CXL_PMU_FILTER_CFG_VALUE_MSK			GENMASK(31, 0)
63 
64 #define CXL_PMU_COUNTER_REG(n)		(0xc00 + 8 * (n))
65 
66 /* CXL rev 3.0 Table 13-5 Events under CXL Vendor ID */
67 #define CXL_PMU_GID_CLOCK_TICKS		0x00
68 #define CXL_PMU_GID_D2H_REQ		0x0010
69 #define CXL_PMU_GID_D2H_RSP		0x0011
70 #define CXL_PMU_GID_H2D_REQ		0x0012
71 #define CXL_PMU_GID_H2D_RSP		0x0013
72 #define CXL_PMU_GID_CACHE_DATA		0x0014
73 #define CXL_PMU_GID_M2S_REQ		0x0020
74 #define CXL_PMU_GID_M2S_RWD		0x0021
75 #define CXL_PMU_GID_M2S_BIRSP		0x0022
76 #define CXL_PMU_GID_S2M_BISNP		0x0023
77 #define CXL_PMU_GID_S2M_NDR		0x0024
78 #define CXL_PMU_GID_S2M_DRS		0x0025
79 #define CXL_PMU_GID_DDR			0x8000
80 #define CXL_PMU_GID_QUEUE_OCC           0x8001
81 #define CXL_PMU_GID_QUEUE_RESID         0x8002
82 #define CXL_PMU_GID_RETRY_EVENTS        0x8003
83 #define CXL_PMU_GID_THROTTLE            0x8004
84 
85 static int cxl_pmu_cpuhp_state_num;
86 
87 struct cxl_pmu_ev_cap {
88 	u16 vid;
89 	u16 gid;
90 	u32 msk;
91 	union {
92 		int counter_idx; /* fixed counters */
93 		int event_idx; /* configurable counters */
94 	};
95 	struct list_head node;
96 };
97 
98 #define CXL_PMU_MAX_COUNTERS 64
99 struct cxl_pmu_info {
100 	struct pmu pmu;
101 	void __iomem *base;
102 	struct perf_event **hw_events;
103 	struct list_head event_caps_configurable;
104 	struct list_head event_caps_fixed;
105 	DECLARE_BITMAP(used_counter_bm, CXL_PMU_MAX_COUNTERS);
106 	DECLARE_BITMAP(conf_counter_bm, CXL_PMU_MAX_COUNTERS);
107 	u16 counter_width;
108 	u8 num_counters;
109 	u8 num_event_capabilities;
110 	int on_cpu;
111 	struct hlist_node node;
112 	bool filter_hdm;
113 	bool filter_crb;
114 	int irq;
115 };
116 
117 #define pmu_to_cxl_pmu_info(_pmu) container_of(_pmu, struct cxl_pmu_info, pmu)
118 
119 /*
120  * All CPMU counters are discoverable via the Event Capabilities Registers.
121  * Each Event Capability register contains a VID / GroupID.
122  * A counter may then count any combination (by summing) of events in
123  * that group which are in the Supported Events Bitmask.
124  * However, there are some complexities to the scheme.
125  *  - Fixed function counters refer to an Event Capabilities register.
126  *    That event capability register is not then used for Configurable
127  *    counters.
128  */
cxl_pmu_parse_caps(struct device * dev,struct cxl_pmu_info * info)129 static int cxl_pmu_parse_caps(struct device *dev, struct cxl_pmu_info *info)
130 {
131 	unsigned long fixed_counter_event_cap_bm = 0;
132 	void __iomem *base = info->base;
133 	bool freeze_for_enable;
134 	u64 val, eval;
135 	int i;
136 
137 	val = readq(base + CXL_PMU_CAP_REG);
138 	freeze_for_enable = FIELD_GET(CXL_PMU_CAP_WRITEABLE_WHEN_FROZEN, val) &&
139 		FIELD_GET(CXL_PMU_CAP_FREEZE, val);
140 	if (!freeze_for_enable) {
141 		dev_err(dev, "Counters not writable while frozen\n");
142 		return -ENODEV;
143 	}
144 
145 	info->num_counters = FIELD_GET(CXL_PMU_CAP_NUM_COUNTERS_MSK, val) + 1;
146 	info->counter_width = FIELD_GET(CXL_PMU_CAP_COUNTER_WIDTH_MSK, val);
147 	info->num_event_capabilities = FIELD_GET(CXL_PMU_CAP_NUM_EVN_CAP_REG_SUP_MSK, val) + 1;
148 
149 	info->filter_hdm = FIELD_GET(CXL_PMU_CAP_FILTERS_SUP_MSK, val) & CXL_PMU_FILTER_HDM;
150 	info->filter_crb = FIELD_GET(CXL_PMU_CAP_FILTERS_SUP_MSK, val) &
151 		CXL_PMU_FILTER_CHAN_RANK_BANK;
152 	if (FIELD_GET(CXL_PMU_CAP_INT, val))
153 		info->irq = FIELD_GET(CXL_PMU_CAP_MSI_N_MSK, val);
154 	else
155 		info->irq = -1;
156 
157 	/* First handle fixed function counters; note if configurable counters found */
158 	for (i = 0; i < info->num_counters; i++) {
159 		struct cxl_pmu_ev_cap *pmu_ev;
160 		u32 events_msk;
161 		u8 group_idx;
162 
163 		val = readq(base + CXL_PMU_COUNTER_CFG_REG(i));
164 
165 		if (FIELD_GET(CXL_PMU_COUNTER_CFG_TYPE_MSK, val) ==
166 			CXL_PMU_COUNTER_CFG_TYPE_CONFIGURABLE) {
167 			set_bit(i, info->conf_counter_bm);
168 		}
169 
170 		if (FIELD_GET(CXL_PMU_COUNTER_CFG_TYPE_MSK, val) !=
171 		    CXL_PMU_COUNTER_CFG_TYPE_FIXED_FUN)
172 			continue;
173 
174 		/* In this case we know which fields are const */
175 		group_idx = FIELD_GET(CXL_PMU_COUNTER_CFG_EVENT_GRP_ID_IDX_MSK, val);
176 		events_msk = FIELD_GET(CXL_PMU_COUNTER_CFG_EVENTS_MSK, val);
177 		eval = readq(base + CXL_PMU_EVENT_CAP_REG(group_idx));
178 		pmu_ev = devm_kzalloc(dev, sizeof(*pmu_ev), GFP_KERNEL);
179 		if (!pmu_ev)
180 			return -ENOMEM;
181 
182 		pmu_ev->vid = FIELD_GET(CXL_PMU_EVENT_CAP_VENDOR_ID_MSK, eval);
183 		pmu_ev->gid = FIELD_GET(CXL_PMU_EVENT_CAP_GROUP_ID_MSK, eval);
184 		/* For a fixed purpose counter use the events mask from the counter CFG */
185 		pmu_ev->msk = events_msk;
186 		pmu_ev->counter_idx = i;
187 		/* This list add is never unwound as all entries deleted on remove */
188 		list_add(&pmu_ev->node, &info->event_caps_fixed);
189 		/*
190 		 * Configurable counters must not use an Event Capability registers that
191 		 * is in use for a Fixed counter
192 		 */
193 		set_bit(group_idx, &fixed_counter_event_cap_bm);
194 	}
195 
196 	if (!bitmap_empty(info->conf_counter_bm, CXL_PMU_MAX_COUNTERS)) {
197 		struct cxl_pmu_ev_cap *pmu_ev;
198 		int j;
199 		/* Walk event capabilities unused by fixed counters */
200 		for_each_clear_bit(j, &fixed_counter_event_cap_bm,
201 				   info->num_event_capabilities) {
202 			pmu_ev = devm_kzalloc(dev, sizeof(*pmu_ev), GFP_KERNEL);
203 			if (!pmu_ev)
204 				return -ENOMEM;
205 
206 			eval = readq(base + CXL_PMU_EVENT_CAP_REG(j));
207 			pmu_ev->vid = FIELD_GET(CXL_PMU_EVENT_CAP_VENDOR_ID_MSK, eval);
208 			pmu_ev->gid = FIELD_GET(CXL_PMU_EVENT_CAP_GROUP_ID_MSK, eval);
209 			pmu_ev->msk = FIELD_GET(CXL_PMU_EVENT_CAP_SUPPORTED_EVENTS_MSK, eval);
210 			pmu_ev->event_idx = j;
211 			list_add(&pmu_ev->node, &info->event_caps_configurable);
212 		}
213 	}
214 
215 	return 0;
216 }
217 
218 #define CXL_PMU_FORMAT_ATTR(_name, _format)\
219 	(&((struct dev_ext_attribute[]) {					\
220 		{								\
221 			.attr = __ATTR(_name, 0444, device_show_string, NULL),	\
222 			.var = (void *)_format					\
223 		}								\
224 		})[0].attr.attr)
225 
226 enum {
227 	cxl_pmu_mask_attr,
228 	cxl_pmu_gid_attr,
229 	cxl_pmu_vid_attr,
230 	cxl_pmu_threshold_attr,
231 	cxl_pmu_invert_attr,
232 	cxl_pmu_edge_attr,
233 	cxl_pmu_hdm_filter_en_attr,
234 	cxl_pmu_hdm_attr,
235 	cxl_pmu_crb_filter_en_attr,
236 	cxl_pmu_crb_attr,
237 };
238 
239 static struct attribute *cxl_pmu_format_attr[] = {
240 	[cxl_pmu_mask_attr] = CXL_PMU_FORMAT_ATTR(mask, "config:0-31"),
241 	[cxl_pmu_gid_attr] = CXL_PMU_FORMAT_ATTR(gid, "config:32-47"),
242 	[cxl_pmu_vid_attr] = CXL_PMU_FORMAT_ATTR(vid, "config:48-63"),
243 	[cxl_pmu_threshold_attr] = CXL_PMU_FORMAT_ATTR(threshold, "config1:0-15"),
244 	[cxl_pmu_invert_attr] = CXL_PMU_FORMAT_ATTR(invert, "config1:16"),
245 	[cxl_pmu_edge_attr] = CXL_PMU_FORMAT_ATTR(edge, "config1:17"),
246 	[cxl_pmu_hdm_filter_en_attr] = CXL_PMU_FORMAT_ATTR(hdm_filter_en, "config1:18"),
247 	[cxl_pmu_hdm_attr] = CXL_PMU_FORMAT_ATTR(hdm, "config2:0-15"),
248 	[cxl_pmu_crb_filter_en_attr] = CXL_PMU_FORMAT_ATTR(crb_filter_en, "config1:19"),
249 	[cxl_pmu_crb_attr] = CXL_PMU_FORMAT_ATTR(crb, "config2:32-63"),
250 	NULL
251 };
252 
253 #define CXL_PMU_ATTR_CONFIG_MASK_MSK		GENMASK_ULL(31, 0)
254 #define CXL_PMU_ATTR_CONFIG_GID_MSK		GENMASK_ULL(47, 32)
255 #define CXL_PMU_ATTR_CONFIG_VID_MSK		GENMASK_ULL(63, 48)
256 #define CXL_PMU_ATTR_CONFIG1_THRESHOLD_MSK	GENMASK_ULL(15, 0)
257 #define CXL_PMU_ATTR_CONFIG1_INVERT_MSK		BIT(16)
258 #define CXL_PMU_ATTR_CONFIG1_EDGE_MSK		BIT(17)
259 #define CXL_PMU_ATTR_CONFIG1_FILTER_EN_MSK	BIT(18)
260 #define CXL_PMU_ATTR_CONFIG1_CRB_FILTER_EN_MSK	BIT(19)
261 #define CXL_PMU_ATTR_CONFIG2_HDM_MSK		GENMASK(15, 0)
262 #define CXL_PMU_ATTR_CONFIG2_CRB_MSK		GENMASK_ULL(63, 32)
263 
cxl_pmu_format_is_visible(struct kobject * kobj,struct attribute * attr,int a)264 static umode_t cxl_pmu_format_is_visible(struct kobject *kobj,
265 					 struct attribute *attr, int a)
266 {
267 	struct device *dev = kobj_to_dev(kobj);
268 	struct cxl_pmu_info *info = dev_get_drvdata(dev);
269 
270 	/*
271 	 * Filter capability at the CPMU level, so hide the attributes if the particular
272 	 * filter is not supported.
273 	 */
274 	if (!info->filter_hdm &&
275 	    (attr == cxl_pmu_format_attr[cxl_pmu_hdm_filter_en_attr] ||
276 	     attr == cxl_pmu_format_attr[cxl_pmu_hdm_attr]))
277 		return 0;
278 
279 	if (!info->filter_crb &&
280 	    (attr == cxl_pmu_format_attr[cxl_pmu_crb_filter_en_attr] ||
281 	     attr == cxl_pmu_format_attr[cxl_pmu_crb_attr]))
282 		return 0;
283 
284 	return attr->mode;
285 }
286 
287 static const struct attribute_group cxl_pmu_format_group = {
288 	.name = "format",
289 	.attrs = cxl_pmu_format_attr,
290 	.is_visible = cxl_pmu_format_is_visible,
291 };
292 
cxl_pmu_config_get_mask(struct perf_event * event)293 static u32 cxl_pmu_config_get_mask(struct perf_event *event)
294 {
295 	return FIELD_GET(CXL_PMU_ATTR_CONFIG_MASK_MSK, event->attr.config);
296 }
297 
cxl_pmu_config_get_gid(struct perf_event * event)298 static u16 cxl_pmu_config_get_gid(struct perf_event *event)
299 {
300 	return FIELD_GET(CXL_PMU_ATTR_CONFIG_GID_MSK, event->attr.config);
301 }
302 
cxl_pmu_config_get_vid(struct perf_event * event)303 static u16 cxl_pmu_config_get_vid(struct perf_event *event)
304 {
305 	return FIELD_GET(CXL_PMU_ATTR_CONFIG_VID_MSK, event->attr.config);
306 }
307 
cxl_pmu_config1_get_threshold(struct perf_event * event)308 static u8 cxl_pmu_config1_get_threshold(struct perf_event *event)
309 {
310 	return FIELD_GET(CXL_PMU_ATTR_CONFIG1_THRESHOLD_MSK, event->attr.config1);
311 }
312 
cxl_pmu_config1_get_invert(struct perf_event * event)313 static bool cxl_pmu_config1_get_invert(struct perf_event *event)
314 {
315 	return FIELD_GET(CXL_PMU_ATTR_CONFIG1_INVERT_MSK, event->attr.config1);
316 }
317 
cxl_pmu_config1_get_edge(struct perf_event * event)318 static bool cxl_pmu_config1_get_edge(struct perf_event *event)
319 {
320 	return FIELD_GET(CXL_PMU_ATTR_CONFIG1_EDGE_MSK, event->attr.config1);
321 }
322 
323 /*
324  * CPMU specification allows for 8 filters, each with a 32 bit value...
325  * So we need to find 8x32bits to store it in.
326  * As the value used for disable is 0xffff_ffff, a separate enable switch
327  * is needed.
328  */
329 
cxl_pmu_config1_hdm_filter_en(struct perf_event * event)330 static bool cxl_pmu_config1_hdm_filter_en(struct perf_event *event)
331 {
332 	return FIELD_GET(CXL_PMU_ATTR_CONFIG1_FILTER_EN_MSK, event->attr.config1);
333 }
334 
cxl_pmu_config2_get_hdm_decoder(struct perf_event * event)335 static u16 cxl_pmu_config2_get_hdm_decoder(struct perf_event *event)
336 {
337 	return FIELD_GET(CXL_PMU_ATTR_CONFIG2_HDM_MSK, event->attr.config2);
338 }
339 
cxl_pmu_config1_crb_filter_en(struct perf_event * event)340 static u16 cxl_pmu_config1_crb_filter_en(struct perf_event *event)
341 {
342 	return FIELD_GET(CXL_PMU_ATTR_CONFIG1_CRB_FILTER_EN_MSK,
343 			 event->attr.config1);
344 }
345 
cxl_pmu_config2_get_crb(struct perf_event * event)346 static u32 cxl_pmu_config2_get_crb(struct perf_event *event)
347 {
348 	return FIELD_GET(CXL_PMU_ATTR_CONFIG2_CRB_MSK, event->attr.config2);
349 }
350 
cxl_pmu_event_sysfs_show(struct device * dev,struct device_attribute * attr,char * buf)351 static ssize_t cxl_pmu_event_sysfs_show(struct device *dev,
352 					struct device_attribute *attr, char *buf)
353 {
354 	struct perf_pmu_events_attr *pmu_attr =
355 		container_of(attr, struct perf_pmu_events_attr, attr);
356 
357 	return sysfs_emit(buf, "config=%#llx\n", pmu_attr->id);
358 }
359 
360 #define CXL_PMU_EVENT_ATTR(_name, _vid, _gid, _msk)			\
361 	PMU_EVENT_ATTR_ID(_name, cxl_pmu_event_sysfs_show,		\
362 			  ((u64)(_vid) << 48) | ((u64)(_gid) << 32) | (u64)(_msk))
363 
364 /* For CXL spec defined events */
365 #define CXL_PMU_EVENT_CXL_ATTR(_name, _gid, _msk)			\
366 	CXL_PMU_EVENT_ATTR(_name, PCI_VENDOR_ID_CXL, _gid, _msk)
367 
368 static struct attribute *cxl_pmu_event_attrs[] = {
369 	CXL_PMU_EVENT_CXL_ATTR(clock_ticks,			CXL_PMU_GID_CLOCK_TICKS, BIT(0)),
370 	/* CXL rev 3.0 Table 3-17 - Device to Host Requests */
371 	CXL_PMU_EVENT_CXL_ATTR(d2h_req_rdcurr,			CXL_PMU_GID_D2H_REQ, BIT(1)),
372 	CXL_PMU_EVENT_CXL_ATTR(d2h_req_rdown,			CXL_PMU_GID_D2H_REQ, BIT(2)),
373 	CXL_PMU_EVENT_CXL_ATTR(d2h_req_rdshared,		CXL_PMU_GID_D2H_REQ, BIT(3)),
374 	CXL_PMU_EVENT_CXL_ATTR(d2h_req_rdany,			CXL_PMU_GID_D2H_REQ, BIT(4)),
375 	CXL_PMU_EVENT_CXL_ATTR(d2h_req_rdownnodata,		CXL_PMU_GID_D2H_REQ, BIT(5)),
376 	CXL_PMU_EVENT_CXL_ATTR(d2h_req_itomwr,			CXL_PMU_GID_D2H_REQ, BIT(6)),
377 	CXL_PMU_EVENT_CXL_ATTR(d2h_req_wrcurr,			CXL_PMU_GID_D2H_REQ, BIT(7)),
378 	CXL_PMU_EVENT_CXL_ATTR(d2h_req_clflush,			CXL_PMU_GID_D2H_REQ, BIT(8)),
379 	CXL_PMU_EVENT_CXL_ATTR(d2h_req_cleanevict,		CXL_PMU_GID_D2H_REQ, BIT(9)),
380 	CXL_PMU_EVENT_CXL_ATTR(d2h_req_dirtyevict,		CXL_PMU_GID_D2H_REQ, BIT(10)),
381 	CXL_PMU_EVENT_CXL_ATTR(d2h_req_cleanevictnodata,	CXL_PMU_GID_D2H_REQ, BIT(11)),
382 	CXL_PMU_EVENT_CXL_ATTR(d2h_req_wowrinv,			CXL_PMU_GID_D2H_REQ, BIT(12)),
383 	CXL_PMU_EVENT_CXL_ATTR(d2h_req_wowrinvf,		CXL_PMU_GID_D2H_REQ, BIT(13)),
384 	CXL_PMU_EVENT_CXL_ATTR(d2h_req_wrinv,			CXL_PMU_GID_D2H_REQ, BIT(14)),
385 	CXL_PMU_EVENT_CXL_ATTR(d2h_req_cacheflushed,		CXL_PMU_GID_D2H_REQ, BIT(16)),
386 	/* CXL rev 3.0 Table 3-20 - D2H Response Encodings */
387 	CXL_PMU_EVENT_CXL_ATTR(d2h_rsp_rspihiti,		CXL_PMU_GID_D2H_RSP, BIT(4)),
388 	CXL_PMU_EVENT_CXL_ATTR(d2h_rsp_rspvhitv,		CXL_PMU_GID_D2H_RSP, BIT(6)),
389 	CXL_PMU_EVENT_CXL_ATTR(d2h_rsp_rspihitse,		CXL_PMU_GID_D2H_RSP, BIT(5)),
390 	CXL_PMU_EVENT_CXL_ATTR(d2h_rsp_rspshitse,		CXL_PMU_GID_D2H_RSP, BIT(1)),
391 	CXL_PMU_EVENT_CXL_ATTR(d2h_rsp_rspsfwdm,		CXL_PMU_GID_D2H_RSP, BIT(7)),
392 	CXL_PMU_EVENT_CXL_ATTR(d2h_rsp_rspifwdm,		CXL_PMU_GID_D2H_RSP, BIT(15)),
393 	CXL_PMU_EVENT_CXL_ATTR(d2h_rsp_rspvfwdv,		CXL_PMU_GID_D2H_RSP, BIT(22)),
394 	/* CXL rev 3.0 Table 3-21 - CXL.cache - Mapping of H2D Requests to D2H Responses */
395 	CXL_PMU_EVENT_CXL_ATTR(h2d_req_snpdata,			CXL_PMU_GID_H2D_REQ, BIT(1)),
396 	CXL_PMU_EVENT_CXL_ATTR(h2d_req_snpinv,			CXL_PMU_GID_H2D_REQ, BIT(2)),
397 	CXL_PMU_EVENT_CXL_ATTR(h2d_req_snpcur,			CXL_PMU_GID_H2D_REQ, BIT(3)),
398 	/* CXL rev 3.0 Table 3-22 - H2D Response Opcode Encodings */
399 	CXL_PMU_EVENT_CXL_ATTR(h2d_rsp_writepull,		CXL_PMU_GID_H2D_RSP, BIT(1)),
400 	CXL_PMU_EVENT_CXL_ATTR(h2d_rsp_go,			CXL_PMU_GID_H2D_RSP, BIT(4)),
401 	CXL_PMU_EVENT_CXL_ATTR(h2d_rsp_gowritepull,		CXL_PMU_GID_H2D_RSP, BIT(5)),
402 	CXL_PMU_EVENT_CXL_ATTR(h2d_rsp_extcmp,			CXL_PMU_GID_H2D_RSP, BIT(6)),
403 	CXL_PMU_EVENT_CXL_ATTR(h2d_rsp_gowritepulldrop,		CXL_PMU_GID_H2D_RSP, BIT(8)),
404 	CXL_PMU_EVENT_CXL_ATTR(h2d_rsp_fastgowritepull,		CXL_PMU_GID_H2D_RSP, BIT(13)),
405 	CXL_PMU_EVENT_CXL_ATTR(h2d_rsp_goerrwritepull,		CXL_PMU_GID_H2D_RSP, BIT(15)),
406 	/* CXL rev 3.0 Table 13-5 directly lists these */
407 	CXL_PMU_EVENT_CXL_ATTR(cachedata_d2h_data,		CXL_PMU_GID_CACHE_DATA, BIT(0)),
408 	CXL_PMU_EVENT_CXL_ATTR(cachedata_h2d_data,		CXL_PMU_GID_CACHE_DATA, BIT(1)),
409 	/* CXL rev 3.1 Table 3-35 M2S Req Memory Opcodes */
410 	CXL_PMU_EVENT_CXL_ATTR(m2s_req_meminv,			CXL_PMU_GID_M2S_REQ, BIT(0)),
411 	CXL_PMU_EVENT_CXL_ATTR(m2s_req_memrd,			CXL_PMU_GID_M2S_REQ, BIT(1)),
412 	CXL_PMU_EVENT_CXL_ATTR(m2s_req_memrddata,		CXL_PMU_GID_M2S_REQ, BIT(2)),
413 	CXL_PMU_EVENT_CXL_ATTR(m2s_req_memrdfwd,		CXL_PMU_GID_M2S_REQ, BIT(3)),
414 	CXL_PMU_EVENT_CXL_ATTR(m2s_req_memwrfwd,		CXL_PMU_GID_M2S_REQ, BIT(4)),
415 	CXL_PMU_EVENT_CXL_ATTR(m2s_req_memrdtee,		CXL_PMU_GID_M2S_REQ, BIT(5)),
416 	CXL_PMU_EVENT_CXL_ATTR(m2s_req_memrddatatee,		CXL_PMU_GID_M2S_REQ, BIT(6)),
417 	CXL_PMU_EVENT_CXL_ATTR(m2s_req_meminvtee,               CXL_PMU_GID_M2S_REQ, BIT(7)),
418 	CXL_PMU_EVENT_CXL_ATTR(m2s_req_memspecrd,		CXL_PMU_GID_M2S_REQ, BIT(8)),
419 	CXL_PMU_EVENT_CXL_ATTR(m2s_req_meminvnt,		CXL_PMU_GID_M2S_REQ, BIT(9)),
420 	CXL_PMU_EVENT_CXL_ATTR(m2s_req_memcleanevict,		CXL_PMU_GID_M2S_REQ, BIT(10)),
421 	CXL_PMU_EVENT_CXL_ATTR(m2s_req_meminvptee,		CXL_PMU_GID_M2S_REQ, BIT(11)),
422 	CXL_PMU_EVENT_CXL_ATTR(m2s_req_memspecrdtee,		CXL_PMU_GID_M2S_REQ, BIT(12)),
423 	CXL_PMU_EVENT_CXL_ATTR(m2s_req_teupdate,		CXL_PMU_GID_M2S_REQ, BIT(13)),
424 	CXL_PMU_EVENT_CXL_ATTR(m2s_req_memclnevcttee,		CXL_PMU_GID_M2S_REQ, BIT(14)),
425 	CXL_PMU_EVENT_CXL_ATTR(m2s_req_memclnevctu,		CXL_PMU_GID_M2S_REQ, BIT(15)),
426 	/* CXL rev 3.0 Table 3-35 M2S RwD Memory Opcodes */
427 	CXL_PMU_EVENT_CXL_ATTR(m2s_rwd_memwr,			CXL_PMU_GID_M2S_RWD, BIT(1)),
428 	CXL_PMU_EVENT_CXL_ATTR(m2s_rwd_memwrptl,		CXL_PMU_GID_M2S_RWD, BIT(2)),
429 	CXL_PMU_EVENT_CXL_ATTR(m2s_rwd_biconflict,		CXL_PMU_GID_M2S_RWD, BIT(4)),
430 	CXL_PMU_EVENT_CXL_ATTR(m2s_rwd_memrdfill,		CXL_PMU_GID_M2S_RWD, BIT(5)),
431 	CXL_PMU_EVENT_CXL_ATTR(m2s_rwd_memwrtee,		CXL_PMU_GID_M2S_RWD, BIT(9)),
432 	CXL_PMU_EVENT_CXL_ATTR(m2s_rwd_memwrptltee,		CXL_PMU_GID_M2S_RWD, BIT(10)),
433 	CXL_PMU_EVENT_CXL_ATTR(m2s_rwd_memrdfilltee,		CXL_PMU_GID_M2S_RWD, BIT(13)),
434 	/* CXL rev 3.0 Table 3-38 M2S BIRsp Memory Opcodes */
435 	CXL_PMU_EVENT_CXL_ATTR(m2s_birsp_i,			CXL_PMU_GID_M2S_BIRSP, BIT(0)),
436 	CXL_PMU_EVENT_CXL_ATTR(m2s_birsp_s,			CXL_PMU_GID_M2S_BIRSP, BIT(1)),
437 	CXL_PMU_EVENT_CXL_ATTR(m2s_birsp_e,			CXL_PMU_GID_M2S_BIRSP, BIT(2)),
438 	CXL_PMU_EVENT_CXL_ATTR(m2s_birsp_iblk,			CXL_PMU_GID_M2S_BIRSP, BIT(4)),
439 	CXL_PMU_EVENT_CXL_ATTR(m2s_birsp_sblk,			CXL_PMU_GID_M2S_BIRSP, BIT(5)),
440 	CXL_PMU_EVENT_CXL_ATTR(m2s_birsp_eblk,			CXL_PMU_GID_M2S_BIRSP, BIT(6)),
441 	/* CXL rev 3.0 Table 3-40 S2M BISnp Opcodes */
442 	CXL_PMU_EVENT_CXL_ATTR(s2m_bisnp_cur,			CXL_PMU_GID_S2M_BISNP, BIT(0)),
443 	CXL_PMU_EVENT_CXL_ATTR(s2m_bisnp_data,			CXL_PMU_GID_S2M_BISNP, BIT(1)),
444 	CXL_PMU_EVENT_CXL_ATTR(s2m_bisnp_inv,			CXL_PMU_GID_S2M_BISNP, BIT(2)),
445 	CXL_PMU_EVENT_CXL_ATTR(s2m_bisnp_curblk,		CXL_PMU_GID_S2M_BISNP, BIT(4)),
446 	CXL_PMU_EVENT_CXL_ATTR(s2m_bisnp_datblk,		CXL_PMU_GID_S2M_BISNP, BIT(5)),
447 	CXL_PMU_EVENT_CXL_ATTR(s2m_bisnp_invblk,		CXL_PMU_GID_S2M_BISNP, BIT(6)),
448 	CXL_PMU_EVENT_CXL_ATTR(s2m_bisnp_curtee,		CXL_PMU_GID_S2M_BISNP, BIT(8)),
449 	CXL_PMU_EVENT_CXL_ATTR(s2m_bisnp_datatee,		CXL_PMU_GID_S2M_BISNP, BIT(9)),
450 	CXL_PMU_EVENT_CXL_ATTR(s2m_bisnp_invtee,		CXL_PMU_GID_S2M_BISNP, BIT(10)),
451 	CXL_PMU_EVENT_CXL_ATTR(s2m_bisnp_curblktee,		CXL_PMU_GID_S2M_BISNP, BIT(12)),
452 	CXL_PMU_EVENT_CXL_ATTR(s2m_bisnp_datablktee,		CXL_PMU_GID_S2M_BISNP, BIT(13)),
453 	CXL_PMU_EVENT_CXL_ATTR(s2m_bisnp_invblktee,		CXL_PMU_GID_S2M_BISNP, BIT(14)),
454 	/* CXL rev 3.1 Table 3-50 S2M NDR Opcodes */
455 	CXL_PMU_EVENT_CXL_ATTR(s2m_ndr_cmp,			CXL_PMU_GID_S2M_NDR, BIT(0)),
456 	CXL_PMU_EVENT_CXL_ATTR(s2m_ndr_cmps,			CXL_PMU_GID_S2M_NDR, BIT(1)),
457 	CXL_PMU_EVENT_CXL_ATTR(s2m_ndr_cmpe,			CXL_PMU_GID_S2M_NDR, BIT(2)),
458 	CXL_PMU_EVENT_CXL_ATTR(s2m_ndr_cmpm,			CXL_PMU_GID_S2M_NDR, BIT(3)),
459 	CXL_PMU_EVENT_CXL_ATTR(s2m_ndr_biconflictack,		CXL_PMU_GID_S2M_NDR, BIT(4)),
460 	CXL_PMU_EVENT_CXL_ATTR(s2m_ndr_cmptee,			CXL_PMU_GID_S2M_NDR, BIT(5)),
461 	CXL_PMU_EVENT_CXL_ATTR(s2m_ndr_cmptee_s,		CXL_PMU_GID_S2M_NDR, BIT(6)),
462 	CXL_PMU_EVENT_CXL_ATTR(s2m_ndr_cmptee_e,		CXL_PMU_GID_S2M_NDR, BIT(7)),
463 	/* CXL rev 3.0 Table 3-46 S2M DRS opcodes */
464 	CXL_PMU_EVENT_CXL_ATTR(s2m_drs_memdata,			CXL_PMU_GID_S2M_DRS, BIT(0)),
465 	CXL_PMU_EVENT_CXL_ATTR(s2m_drs_memdatanxm,		CXL_PMU_GID_S2M_DRS, BIT(1)),
466 	CXL_PMU_EVENT_CXL_ATTR(s2m_drs_memdatatee,		CXL_PMU_GID_S2M_DRS, BIT(2)),
467 	/* CXL rev 3.0 Table 13-5 directly lists these */
468 	CXL_PMU_EVENT_CXL_ATTR(ddr_act,				CXL_PMU_GID_DDR, BIT(0)),
469 	CXL_PMU_EVENT_CXL_ATTR(ddr_pre,				CXL_PMU_GID_DDR, BIT(1)),
470 	CXL_PMU_EVENT_CXL_ATTR(ddr_casrd,			CXL_PMU_GID_DDR, BIT(2)),
471 	CXL_PMU_EVENT_CXL_ATTR(ddr_caswr,			CXL_PMU_GID_DDR, BIT(3)),
472 	CXL_PMU_EVENT_CXL_ATTR(ddr_refresh,			CXL_PMU_GID_DDR, BIT(4)),
473 	CXL_PMU_EVENT_CXL_ATTR(ddr_selfrefreshent,		CXL_PMU_GID_DDR, BIT(5)),
474 	CXL_PMU_EVENT_CXL_ATTR(ddr_rfm,				CXL_PMU_GID_DDR, BIT(6)),
475 	/* CXL 4.0 Table 13-5 DDR add-on events opcodes */
476 	CXL_PMU_EVENT_CXL_ATTR(ddr_cas_rd_ap,			CXL_PMU_GID_DDR, BIT(7)),
477 	CXL_PMU_EVENT_CXL_ATTR(ddr_cas_wr_ap,			CXL_PMU_GID_DDR, BIT(8)),
478 	CXL_PMU_EVENT_CXL_ATTR(ddr_refresh_all_banks,		CXL_PMU_GID_DDR, BIT(9)),
479 	CXL_PMU_EVENT_CXL_ATTR(ddr_refresh_same_bank,		CXL_PMU_GID_DDR, BIT(10)),
480 	CXL_PMU_EVENT_CXL_ATTR(ddr_pwrdn_entry,			CXL_PMU_GID_DDR, BIT(11)),
481 	CXL_PMU_EVENT_CXL_ATTR(ddr_pwrdn_exit,			CXL_PMU_GID_DDR, BIT(12)),
482 	CXL_PMU_EVENT_CXL_ATTR(ddr_rd_wr_ddr_bus_switching,	CXL_PMU_GID_DDR, BIT(13)),
483 	CXL_PMU_EVENT_CXL_ATTR(ddr_incoming_rd_req,		CXL_PMU_GID_DDR, BIT(14)),
484 	CXL_PMU_EVENT_CXL_ATTR(ddr_incoming_wr_req,		CXL_PMU_GID_DDR, BIT(15)),
485 	/* CXL 4.0 Table 13-5 QUEUE OCCUPANCY events opcodes */
486 	CXL_PMU_EVENT_CXL_ATTR(rd_queue_occ,			CXL_PMU_GID_QUEUE_OCC, BIT(0)),
487 	CXL_PMU_EVENT_CXL_ATTR(wr_queue_occ,			CXL_PMU_GID_QUEUE_OCC, BIT(1)),
488 	CXL_PMU_EVENT_CXL_ATTR(rd_wr_merged_queue_occ,		CXL_PMU_GID_QUEUE_OCC, BIT(2)),
489 	CXL_PMU_EVENT_CXL_ATTR(pwrdn_event,			CXL_PMU_GID_QUEUE_OCC, BIT(3)),
490 	/* CXL 4.0 Table 13-5 QUEUE RESIDENCY events opcodes */
491 	CXL_PMU_EVENT_CXL_ATTR(mc_rd_resid_cnt,			CXL_PMU_GID_QUEUE_RESID, BIT(0)),
492 	CXL_PMU_EVENT_CXL_ATTR(mc_wr_resid_cnt,			CXL_PMU_GID_QUEUE_RESID, BIT(1)),
493 	/* CXL 4.0 Table 13-5 RETRY events opcodes */
494 	CXL_PMU_EVENT_CXL_ATTR(retry_event_trig_by_rd_crc,      CXL_PMU_GID_RETRY_EVENTS, BIT(0)),
495 	CXL_PMU_EVENT_CXL_ATTR(retry_event_trig_by_wr_crc,      CXL_PMU_GID_RETRY_EVENTS, BIT(1)),
496 	CXL_PMU_EVENT_CXL_ATTR(retry_event_trig_by_ca_parity,   CXL_PMU_GID_RETRY_EVENTS, BIT(2)),
497 	CXL_PMU_EVENT_CXL_ATTR(retry_event_trig_by_ecc,         CXL_PMU_GID_RETRY_EVENTS, BIT(3)),
498 	/* CXL 4.0 Table 13-5 THROTTLE events opcodes */
499 	CXL_PMU_EVENT_CXL_ATTR(thermal_throttle_event,		CXL_PMU_GID_THROTTLE, BIT(0)),
500 	CXL_PMU_EVENT_CXL_ATTR(power_throttle_event,		CXL_PMU_GID_THROTTLE, BIT(1)),
501 	NULL
502 };
503 
cxl_pmu_find_fixed_counter_ev_cap(struct cxl_pmu_info * info,int vid,int gid,int msk)504 static struct cxl_pmu_ev_cap *cxl_pmu_find_fixed_counter_ev_cap(struct cxl_pmu_info *info,
505 								int vid, int gid, int msk)
506 {
507 	struct cxl_pmu_ev_cap *pmu_ev;
508 
509 	list_for_each_entry(pmu_ev, &info->event_caps_fixed, node) {
510 		if (vid != pmu_ev->vid || gid != pmu_ev->gid)
511 			continue;
512 
513 		/* Precise match for fixed counter */
514 		if (msk == pmu_ev->msk)
515 			return pmu_ev;
516 	}
517 
518 	return ERR_PTR(-EINVAL);
519 }
520 
cxl_pmu_find_config_counter_ev_cap(struct cxl_pmu_info * info,int vid,int gid,int msk)521 static struct cxl_pmu_ev_cap *cxl_pmu_find_config_counter_ev_cap(struct cxl_pmu_info *info,
522 								 int vid, int gid, int msk)
523 {
524 	struct cxl_pmu_ev_cap *pmu_ev;
525 
526 	list_for_each_entry(pmu_ev, &info->event_caps_configurable, node) {
527 		if (vid != pmu_ev->vid || gid != pmu_ev->gid)
528 			continue;
529 
530 		/* Request mask must be subset of supported */
531 		if (msk & ~pmu_ev->msk)
532 			continue;
533 
534 		return pmu_ev;
535 	}
536 
537 	return ERR_PTR(-EINVAL);
538 }
539 
cxl_pmu_event_is_visible(struct kobject * kobj,struct attribute * attr,int a)540 static umode_t cxl_pmu_event_is_visible(struct kobject *kobj, struct attribute *attr, int a)
541 {
542 	struct device_attribute *dev_attr = container_of(attr, struct device_attribute, attr);
543 	struct perf_pmu_events_attr *pmu_attr =
544 		container_of(dev_attr, struct perf_pmu_events_attr, attr);
545 	struct device *dev = kobj_to_dev(kobj);
546 	struct cxl_pmu_info *info = dev_get_drvdata(dev);
547 	int vid = FIELD_GET(CXL_PMU_ATTR_CONFIG_VID_MSK, pmu_attr->id);
548 	int gid = FIELD_GET(CXL_PMU_ATTR_CONFIG_GID_MSK, pmu_attr->id);
549 	int msk = FIELD_GET(CXL_PMU_ATTR_CONFIG_MASK_MSK, pmu_attr->id);
550 
551 	if (!IS_ERR(cxl_pmu_find_fixed_counter_ev_cap(info, vid, gid, msk)))
552 		return attr->mode;
553 
554 	if (!IS_ERR(cxl_pmu_find_config_counter_ev_cap(info, vid, gid, msk)))
555 		return attr->mode;
556 
557 	return 0;
558 }
559 
560 static const struct attribute_group cxl_pmu_events = {
561 	.name = "events",
562 	.attrs = cxl_pmu_event_attrs,
563 	.is_visible = cxl_pmu_event_is_visible,
564 };
565 
cpumask_show(struct device * dev,struct device_attribute * attr,char * buf)566 static ssize_t cpumask_show(struct device *dev, struct device_attribute *attr,
567 			    char *buf)
568 {
569 	struct cxl_pmu_info *info = dev_get_drvdata(dev);
570 
571 	return sysfs_emit(buf, "%*pbl\n", cpumask_pr_args(cpumask_of(info->on_cpu)));
572 }
573 static DEVICE_ATTR_RO(cpumask);
574 
575 static struct attribute *cxl_pmu_cpumask_attrs[] = {
576 	&dev_attr_cpumask.attr,
577 	NULL
578 };
579 
580 static const struct attribute_group cxl_pmu_cpumask_group = {
581 	.attrs = cxl_pmu_cpumask_attrs,
582 };
583 
584 static const struct attribute_group *cxl_pmu_attr_groups[] = {
585 	&cxl_pmu_events,
586 	&cxl_pmu_format_group,
587 	&cxl_pmu_cpumask_group,
588 	NULL
589 };
590 
591 /* If counter_idx == NULL, don't try to allocate a counter. */
cxl_pmu_get_event_idx(struct perf_event * event,int * counter_idx,int * event_idx)592 static int cxl_pmu_get_event_idx(struct perf_event *event, int *counter_idx,
593 				 int *event_idx)
594 {
595 	struct cxl_pmu_info *info = pmu_to_cxl_pmu_info(event->pmu);
596 	DECLARE_BITMAP(configurable_and_free, CXL_PMU_MAX_COUNTERS);
597 	struct cxl_pmu_ev_cap *pmu_ev;
598 	u32 mask;
599 	u16 gid, vid;
600 	int i;
601 
602 	vid = cxl_pmu_config_get_vid(event);
603 	gid = cxl_pmu_config_get_gid(event);
604 	mask = cxl_pmu_config_get_mask(event);
605 
606 	pmu_ev = cxl_pmu_find_fixed_counter_ev_cap(info, vid, gid, mask);
607 	if (!IS_ERR(pmu_ev)) {
608 		if (!counter_idx)
609 			return 0;
610 		if (!test_bit(pmu_ev->counter_idx, info->used_counter_bm)) {
611 			*counter_idx = pmu_ev->counter_idx;
612 			return 0;
613 		}
614 		/* Fixed counter is in use, but maybe a configurable one? */
615 	}
616 
617 	pmu_ev = cxl_pmu_find_config_counter_ev_cap(info, vid, gid, mask);
618 	if (!IS_ERR(pmu_ev)) {
619 		if (!counter_idx)
620 			return 0;
621 
622 		bitmap_andnot(configurable_and_free, info->conf_counter_bm,
623 			info->used_counter_bm, CXL_PMU_MAX_COUNTERS);
624 
625 		i = find_first_bit(configurable_and_free, CXL_PMU_MAX_COUNTERS);
626 		if (i == CXL_PMU_MAX_COUNTERS)
627 			return -EINVAL;
628 
629 		*counter_idx = i;
630 		return 0;
631 	}
632 
633 	return -EINVAL;
634 }
635 
cxl_pmu_event_init(struct perf_event * event)636 static int cxl_pmu_event_init(struct perf_event *event)
637 {
638 	struct cxl_pmu_info *info = pmu_to_cxl_pmu_info(event->pmu);
639 	int rc;
640 
641 	/* Top level type sanity check - is this a Hardware Event being requested */
642 	if (event->attr.type != event->pmu->type)
643 		return -ENOENT;
644 
645 	if (is_sampling_event(event) || event->attach_state & PERF_ATTACH_TASK)
646 		return -EOPNOTSUPP;
647 	/* TODO: Validation of any filter */
648 
649 	if (cxl_pmu_config1_crb_filter_en(event)) {
650 		if (!info->filter_crb)
651 			return -EINVAL;
652 		/* event group IDs are scoped by the CXL vendor ID */
653 		if (cxl_pmu_config_get_vid(event) != PCI_VENDOR_ID_CXL)
654 			return -EINVAL;
655 
656 		/*
657 		 * CRB filtering (Filter ID 1) is only valid for the DDR
658 		 * Interface, Queue Occupancy, Queue Residency and Retry
659 		 * event groups (CXL 4.0 Table 13-5).
660 		 */
661 		switch (cxl_pmu_config_get_gid(event)) {
662 		case CXL_PMU_GID_DDR:
663 		case CXL_PMU_GID_QUEUE_OCC:
664 		case CXL_PMU_GID_QUEUE_RESID:
665 		case CXL_PMU_GID_RETRY_EVENTS:
666 			break;
667 		default:
668 			return -EINVAL;
669 		}
670 
671 		/*
672 		 * Filtering while counting multiple events is
673 		 * undefined behavior.
674 		 */
675 		if (hweight32(cxl_pmu_config_get_mask(event)) > 1)
676 			return -EINVAL;
677 	}
678 
679 	/*
680 	 * Verify that it is possible to count what was requested. Either must
681 	 * be a fixed counter that is a precise match or a configurable counter
682 	 * where this is a subset.
683 	 */
684 	rc = cxl_pmu_get_event_idx(event, NULL, NULL);
685 	if (rc < 0)
686 		return rc;
687 
688 	event->cpu = info->on_cpu;
689 
690 	return 0;
691 }
692 
cxl_pmu_enable(struct pmu * pmu)693 static void cxl_pmu_enable(struct pmu *pmu)
694 {
695 	struct cxl_pmu_info *info = pmu_to_cxl_pmu_info(pmu);
696 	void __iomem *base = info->base;
697 
698 	/* Can assume frozen at this stage */
699 	writeq(0, base + CXL_PMU_FREEZE_REG);
700 }
701 
cxl_pmu_disable(struct pmu * pmu)702 static void cxl_pmu_disable(struct pmu *pmu)
703 {
704 	struct cxl_pmu_info *info = pmu_to_cxl_pmu_info(pmu);
705 	void __iomem *base = info->base;
706 
707 	/*
708 	 * Whilst bits above number of counters are RsvdZ
709 	 * they are unlikely to be repurposed given
710 	 * number of counters is allowed to be 64 leaving
711 	 * no reserved bits.  Hence this is only slightly
712 	 * naughty.
713 	 */
714 	writeq(GENMASK_ULL(63, 0), base + CXL_PMU_FREEZE_REG);
715 }
716 
cxl_pmu_event_start(struct perf_event * event,int flags)717 static void cxl_pmu_event_start(struct perf_event *event, int flags)
718 {
719 	struct cxl_pmu_info *info = pmu_to_cxl_pmu_info(event->pmu);
720 	struct hw_perf_event *hwc = &event->hw;
721 	void __iomem *base = info->base;
722 	u64 cfg;
723 
724 	/*
725 	 * All paths to here should either set these flags directly or
726 	 * call cxl_pmu_event_stop() which will ensure the correct state.
727 	 */
728 	if (WARN_ON_ONCE(!(hwc->state & PERF_HES_STOPPED)))
729 		return;
730 
731 	WARN_ON_ONCE(!(hwc->state & PERF_HES_UPTODATE));
732 	hwc->state = 0;
733 
734 	/*
735 	 * Filter ID=0: HDM decoder filter
736 	 * Filter ID=1: Channel/Rank/Bank (CRB) filter
737 	 */
738 	if (info->filter_hdm) {
739 		if (cxl_pmu_config1_hdm_filter_en(event))
740 			cfg = cxl_pmu_config2_get_hdm_decoder(event);
741 		else
742 			cfg = GENMASK(31, 0); /* No filtering if 0xFFFF_FFFF */
743 		writel(cfg, base + CXL_PMU_FILTER_CFG_REG(hwc->idx, 0));
744 	}
745 
746 	if (info->filter_crb) {
747 		if (cxl_pmu_config1_crb_filter_en(event))
748 			cfg = cxl_pmu_config2_get_crb(event);
749 		else
750 			cfg = GENMASK(31, 0); /* no filtering if 0xFFFF_FFFF */
751 		writel(cfg, base + CXL_PMU_FILTER_CFG_REG(hwc->idx, 1));
752 	}
753 
754 	cfg = readq(base + CXL_PMU_COUNTER_CFG_REG(hwc->idx));
755 	cfg |= FIELD_PREP(CXL_PMU_COUNTER_CFG_INT_ON_OVRFLW, 1);
756 	cfg |= FIELD_PREP(CXL_PMU_COUNTER_CFG_FREEZE_ON_OVRFLW, 1);
757 	cfg |= FIELD_PREP(CXL_PMU_COUNTER_CFG_ENABLE, 1);
758 	cfg |= FIELD_PREP(CXL_PMU_COUNTER_CFG_EDGE,
759 			  cxl_pmu_config1_get_edge(event) ? 1 : 0);
760 	cfg |= FIELD_PREP(CXL_PMU_COUNTER_CFG_INVERT,
761 			  cxl_pmu_config1_get_invert(event) ? 1 : 0);
762 
763 	/* Fixed purpose counters have next two fields RO */
764 	if (test_bit(hwc->idx, info->conf_counter_bm)) {
765 		cfg |= FIELD_PREP(CXL_PMU_COUNTER_CFG_EVENT_GRP_ID_IDX_MSK,
766 				  hwc->event_base);
767 		cfg |= FIELD_PREP(CXL_PMU_COUNTER_CFG_EVENTS_MSK,
768 				  cxl_pmu_config_get_mask(event));
769 	}
770 	cfg &= ~CXL_PMU_COUNTER_CFG_THRESHOLD_MSK;
771 	/*
772 	 * For events that generate only 1 count per clock the CXL 3.0 spec
773 	 * states the threshold shall be set to 1 but if set to 0 it will
774 	 * count the raw value anwyay?
775 	 * There is no definition of what events will count multiple per cycle
776 	 * and hence to which non 1 values of threshold can apply.
777 	 * (CXL 3.0 8.2.7.2.1 Counter Configuration - threshold field definition)
778 	 */
779 	cfg |= FIELD_PREP(CXL_PMU_COUNTER_CFG_THRESHOLD_MSK,
780 			  cxl_pmu_config1_get_threshold(event));
781 	writeq(cfg, base + CXL_PMU_COUNTER_CFG_REG(hwc->idx));
782 
783 	local64_set(&hwc->prev_count, 0);
784 	writeq(0, base + CXL_PMU_COUNTER_REG(hwc->idx));
785 
786 	perf_event_update_userpage(event);
787 }
788 
cxl_pmu_read_counter(struct perf_event * event)789 static u64 cxl_pmu_read_counter(struct perf_event *event)
790 {
791 	struct cxl_pmu_info *info = pmu_to_cxl_pmu_info(event->pmu);
792 	void __iomem *base = info->base;
793 
794 	return readq(base + CXL_PMU_COUNTER_REG(event->hw.idx));
795 }
796 
__cxl_pmu_read(struct perf_event * event,bool overflow)797 static void __cxl_pmu_read(struct perf_event *event, bool overflow)
798 {
799 	struct cxl_pmu_info *info = pmu_to_cxl_pmu_info(event->pmu);
800 	struct hw_perf_event *hwc = &event->hw;
801 	u64 new_cnt, prev_cnt, delta;
802 
803 	do {
804 		prev_cnt = local64_read(&hwc->prev_count);
805 		new_cnt = cxl_pmu_read_counter(event);
806 	} while (local64_cmpxchg(&hwc->prev_count, prev_cnt, new_cnt) != prev_cnt);
807 
808 	/*
809 	 * If we know an overflow occur then take that into account.
810 	 * Note counter is not reset as that would lose events
811 	 */
812 	delta = (new_cnt - prev_cnt) & GENMASK_ULL(info->counter_width - 1, 0);
813 	if (overflow && delta < GENMASK_ULL(info->counter_width - 1, 0))
814 		delta += (1UL << info->counter_width);
815 
816 	local64_add(delta, &event->count);
817 }
818 
cxl_pmu_read(struct perf_event * event)819 static void cxl_pmu_read(struct perf_event *event)
820 {
821 	__cxl_pmu_read(event, false);
822 }
823 
cxl_pmu_event_stop(struct perf_event * event,int flags)824 static void cxl_pmu_event_stop(struct perf_event *event, int flags)
825 {
826 	struct cxl_pmu_info *info = pmu_to_cxl_pmu_info(event->pmu);
827 	void __iomem *base = info->base;
828 	struct hw_perf_event *hwc = &event->hw;
829 	u64 cfg;
830 
831 	cxl_pmu_read(event);
832 	WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
833 	hwc->state |= PERF_HES_STOPPED;
834 
835 	cfg = readq(base + CXL_PMU_COUNTER_CFG_REG(hwc->idx));
836 	cfg &= ~(FIELD_PREP(CXL_PMU_COUNTER_CFG_INT_ON_OVRFLW, 1) |
837 		 FIELD_PREP(CXL_PMU_COUNTER_CFG_ENABLE, 1));
838 	writeq(cfg, base + CXL_PMU_COUNTER_CFG_REG(hwc->idx));
839 
840 	hwc->state |= PERF_HES_UPTODATE;
841 }
842 
cxl_pmu_event_add(struct perf_event * event,int flags)843 static int cxl_pmu_event_add(struct perf_event *event, int flags)
844 {
845 	struct cxl_pmu_info *info = pmu_to_cxl_pmu_info(event->pmu);
846 	struct hw_perf_event *hwc = &event->hw;
847 	int idx, rc;
848 	int event_idx = 0;
849 
850 	hwc->state = PERF_HES_STOPPED | PERF_HES_UPTODATE;
851 
852 	rc = cxl_pmu_get_event_idx(event, &idx, &event_idx);
853 	if (rc < 0)
854 		return rc;
855 
856 	hwc->idx = idx;
857 
858 	/* Only set for configurable counters */
859 	hwc->event_base = event_idx;
860 	info->hw_events[idx] = event;
861 	set_bit(idx, info->used_counter_bm);
862 
863 	if (flags & PERF_EF_START)
864 		cxl_pmu_event_start(event, PERF_EF_RELOAD);
865 
866 	return 0;
867 }
868 
cxl_pmu_event_del(struct perf_event * event,int flags)869 static void cxl_pmu_event_del(struct perf_event *event, int flags)
870 {
871 	struct cxl_pmu_info *info = pmu_to_cxl_pmu_info(event->pmu);
872 	struct hw_perf_event *hwc = &event->hw;
873 
874 	cxl_pmu_event_stop(event, PERF_EF_UPDATE);
875 	clear_bit(hwc->idx, info->used_counter_bm);
876 	info->hw_events[hwc->idx] = NULL;
877 	perf_event_update_userpage(event);
878 }
879 
cxl_pmu_irq(int irq,void * data)880 static irqreturn_t cxl_pmu_irq(int irq, void *data)
881 {
882 	struct cxl_pmu_info *info = data;
883 	void __iomem *base = info->base;
884 	u64 overflowed;
885 	DECLARE_BITMAP(overflowedbm, 64);
886 	int i;
887 
888 	overflowed = readq(base + CXL_PMU_OVERFLOW_REG);
889 
890 	/* Interrupt may be shared, so maybe it isn't ours */
891 	if (!overflowed)
892 		return IRQ_NONE;
893 
894 	bitmap_from_arr64(overflowedbm, &overflowed, 64);
895 	for_each_set_bit(i, overflowedbm, info->num_counters) {
896 		struct perf_event *event = info->hw_events[i];
897 
898 		if (!event) {
899 			dev_dbg(info->pmu.dev,
900 				"overflow but on non enabled counter %d\n", i);
901 			continue;
902 		}
903 
904 		__cxl_pmu_read(event, true);
905 	}
906 
907 	writeq(overflowed, base + CXL_PMU_OVERFLOW_REG);
908 
909 	return IRQ_HANDLED;
910 }
911 
cxl_pmu_perf_unregister(void * _info)912 static void cxl_pmu_perf_unregister(void *_info)
913 {
914 	struct cxl_pmu_info *info = _info;
915 
916 	perf_pmu_unregister(&info->pmu);
917 }
918 
cxl_pmu_cpuhp_remove(void * _info)919 static void cxl_pmu_cpuhp_remove(void *_info)
920 {
921 	struct cxl_pmu_info *info = _info;
922 
923 	cpuhp_state_remove_instance_nocalls(cxl_pmu_cpuhp_state_num, &info->node);
924 }
925 
cxl_pmu_probe(struct device * dev)926 static int cxl_pmu_probe(struct device *dev)
927 {
928 	struct cxl_pmu *pmu = to_cxl_pmu(dev);
929 	struct pci_dev *pdev = to_pci_dev(dev->parent);
930 	struct cxl_pmu_info *info;
931 	char *irq_name;
932 	char *dev_name;
933 	int rc, irq;
934 
935 	info = devm_kzalloc(dev, sizeof(*info), GFP_KERNEL);
936 	if (!info)
937 		return -ENOMEM;
938 
939 	dev_set_drvdata(dev, info);
940 	INIT_LIST_HEAD(&info->event_caps_fixed);
941 	INIT_LIST_HEAD(&info->event_caps_configurable);
942 
943 	info->base = pmu->base;
944 
945 	info->on_cpu = -1;
946 	rc = cxl_pmu_parse_caps(dev, info);
947 	if (rc)
948 		return rc;
949 
950 	info->hw_events = devm_kcalloc(dev, info->num_counters,
951 				       sizeof(*info->hw_events), GFP_KERNEL);
952 	if (!info->hw_events)
953 		return -ENOMEM;
954 
955 	switch (pmu->type) {
956 	case CXL_PMU_MEMDEV:
957 		dev_name = devm_kasprintf(dev, GFP_KERNEL, "cxl_pmu_mem%d.%d",
958 					  pmu->assoc_id, pmu->index);
959 		break;
960 	}
961 	if (!dev_name)
962 		return -ENOMEM;
963 
964 	info->pmu = (struct pmu) {
965 		.name = dev_name,
966 		.parent = dev,
967 		.module = THIS_MODULE,
968 		.event_init = cxl_pmu_event_init,
969 		.pmu_enable = cxl_pmu_enable,
970 		.pmu_disable = cxl_pmu_disable,
971 		.add = cxl_pmu_event_add,
972 		.del = cxl_pmu_event_del,
973 		.start = cxl_pmu_event_start,
974 		.stop = cxl_pmu_event_stop,
975 		.read = cxl_pmu_read,
976 		.task_ctx_nr = perf_invalid_context,
977 		.attr_groups = cxl_pmu_attr_groups,
978 		.capabilities = PERF_PMU_CAP_NO_EXCLUDE,
979 	};
980 
981 	if (info->irq <= 0)
982 		return -EINVAL;
983 
984 	rc = pci_irq_vector(pdev, info->irq);
985 	if (rc < 0)
986 		return rc;
987 	irq = rc;
988 
989 	irq_name = devm_kasprintf(dev, GFP_KERNEL, "%s_overflow", dev_name);
990 	if (!irq_name)
991 		return -ENOMEM;
992 
993 	rc = devm_request_irq(dev, irq, cxl_pmu_irq, IRQF_SHARED | IRQF_NO_THREAD,
994 			      irq_name, info);
995 	if (rc)
996 		return rc;
997 	info->irq = irq;
998 
999 	rc = cpuhp_state_add_instance(cxl_pmu_cpuhp_state_num, &info->node);
1000 	if (rc)
1001 		return rc;
1002 
1003 	rc = devm_add_action_or_reset(dev, cxl_pmu_cpuhp_remove, info);
1004 	if (rc)
1005 		return rc;
1006 
1007 	rc = perf_pmu_register(&info->pmu, info->pmu.name, -1);
1008 	if (rc)
1009 		return rc;
1010 
1011 	rc = devm_add_action_or_reset(dev, cxl_pmu_perf_unregister, info);
1012 	if (rc)
1013 		return rc;
1014 
1015 	return 0;
1016 }
1017 
1018 static struct cxl_driver cxl_pmu_driver = {
1019 	.name = "cxl_pmu",
1020 	.probe = cxl_pmu_probe,
1021 	.id = CXL_DEVICE_PMU,
1022 };
1023 
cxl_pmu_online_cpu(unsigned int cpu,struct hlist_node * node)1024 static int cxl_pmu_online_cpu(unsigned int cpu, struct hlist_node *node)
1025 {
1026 	struct cxl_pmu_info *info = hlist_entry_safe(node, struct cxl_pmu_info, node);
1027 
1028 	if (info->on_cpu != -1)
1029 		return 0;
1030 
1031 	info->on_cpu = cpu;
1032 	/*
1033 	 * CPU HP lock is held so we should be guaranteed that the CPU hasn't yet
1034 	 * gone away again.
1035 	 */
1036 	WARN_ON(irq_set_affinity(info->irq, cpumask_of(cpu)));
1037 
1038 	return 0;
1039 }
1040 
cxl_pmu_offline_cpu(unsigned int cpu,struct hlist_node * node)1041 static int cxl_pmu_offline_cpu(unsigned int cpu, struct hlist_node *node)
1042 {
1043 	struct cxl_pmu_info *info = hlist_entry_safe(node, struct cxl_pmu_info, node);
1044 	unsigned int target;
1045 
1046 	if (info->on_cpu != cpu)
1047 		return 0;
1048 
1049 	info->on_cpu = -1;
1050 	target = cpumask_any_but(cpu_online_mask, cpu);
1051 	if (target >= nr_cpu_ids) {
1052 		dev_err(info->pmu.dev, "Unable to find a suitable CPU\n");
1053 		return 0;
1054 	}
1055 
1056 	perf_pmu_migrate_context(&info->pmu, cpu, target);
1057 	info->on_cpu = target;
1058 	/*
1059 	 * CPU HP lock is held so we should be guaranteed that this CPU hasn't yet
1060 	 * gone away.
1061 	 */
1062 	WARN_ON(irq_set_affinity(info->irq, cpumask_of(target)));
1063 
1064 	return 0;
1065 }
1066 
cxl_pmu_init(void)1067 static __init int cxl_pmu_init(void)
1068 {
1069 	int rc;
1070 
1071 	rc = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN,
1072 				     "AP_PERF_CXL_PMU_ONLINE",
1073 				     cxl_pmu_online_cpu, cxl_pmu_offline_cpu);
1074 	if (rc < 0)
1075 		return rc;
1076 	cxl_pmu_cpuhp_state_num = rc;
1077 
1078 	rc = cxl_driver_register(&cxl_pmu_driver);
1079 	if (rc)
1080 		cpuhp_remove_multi_state(cxl_pmu_cpuhp_state_num);
1081 
1082 	return rc;
1083 }
1084 
cxl_pmu_exit(void)1085 static __exit void cxl_pmu_exit(void)
1086 {
1087 	cxl_driver_unregister(&cxl_pmu_driver);
1088 	cpuhp_remove_multi_state(cxl_pmu_cpuhp_state_num);
1089 }
1090 
1091 MODULE_DESCRIPTION("CXL Performance Monitor Driver");
1092 MODULE_LICENSE("GPL");
1093 MODULE_IMPORT_NS("CXL");
1094 module_init(cxl_pmu_init);
1095 module_exit(cxl_pmu_exit);
1096 MODULE_ALIAS_CXL(CXL_DEVICE_PMU);
1097