xref: /openbmc/linux/drivers/perf/cxl_pmu.c (revision b608dd67)
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(4, 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(15, 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 
81 static int cxl_pmu_cpuhp_state_num;
82 
83 struct cxl_pmu_ev_cap {
84 	u16 vid;
85 	u16 gid;
86 	u32 msk;
87 	union {
88 		int counter_idx; /* fixed counters */
89 		int event_idx; /* configurable counters */
90 	};
91 	struct list_head node;
92 };
93 
94 #define CXL_PMU_MAX_COUNTERS 64
95 struct cxl_pmu_info {
96 	struct pmu pmu;
97 	void __iomem *base;
98 	struct perf_event **hw_events;
99 	struct list_head event_caps_configurable;
100 	struct list_head event_caps_fixed;
101 	DECLARE_BITMAP(used_counter_bm, CXL_PMU_MAX_COUNTERS);
102 	DECLARE_BITMAP(conf_counter_bm, CXL_PMU_MAX_COUNTERS);
103 	u16 counter_width;
104 	u8 num_counters;
105 	u8 num_event_capabilities;
106 	int on_cpu;
107 	struct hlist_node node;
108 	bool filter_hdm;
109 	int irq;
110 };
111 
112 #define pmu_to_cxl_pmu_info(_pmu) container_of(_pmu, struct cxl_pmu_info, pmu)
113 
114 /*
115  * All CPMU counters are discoverable via the Event Capabilities Registers.
116  * Each Event Capability register contains a a VID / GroupID.
117  * A counter may then count any combination (by summing) of events in
118  * that group which are in the Supported Events Bitmask.
119  * However, there are some complexities to the scheme.
120  *  - Fixed function counters refer to an Event Capabilities register.
121  *    That event capability register is not then used for Configurable
122  *    counters.
123  */
124 static int cxl_pmu_parse_caps(struct device *dev, struct cxl_pmu_info *info)
125 {
126 	unsigned long fixed_counter_event_cap_bm = 0;
127 	void __iomem *base = info->base;
128 	bool freeze_for_enable;
129 	u64 val, eval;
130 	int i;
131 
132 	val = readq(base + CXL_PMU_CAP_REG);
133 	freeze_for_enable = FIELD_GET(CXL_PMU_CAP_WRITEABLE_WHEN_FROZEN, val) &&
134 		FIELD_GET(CXL_PMU_CAP_FREEZE, val);
135 	if (!freeze_for_enable) {
136 		dev_err(dev, "Counters not writable while frozen\n");
137 		return -ENODEV;
138 	}
139 
140 	info->num_counters = FIELD_GET(CXL_PMU_CAP_NUM_COUNTERS_MSK, val) + 1;
141 	info->counter_width = FIELD_GET(CXL_PMU_CAP_COUNTER_WIDTH_MSK, val);
142 	info->num_event_capabilities = FIELD_GET(CXL_PMU_CAP_NUM_EVN_CAP_REG_SUP_MSK, val) + 1;
143 
144 	info->filter_hdm = FIELD_GET(CXL_PMU_CAP_FILTERS_SUP_MSK, val) & CXL_PMU_FILTER_HDM;
145 	if (FIELD_GET(CXL_PMU_CAP_INT, val))
146 		info->irq = FIELD_GET(CXL_PMU_CAP_MSI_N_MSK, val);
147 	else
148 		info->irq = -1;
149 
150 	/* First handle fixed function counters; note if configurable counters found */
151 	for (i = 0; i < info->num_counters; i++) {
152 		struct cxl_pmu_ev_cap *pmu_ev;
153 		u32 events_msk;
154 		u8 group_idx;
155 
156 		val = readq(base + CXL_PMU_COUNTER_CFG_REG(i));
157 
158 		if (FIELD_GET(CXL_PMU_COUNTER_CFG_TYPE_MSK, val) ==
159 			CXL_PMU_COUNTER_CFG_TYPE_CONFIGURABLE) {
160 			set_bit(i, info->conf_counter_bm);
161 		}
162 
163 		if (FIELD_GET(CXL_PMU_COUNTER_CFG_TYPE_MSK, val) !=
164 		    CXL_PMU_COUNTER_CFG_TYPE_FIXED_FUN)
165 			continue;
166 
167 		/* In this case we know which fields are const */
168 		group_idx = FIELD_GET(CXL_PMU_COUNTER_CFG_EVENT_GRP_ID_IDX_MSK, val);
169 		events_msk = FIELD_GET(CXL_PMU_COUNTER_CFG_EVENTS_MSK, val);
170 		eval = readq(base + CXL_PMU_EVENT_CAP_REG(group_idx));
171 		pmu_ev = devm_kzalloc(dev, sizeof(*pmu_ev), GFP_KERNEL);
172 		if (!pmu_ev)
173 			return -ENOMEM;
174 
175 		pmu_ev->vid = FIELD_GET(CXL_PMU_EVENT_CAP_VENDOR_ID_MSK, eval);
176 		pmu_ev->gid = FIELD_GET(CXL_PMU_EVENT_CAP_GROUP_ID_MSK, eval);
177 		/* For a fixed purpose counter use the events mask from the counter CFG */
178 		pmu_ev->msk = events_msk;
179 		pmu_ev->counter_idx = i;
180 		/* This list add is never unwound as all entries deleted on remove */
181 		list_add(&pmu_ev->node, &info->event_caps_fixed);
182 		/*
183 		 * Configurable counters must not use an Event Capability registers that
184 		 * is in use for a Fixed counter
185 		 */
186 		set_bit(group_idx, &fixed_counter_event_cap_bm);
187 	}
188 
189 	if (!bitmap_empty(info->conf_counter_bm, CXL_PMU_MAX_COUNTERS)) {
190 		struct cxl_pmu_ev_cap *pmu_ev;
191 		int j;
192 		/* Walk event capabilities unused by fixed counters */
193 		for_each_clear_bit(j, &fixed_counter_event_cap_bm,
194 				   info->num_event_capabilities) {
195 			pmu_ev = devm_kzalloc(dev, sizeof(*pmu_ev), GFP_KERNEL);
196 			if (!pmu_ev)
197 				return -ENOMEM;
198 
199 			eval = readq(base + CXL_PMU_EVENT_CAP_REG(j));
200 			pmu_ev->vid = FIELD_GET(CXL_PMU_EVENT_CAP_VENDOR_ID_MSK, eval);
201 			pmu_ev->gid = FIELD_GET(CXL_PMU_EVENT_CAP_GROUP_ID_MSK, eval);
202 			pmu_ev->msk = FIELD_GET(CXL_PMU_EVENT_CAP_SUPPORTED_EVENTS_MSK, eval);
203 			pmu_ev->event_idx = j;
204 			list_add(&pmu_ev->node, &info->event_caps_configurable);
205 		}
206 	}
207 
208 	return 0;
209 }
210 
211 static ssize_t cxl_pmu_format_sysfs_show(struct device *dev,
212 					 struct device_attribute *attr, char *buf)
213 {
214 	struct dev_ext_attribute *eattr;
215 
216 	eattr = container_of(attr, struct dev_ext_attribute, attr);
217 
218 	return sysfs_emit(buf, "%s\n", (char *)eattr->var);
219 }
220 
221 #define CXL_PMU_FORMAT_ATTR(_name, _format)\
222 	(&((struct dev_ext_attribute[]) {					\
223 		{								\
224 			.attr = __ATTR(_name, 0444,				\
225 				       cxl_pmu_format_sysfs_show, NULL),	\
226 			.var = (void *)_format					\
227 		}								\
228 		})[0].attr.attr)
229 
230 enum {
231 	cxl_pmu_mask_attr,
232 	cxl_pmu_gid_attr,
233 	cxl_pmu_vid_attr,
234 	cxl_pmu_threshold_attr,
235 	cxl_pmu_invert_attr,
236 	cxl_pmu_edge_attr,
237 	cxl_pmu_hdm_filter_en_attr,
238 	cxl_pmu_hdm_attr,
239 };
240 
241 static struct attribute *cxl_pmu_format_attr[] = {
242 	[cxl_pmu_mask_attr] = CXL_PMU_FORMAT_ATTR(mask, "config:0-31"),
243 	[cxl_pmu_gid_attr] = CXL_PMU_FORMAT_ATTR(gid, "config:32-47"),
244 	[cxl_pmu_vid_attr] = CXL_PMU_FORMAT_ATTR(vid, "config:48-63"),
245 	[cxl_pmu_threshold_attr] = CXL_PMU_FORMAT_ATTR(threshold, "config1:0-15"),
246 	[cxl_pmu_invert_attr] = CXL_PMU_FORMAT_ATTR(invert, "config1:16"),
247 	[cxl_pmu_edge_attr] = CXL_PMU_FORMAT_ATTR(edge, "config1:17"),
248 	[cxl_pmu_hdm_filter_en_attr] = CXL_PMU_FORMAT_ATTR(hdm_filter_en, "config1:18"),
249 	[cxl_pmu_hdm_attr] = CXL_PMU_FORMAT_ATTR(hdm, "config2:0-15"),
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_CONFIG2_HDM_MSK		GENMASK(15, 0)
261 
262 static umode_t cxl_pmu_format_is_visible(struct kobject *kobj,
263 					 struct attribute *attr, int a)
264 {
265 	struct device *dev = kobj_to_dev(kobj);
266 	struct cxl_pmu_info *info = dev_get_drvdata(dev);
267 
268 	/*
269 	 * Filter capability at the CPMU level, so hide the attributes if the particular
270 	 * filter is not supported.
271 	 */
272 	if (!info->filter_hdm &&
273 	    (attr == cxl_pmu_format_attr[cxl_pmu_hdm_filter_en_attr] ||
274 	     attr == cxl_pmu_format_attr[cxl_pmu_hdm_attr]))
275 		return 0;
276 
277 	return attr->mode;
278 }
279 
280 static const struct attribute_group cxl_pmu_format_group = {
281 	.name = "format",
282 	.attrs = cxl_pmu_format_attr,
283 	.is_visible = cxl_pmu_format_is_visible,
284 };
285 
286 static u32 cxl_pmu_config_get_mask(struct perf_event *event)
287 {
288 	return FIELD_GET(CXL_PMU_ATTR_CONFIG_MASK_MSK, event->attr.config);
289 }
290 
291 static u16 cxl_pmu_config_get_gid(struct perf_event *event)
292 {
293 	return FIELD_GET(CXL_PMU_ATTR_CONFIG_GID_MSK, event->attr.config);
294 }
295 
296 static u16 cxl_pmu_config_get_vid(struct perf_event *event)
297 {
298 	return FIELD_GET(CXL_PMU_ATTR_CONFIG_VID_MSK, event->attr.config);
299 }
300 
301 static u8 cxl_pmu_config1_get_threshold(struct perf_event *event)
302 {
303 	return FIELD_GET(CXL_PMU_ATTR_CONFIG1_THRESHOLD_MSK, event->attr.config1);
304 }
305 
306 static bool cxl_pmu_config1_get_invert(struct perf_event *event)
307 {
308 	return FIELD_GET(CXL_PMU_ATTR_CONFIG1_INVERT_MSK, event->attr.config1);
309 }
310 
311 static bool cxl_pmu_config1_get_edge(struct perf_event *event)
312 {
313 	return FIELD_GET(CXL_PMU_ATTR_CONFIG1_EDGE_MSK, event->attr.config1);
314 }
315 
316 /*
317  * CPMU specification allows for 8 filters, each with a 16 bit value...
318  * So we need to find 8x16bits to store it in.
319  * As the value used for disable is 0xffff, a separate enable switch
320  * is needed.
321  */
322 
323 static bool cxl_pmu_config1_hdm_filter_en(struct perf_event *event)
324 {
325 	return FIELD_GET(CXL_PMU_ATTR_CONFIG1_FILTER_EN_MSK, event->attr.config1);
326 }
327 
328 static u16 cxl_pmu_config2_get_hdm_decoder(struct perf_event *event)
329 {
330 	return FIELD_GET(CXL_PMU_ATTR_CONFIG2_HDM_MSK, event->attr.config2);
331 }
332 
333 static ssize_t cxl_pmu_event_sysfs_show(struct device *dev,
334 					struct device_attribute *attr, char *buf)
335 {
336 	struct perf_pmu_events_attr *pmu_attr =
337 		container_of(attr, struct perf_pmu_events_attr, attr);
338 
339 	return sysfs_emit(buf, "config=%#llx\n", pmu_attr->id);
340 }
341 
342 #define CXL_PMU_EVENT_ATTR(_name, _vid, _gid, _msk)			\
343 	PMU_EVENT_ATTR_ID(_name, cxl_pmu_event_sysfs_show,		\
344 			  ((u64)(_vid) << 48) | ((u64)(_gid) << 32) | (u64)(_msk))
345 
346 /* For CXL spec defined events */
347 #define CXL_PMU_EVENT_CXL_ATTR(_name, _gid, _msk)			\
348 	CXL_PMU_EVENT_ATTR(_name, PCI_DVSEC_VENDOR_ID_CXL, _gid, _msk)
349 
350 static struct attribute *cxl_pmu_event_attrs[] = {
351 	CXL_PMU_EVENT_CXL_ATTR(clock_ticks,			CXL_PMU_GID_CLOCK_TICKS, BIT(0)),
352 	/* CXL rev 3.0 Table 3-17 - Device to Host Requests */
353 	CXL_PMU_EVENT_CXL_ATTR(d2h_req_rdcurr,			CXL_PMU_GID_D2H_REQ, BIT(1)),
354 	CXL_PMU_EVENT_CXL_ATTR(d2h_req_rdown,			CXL_PMU_GID_D2H_REQ, BIT(2)),
355 	CXL_PMU_EVENT_CXL_ATTR(d2h_req_rdshared,		CXL_PMU_GID_D2H_REQ, BIT(3)),
356 	CXL_PMU_EVENT_CXL_ATTR(d2h_req_rdany,			CXL_PMU_GID_D2H_REQ, BIT(4)),
357 	CXL_PMU_EVENT_CXL_ATTR(d2h_req_rdownnodata,		CXL_PMU_GID_D2H_REQ, BIT(5)),
358 	CXL_PMU_EVENT_CXL_ATTR(d2h_req_itomwr,			CXL_PMU_GID_D2H_REQ, BIT(6)),
359 	CXL_PMU_EVENT_CXL_ATTR(d2h_req_wrcurr,			CXL_PMU_GID_D2H_REQ, BIT(7)),
360 	CXL_PMU_EVENT_CXL_ATTR(d2h_req_clflush,			CXL_PMU_GID_D2H_REQ, BIT(8)),
361 	CXL_PMU_EVENT_CXL_ATTR(d2h_req_cleanevict,		CXL_PMU_GID_D2H_REQ, BIT(9)),
362 	CXL_PMU_EVENT_CXL_ATTR(d2h_req_dirtyevict,		CXL_PMU_GID_D2H_REQ, BIT(10)),
363 	CXL_PMU_EVENT_CXL_ATTR(d2h_req_cleanevictnodata,	CXL_PMU_GID_D2H_REQ, BIT(11)),
364 	CXL_PMU_EVENT_CXL_ATTR(d2h_req_wowrinv,			CXL_PMU_GID_D2H_REQ, BIT(12)),
365 	CXL_PMU_EVENT_CXL_ATTR(d2h_req_wowrinvf,		CXL_PMU_GID_D2H_REQ, BIT(13)),
366 	CXL_PMU_EVENT_CXL_ATTR(d2h_req_wrinv,			CXL_PMU_GID_D2H_REQ, BIT(14)),
367 	CXL_PMU_EVENT_CXL_ATTR(d2h_req_cacheflushed,		CXL_PMU_GID_D2H_REQ, BIT(16)),
368 	/* CXL rev 3.0 Table 3-20 - D2H Repsonse Encodings */
369 	CXL_PMU_EVENT_CXL_ATTR(d2h_rsp_rspihiti,		CXL_PMU_GID_D2H_RSP, BIT(4)),
370 	CXL_PMU_EVENT_CXL_ATTR(d2h_rsp_rspvhitv,		CXL_PMU_GID_D2H_RSP, BIT(6)),
371 	CXL_PMU_EVENT_CXL_ATTR(d2h_rsp_rspihitse,		CXL_PMU_GID_D2H_RSP, BIT(5)),
372 	CXL_PMU_EVENT_CXL_ATTR(d2h_rsp_rspshitse,		CXL_PMU_GID_D2H_RSP, BIT(1)),
373 	CXL_PMU_EVENT_CXL_ATTR(d2h_rsp_rspsfwdm,		CXL_PMU_GID_D2H_RSP, BIT(7)),
374 	CXL_PMU_EVENT_CXL_ATTR(d2h_rsp_rspifwdm,		CXL_PMU_GID_D2H_RSP, BIT(15)),
375 	CXL_PMU_EVENT_CXL_ATTR(d2h_rsp_rspvfwdv,		CXL_PMU_GID_D2H_RSP, BIT(22)),
376 	/* CXL rev 3.0 Table 3-21 - CXL.cache - Mapping of H2D Requests to D2H Responses */
377 	CXL_PMU_EVENT_CXL_ATTR(h2d_req_snpdata,			CXL_PMU_GID_H2D_REQ, BIT(1)),
378 	CXL_PMU_EVENT_CXL_ATTR(h2d_req_snpinv,			CXL_PMU_GID_H2D_REQ, BIT(2)),
379 	CXL_PMU_EVENT_CXL_ATTR(h2d_req_snpcur,			CXL_PMU_GID_H2D_REQ, BIT(3)),
380 	/* CXL rev 3.0 Table 3-22 - H2D Response Opcode Encodings */
381 	CXL_PMU_EVENT_CXL_ATTR(h2d_rsp_writepull,		CXL_PMU_GID_H2D_RSP, BIT(1)),
382 	CXL_PMU_EVENT_CXL_ATTR(h2d_rsp_go,			CXL_PMU_GID_H2D_RSP, BIT(4)),
383 	CXL_PMU_EVENT_CXL_ATTR(h2d_rsp_gowritepull,		CXL_PMU_GID_H2D_RSP, BIT(5)),
384 	CXL_PMU_EVENT_CXL_ATTR(h2d_rsp_extcmp,			CXL_PMU_GID_H2D_RSP, BIT(6)),
385 	CXL_PMU_EVENT_CXL_ATTR(h2d_rsp_gowritepulldrop,		CXL_PMU_GID_H2D_RSP, BIT(8)),
386 	CXL_PMU_EVENT_CXL_ATTR(h2d_rsp_fastgowritepull,		CXL_PMU_GID_H2D_RSP, BIT(13)),
387 	CXL_PMU_EVENT_CXL_ATTR(h2d_rsp_goerrwritepull,		CXL_PMU_GID_H2D_RSP, BIT(15)),
388 	/* CXL rev 3.0 Table 13-5 directly lists these */
389 	CXL_PMU_EVENT_CXL_ATTR(cachedata_d2h_data,		CXL_PMU_GID_CACHE_DATA, BIT(0)),
390 	CXL_PMU_EVENT_CXL_ATTR(cachedata_h2d_data,		CXL_PMU_GID_CACHE_DATA, BIT(1)),
391 	/* CXL rev 3.0 Table 3-29 M2S Req Memory Opcodes */
392 	CXL_PMU_EVENT_CXL_ATTR(m2s_req_meminv,			CXL_PMU_GID_M2S_REQ, BIT(0)),
393 	CXL_PMU_EVENT_CXL_ATTR(m2s_req_memrd,			CXL_PMU_GID_M2S_REQ, BIT(1)),
394 	CXL_PMU_EVENT_CXL_ATTR(m2s_req_memrddata,		CXL_PMU_GID_M2S_REQ, BIT(2)),
395 	CXL_PMU_EVENT_CXL_ATTR(m2s_req_memrdfwd,		CXL_PMU_GID_M2S_REQ, BIT(3)),
396 	CXL_PMU_EVENT_CXL_ATTR(m2s_req_memwrfwd,		CXL_PMU_GID_M2S_REQ, BIT(4)),
397 	CXL_PMU_EVENT_CXL_ATTR(m2s_req_memspecrd,		CXL_PMU_GID_M2S_REQ, BIT(8)),
398 	CXL_PMU_EVENT_CXL_ATTR(m2s_req_meminvnt,		CXL_PMU_GID_M2S_REQ, BIT(9)),
399 	CXL_PMU_EVENT_CXL_ATTR(m2s_req_memcleanevict,		CXL_PMU_GID_M2S_REQ, BIT(10)),
400 	/* CXL rev 3.0 Table 3-35 M2S RwD Memory Opcodes */
401 	CXL_PMU_EVENT_CXL_ATTR(m2s_rwd_memwr,			CXL_PMU_GID_M2S_RWD, BIT(1)),
402 	CXL_PMU_EVENT_CXL_ATTR(m2s_rwd_memwrptl,		CXL_PMU_GID_M2S_RWD, BIT(2)),
403 	CXL_PMU_EVENT_CXL_ATTR(m2s_rwd_biconflict,		CXL_PMU_GID_M2S_RWD, BIT(4)),
404 	/* CXL rev 3.0 Table 3-38 M2S BIRsp Memory Opcodes */
405 	CXL_PMU_EVENT_CXL_ATTR(m2s_birsp_i,			CXL_PMU_GID_M2S_BIRSP, BIT(0)),
406 	CXL_PMU_EVENT_CXL_ATTR(m2s_birsp_s,			CXL_PMU_GID_M2S_BIRSP, BIT(1)),
407 	CXL_PMU_EVENT_CXL_ATTR(m2s_birsp_e,			CXL_PMU_GID_M2S_BIRSP, BIT(2)),
408 	CXL_PMU_EVENT_CXL_ATTR(m2s_birsp_iblk,			CXL_PMU_GID_M2S_BIRSP, BIT(4)),
409 	CXL_PMU_EVENT_CXL_ATTR(m2s_birsp_sblk,			CXL_PMU_GID_M2S_BIRSP, BIT(5)),
410 	CXL_PMU_EVENT_CXL_ATTR(m2s_birsp_eblk,			CXL_PMU_GID_M2S_BIRSP, BIT(6)),
411 	/* CXL rev 3.0 Table 3-40 S2M BISnp Opcodes */
412 	CXL_PMU_EVENT_CXL_ATTR(s2m_bisnp_cur,			CXL_PMU_GID_S2M_BISNP, BIT(0)),
413 	CXL_PMU_EVENT_CXL_ATTR(s2m_bisnp_data,			CXL_PMU_GID_S2M_BISNP, BIT(1)),
414 	CXL_PMU_EVENT_CXL_ATTR(s2m_bisnp_inv,			CXL_PMU_GID_S2M_BISNP, BIT(2)),
415 	CXL_PMU_EVENT_CXL_ATTR(s2m_bisnp_curblk,		CXL_PMU_GID_S2M_BISNP, BIT(4)),
416 	CXL_PMU_EVENT_CXL_ATTR(s2m_bisnp_datblk,		CXL_PMU_GID_S2M_BISNP, BIT(5)),
417 	CXL_PMU_EVENT_CXL_ATTR(s2m_bisnp_invblk,		CXL_PMU_GID_S2M_BISNP, BIT(6)),
418 	/* CXL rev 3.0 Table 3-43 S2M NDR Opcopdes */
419 	CXL_PMU_EVENT_CXL_ATTR(s2m_ndr_cmp,			CXL_PMU_GID_S2M_NDR, BIT(0)),
420 	CXL_PMU_EVENT_CXL_ATTR(s2m_ndr_cmps,			CXL_PMU_GID_S2M_NDR, BIT(1)),
421 	CXL_PMU_EVENT_CXL_ATTR(s2m_ndr_cmpe,			CXL_PMU_GID_S2M_NDR, BIT(2)),
422 	CXL_PMU_EVENT_CXL_ATTR(s2m_ndr_biconflictack,		CXL_PMU_GID_S2M_NDR, BIT(3)),
423 	/* CXL rev 3.0 Table 3-46 S2M DRS opcodes */
424 	CXL_PMU_EVENT_CXL_ATTR(s2m_drs_memdata,			CXL_PMU_GID_S2M_DRS, BIT(0)),
425 	CXL_PMU_EVENT_CXL_ATTR(s2m_drs_memdatanxm,		CXL_PMU_GID_S2M_DRS, BIT(1)),
426 	/* CXL rev 3.0 Table 13-5 directly lists these */
427 	CXL_PMU_EVENT_CXL_ATTR(ddr_act,				CXL_PMU_GID_DDR, BIT(0)),
428 	CXL_PMU_EVENT_CXL_ATTR(ddr_pre,				CXL_PMU_GID_DDR, BIT(1)),
429 	CXL_PMU_EVENT_CXL_ATTR(ddr_casrd,			CXL_PMU_GID_DDR, BIT(2)),
430 	CXL_PMU_EVENT_CXL_ATTR(ddr_caswr,			CXL_PMU_GID_DDR, BIT(3)),
431 	CXL_PMU_EVENT_CXL_ATTR(ddr_refresh,			CXL_PMU_GID_DDR, BIT(4)),
432 	CXL_PMU_EVENT_CXL_ATTR(ddr_selfrefreshent,		CXL_PMU_GID_DDR, BIT(5)),
433 	CXL_PMU_EVENT_CXL_ATTR(ddr_rfm,				CXL_PMU_GID_DDR, BIT(6)),
434 	NULL
435 };
436 
437 static struct cxl_pmu_ev_cap *cxl_pmu_find_fixed_counter_ev_cap(struct cxl_pmu_info *info,
438 								int vid, int gid, int msk)
439 {
440 	struct cxl_pmu_ev_cap *pmu_ev;
441 
442 	list_for_each_entry(pmu_ev, &info->event_caps_fixed, node) {
443 		if (vid != pmu_ev->vid || gid != pmu_ev->gid)
444 			continue;
445 
446 		/* Precise match for fixed counter */
447 		if (msk == pmu_ev->msk)
448 			return pmu_ev;
449 	}
450 
451 	return ERR_PTR(-EINVAL);
452 }
453 
454 static struct cxl_pmu_ev_cap *cxl_pmu_find_config_counter_ev_cap(struct cxl_pmu_info *info,
455 								 int vid, int gid, int msk)
456 {
457 	struct cxl_pmu_ev_cap *pmu_ev;
458 
459 	list_for_each_entry(pmu_ev, &info->event_caps_configurable, node) {
460 		if (vid != pmu_ev->vid || gid != pmu_ev->gid)
461 			continue;
462 
463 		/* Request mask must be subset of supported */
464 		if (msk & ~pmu_ev->msk)
465 			continue;
466 
467 		return pmu_ev;
468 	}
469 
470 	return ERR_PTR(-EINVAL);
471 }
472 
473 static umode_t cxl_pmu_event_is_visible(struct kobject *kobj, struct attribute *attr, int a)
474 {
475 	struct device_attribute *dev_attr = container_of(attr, struct device_attribute, attr);
476 	struct perf_pmu_events_attr *pmu_attr =
477 		container_of(dev_attr, struct perf_pmu_events_attr, attr);
478 	struct device *dev = kobj_to_dev(kobj);
479 	struct cxl_pmu_info *info = dev_get_drvdata(dev);
480 	int vid = FIELD_GET(CXL_PMU_ATTR_CONFIG_VID_MSK, pmu_attr->id);
481 	int gid = FIELD_GET(CXL_PMU_ATTR_CONFIG_GID_MSK, pmu_attr->id);
482 	int msk = FIELD_GET(CXL_PMU_ATTR_CONFIG_MASK_MSK, pmu_attr->id);
483 
484 	if (!IS_ERR(cxl_pmu_find_fixed_counter_ev_cap(info, vid, gid, msk)))
485 		return attr->mode;
486 
487 	if (!IS_ERR(cxl_pmu_find_config_counter_ev_cap(info, vid, gid, msk)))
488 		return attr->mode;
489 
490 	return 0;
491 }
492 
493 static const struct attribute_group cxl_pmu_events = {
494 	.name = "events",
495 	.attrs = cxl_pmu_event_attrs,
496 	.is_visible = cxl_pmu_event_is_visible,
497 };
498 
499 static ssize_t cpumask_show(struct device *dev, struct device_attribute *attr,
500 			    char *buf)
501 {
502 	struct cxl_pmu_info *info = dev_get_drvdata(dev);
503 
504 	return cpumap_print_to_pagebuf(true, buf, cpumask_of(info->on_cpu));
505 }
506 static DEVICE_ATTR_RO(cpumask);
507 
508 static struct attribute *cxl_pmu_cpumask_attrs[] = {
509 	&dev_attr_cpumask.attr,
510 	NULL
511 };
512 
513 static const struct attribute_group cxl_pmu_cpumask_group = {
514 	.attrs = cxl_pmu_cpumask_attrs,
515 };
516 
517 static const struct attribute_group *cxl_pmu_attr_groups[] = {
518 	&cxl_pmu_events,
519 	&cxl_pmu_format_group,
520 	&cxl_pmu_cpumask_group,
521 	NULL
522 };
523 
524 /* If counter_idx == NULL, don't try to allocate a counter. */
525 static int cxl_pmu_get_event_idx(struct perf_event *event, int *counter_idx,
526 				 int *event_idx)
527 {
528 	struct cxl_pmu_info *info = pmu_to_cxl_pmu_info(event->pmu);
529 	DECLARE_BITMAP(configurable_and_free, CXL_PMU_MAX_COUNTERS);
530 	struct cxl_pmu_ev_cap *pmu_ev;
531 	u32 mask;
532 	u16 gid, vid;
533 	int i;
534 
535 	vid = cxl_pmu_config_get_vid(event);
536 	gid = cxl_pmu_config_get_gid(event);
537 	mask = cxl_pmu_config_get_mask(event);
538 
539 	pmu_ev = cxl_pmu_find_fixed_counter_ev_cap(info, vid, gid, mask);
540 	if (!IS_ERR(pmu_ev)) {
541 		if (!counter_idx)
542 			return 0;
543 		if (!test_bit(pmu_ev->counter_idx, info->used_counter_bm)) {
544 			*counter_idx = pmu_ev->counter_idx;
545 			return 0;
546 		}
547 		/* Fixed counter is in use, but maybe a configurable one? */
548 	}
549 
550 	pmu_ev = cxl_pmu_find_config_counter_ev_cap(info, vid, gid, mask);
551 	if (!IS_ERR(pmu_ev)) {
552 		if (!counter_idx)
553 			return 0;
554 
555 		bitmap_andnot(configurable_and_free, info->conf_counter_bm,
556 			info->used_counter_bm, CXL_PMU_MAX_COUNTERS);
557 
558 		i = find_first_bit(configurable_and_free, CXL_PMU_MAX_COUNTERS);
559 		if (i == CXL_PMU_MAX_COUNTERS)
560 			return -EINVAL;
561 
562 		*counter_idx = i;
563 		return 0;
564 	}
565 
566 	return -EINVAL;
567 }
568 
569 static int cxl_pmu_event_init(struct perf_event *event)
570 {
571 	struct cxl_pmu_info *info = pmu_to_cxl_pmu_info(event->pmu);
572 	int rc;
573 
574 	/* Top level type sanity check - is this a Hardware Event being requested */
575 	if (event->attr.type != event->pmu->type)
576 		return -ENOENT;
577 
578 	if (is_sampling_event(event) || event->attach_state & PERF_ATTACH_TASK)
579 		return -EOPNOTSUPP;
580 	/* TODO: Validation of any filter */
581 
582 	/*
583 	 * Verify that it is possible to count what was requested. Either must
584 	 * be a fixed counter that is a precise match or a configurable counter
585 	 * where this is a subset.
586 	 */
587 	rc = cxl_pmu_get_event_idx(event, NULL, NULL);
588 	if (rc < 0)
589 		return rc;
590 
591 	event->cpu = info->on_cpu;
592 
593 	return 0;
594 }
595 
596 static void cxl_pmu_enable(struct pmu *pmu)
597 {
598 	struct cxl_pmu_info *info = pmu_to_cxl_pmu_info(pmu);
599 	void __iomem *base = info->base;
600 
601 	/* Can assume frozen at this stage */
602 	writeq(0, base + CXL_PMU_FREEZE_REG);
603 }
604 
605 static void cxl_pmu_disable(struct pmu *pmu)
606 {
607 	struct cxl_pmu_info *info = pmu_to_cxl_pmu_info(pmu);
608 	void __iomem *base = info->base;
609 
610 	/*
611 	 * Whilst bits above number of counters are RsvdZ
612 	 * they are unlikely to be repurposed given
613 	 * number of counters is allowed to be 64 leaving
614 	 * no reserved bits.  Hence this is only slightly
615 	 * naughty.
616 	 */
617 	writeq(GENMASK_ULL(63, 0), base + CXL_PMU_FREEZE_REG);
618 }
619 
620 static void cxl_pmu_event_start(struct perf_event *event, int flags)
621 {
622 	struct cxl_pmu_info *info = pmu_to_cxl_pmu_info(event->pmu);
623 	struct hw_perf_event *hwc = &event->hw;
624 	void __iomem *base = info->base;
625 	u64 cfg;
626 
627 	/*
628 	 * All paths to here should either set these flags directly or
629 	 * call cxl_pmu_event_stop() which will ensure the correct state.
630 	 */
631 	if (WARN_ON_ONCE(!(hwc->state & PERF_HES_STOPPED)))
632 		return;
633 
634 	WARN_ON_ONCE(!(hwc->state & PERF_HES_UPTODATE));
635 	hwc->state = 0;
636 
637 	/*
638 	 * Currently only hdm filter control is implemnted, this code will
639 	 * want generalizing when more filters are added.
640 	 */
641 	if (info->filter_hdm) {
642 		if (cxl_pmu_config1_hdm_filter_en(event))
643 			cfg = cxl_pmu_config2_get_hdm_decoder(event);
644 		else
645 			cfg = GENMASK(15, 0); /* No filtering if 0xFFFF_FFFF */
646 		writeq(cfg, base + CXL_PMU_FILTER_CFG_REG(hwc->idx, 0));
647 	}
648 
649 	cfg = readq(base + CXL_PMU_COUNTER_CFG_REG(hwc->idx));
650 	cfg |= FIELD_PREP(CXL_PMU_COUNTER_CFG_INT_ON_OVRFLW, 1);
651 	cfg |= FIELD_PREP(CXL_PMU_COUNTER_CFG_FREEZE_ON_OVRFLW, 1);
652 	cfg |= FIELD_PREP(CXL_PMU_COUNTER_CFG_ENABLE, 1);
653 	cfg |= FIELD_PREP(CXL_PMU_COUNTER_CFG_EDGE,
654 			  cxl_pmu_config1_get_edge(event) ? 1 : 0);
655 	cfg |= FIELD_PREP(CXL_PMU_COUNTER_CFG_INVERT,
656 			  cxl_pmu_config1_get_invert(event) ? 1 : 0);
657 
658 	/* Fixed purpose counters have next two fields RO */
659 	if (test_bit(hwc->idx, info->conf_counter_bm)) {
660 		cfg |= FIELD_PREP(CXL_PMU_COUNTER_CFG_EVENT_GRP_ID_IDX_MSK,
661 				  hwc->event_base);
662 		cfg |= FIELD_PREP(CXL_PMU_COUNTER_CFG_EVENTS_MSK,
663 				  cxl_pmu_config_get_mask(event));
664 	}
665 	cfg &= ~CXL_PMU_COUNTER_CFG_THRESHOLD_MSK;
666 	/*
667 	 * For events that generate only 1 count per clock the CXL 3.0 spec
668 	 * states the threshold shall be set to 1 but if set to 0 it will
669 	 * count the raw value anwyay?
670 	 * There is no definition of what events will count multiple per cycle
671 	 * and hence to which non 1 values of threshold can apply.
672 	 * (CXL 3.0 8.2.7.2.1 Counter Configuration - threshold field definition)
673 	 */
674 	cfg |= FIELD_PREP(CXL_PMU_COUNTER_CFG_THRESHOLD_MSK,
675 			  cxl_pmu_config1_get_threshold(event));
676 	writeq(cfg, base + CXL_PMU_COUNTER_CFG_REG(hwc->idx));
677 
678 	local64_set(&hwc->prev_count, 0);
679 	writeq(0, base + CXL_PMU_COUNTER_REG(hwc->idx));
680 
681 	perf_event_update_userpage(event);
682 }
683 
684 static u64 cxl_pmu_read_counter(struct perf_event *event)
685 {
686 	struct cxl_pmu_info *info = pmu_to_cxl_pmu_info(event->pmu);
687 	void __iomem *base = info->base;
688 
689 	return readq(base + CXL_PMU_COUNTER_REG(event->hw.idx));
690 }
691 
692 static void __cxl_pmu_read(struct perf_event *event, bool overflow)
693 {
694 	struct cxl_pmu_info *info = pmu_to_cxl_pmu_info(event->pmu);
695 	struct hw_perf_event *hwc = &event->hw;
696 	u64 new_cnt, prev_cnt, delta;
697 
698 	do {
699 		prev_cnt = local64_read(&hwc->prev_count);
700 		new_cnt = cxl_pmu_read_counter(event);
701 	} while (local64_cmpxchg(&hwc->prev_count, prev_cnt, new_cnt) != prev_cnt);
702 
703 	/*
704 	 * If we know an overflow occur then take that into account.
705 	 * Note counter is not reset as that would lose events
706 	 */
707 	delta = (new_cnt - prev_cnt) & GENMASK_ULL(info->counter_width - 1, 0);
708 	if (overflow && delta < GENMASK_ULL(info->counter_width - 1, 0))
709 		delta += (1UL << info->counter_width);
710 
711 	local64_add(delta, &event->count);
712 }
713 
714 static void cxl_pmu_read(struct perf_event *event)
715 {
716 	__cxl_pmu_read(event, false);
717 }
718 
719 static void cxl_pmu_event_stop(struct perf_event *event, int flags)
720 {
721 	struct cxl_pmu_info *info = pmu_to_cxl_pmu_info(event->pmu);
722 	void __iomem *base = info->base;
723 	struct hw_perf_event *hwc = &event->hw;
724 	u64 cfg;
725 
726 	cxl_pmu_read(event);
727 	WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
728 	hwc->state |= PERF_HES_STOPPED;
729 
730 	cfg = readq(base + CXL_PMU_COUNTER_CFG_REG(hwc->idx));
731 	cfg &= ~(FIELD_PREP(CXL_PMU_COUNTER_CFG_INT_ON_OVRFLW, 1) |
732 		 FIELD_PREP(CXL_PMU_COUNTER_CFG_ENABLE, 1));
733 	writeq(cfg, base + CXL_PMU_COUNTER_CFG_REG(hwc->idx));
734 
735 	hwc->state |= PERF_HES_UPTODATE;
736 }
737 
738 static int cxl_pmu_event_add(struct perf_event *event, int flags)
739 {
740 	struct cxl_pmu_info *info = pmu_to_cxl_pmu_info(event->pmu);
741 	struct hw_perf_event *hwc = &event->hw;
742 	int idx, rc;
743 	int event_idx = 0;
744 
745 	hwc->state = PERF_HES_STOPPED | PERF_HES_UPTODATE;
746 
747 	rc = cxl_pmu_get_event_idx(event, &idx, &event_idx);
748 	if (rc < 0)
749 		return rc;
750 
751 	hwc->idx = idx;
752 
753 	/* Only set for configurable counters */
754 	hwc->event_base = event_idx;
755 	info->hw_events[idx] = event;
756 	set_bit(idx, info->used_counter_bm);
757 
758 	if (flags & PERF_EF_START)
759 		cxl_pmu_event_start(event, PERF_EF_RELOAD);
760 
761 	return 0;
762 }
763 
764 static void cxl_pmu_event_del(struct perf_event *event, int flags)
765 {
766 	struct cxl_pmu_info *info = pmu_to_cxl_pmu_info(event->pmu);
767 	struct hw_perf_event *hwc = &event->hw;
768 
769 	cxl_pmu_event_stop(event, PERF_EF_UPDATE);
770 	clear_bit(hwc->idx, info->used_counter_bm);
771 	info->hw_events[hwc->idx] = NULL;
772 	perf_event_update_userpage(event);
773 }
774 
775 static irqreturn_t cxl_pmu_irq(int irq, void *data)
776 {
777 	struct cxl_pmu_info *info = data;
778 	void __iomem *base = info->base;
779 	u64 overflowed;
780 	DECLARE_BITMAP(overflowedbm, 64);
781 	int i;
782 
783 	overflowed = readq(base + CXL_PMU_OVERFLOW_REG);
784 
785 	/* Interrupt may be shared, so maybe it isn't ours */
786 	if (!overflowed)
787 		return IRQ_NONE;
788 
789 	bitmap_from_arr64(overflowedbm, &overflowed, 64);
790 	for_each_set_bit(i, overflowedbm, info->num_counters) {
791 		struct perf_event *event = info->hw_events[i];
792 
793 		if (!event) {
794 			dev_dbg(info->pmu.dev,
795 				"overflow but on non enabled counter %d\n", i);
796 			continue;
797 		}
798 
799 		__cxl_pmu_read(event, true);
800 	}
801 
802 	writeq(overflowed, base + CXL_PMU_OVERFLOW_REG);
803 
804 	return IRQ_HANDLED;
805 }
806 
807 static void cxl_pmu_perf_unregister(void *_info)
808 {
809 	struct cxl_pmu_info *info = _info;
810 
811 	perf_pmu_unregister(&info->pmu);
812 }
813 
814 static void cxl_pmu_cpuhp_remove(void *_info)
815 {
816 	struct cxl_pmu_info *info = _info;
817 
818 	cpuhp_state_remove_instance_nocalls(cxl_pmu_cpuhp_state_num, &info->node);
819 }
820 
821 static int cxl_pmu_probe(struct device *dev)
822 {
823 	struct cxl_pmu *pmu = to_cxl_pmu(dev);
824 	struct pci_dev *pdev = to_pci_dev(dev->parent);
825 	struct cxl_pmu_info *info;
826 	char *irq_name;
827 	char *dev_name;
828 	int rc, irq;
829 
830 	info = devm_kzalloc(dev, sizeof(*info), GFP_KERNEL);
831 	if (!info)
832 		return -ENOMEM;
833 
834 	dev_set_drvdata(dev, info);
835 	INIT_LIST_HEAD(&info->event_caps_fixed);
836 	INIT_LIST_HEAD(&info->event_caps_configurable);
837 
838 	info->base = pmu->base;
839 
840 	info->on_cpu = -1;
841 	rc = cxl_pmu_parse_caps(dev, info);
842 	if (rc)
843 		return rc;
844 
845 	info->hw_events = devm_kcalloc(dev, sizeof(*info->hw_events),
846 				       info->num_counters, GFP_KERNEL);
847 	if (!info->hw_events)
848 		return -ENOMEM;
849 
850 	switch (pmu->type) {
851 	case CXL_PMU_MEMDEV:
852 		dev_name = devm_kasprintf(dev, GFP_KERNEL, "cxl_pmu_mem%d.%d",
853 					  pmu->assoc_id, pmu->index);
854 		break;
855 	}
856 	if (!dev_name)
857 		return -ENOMEM;
858 
859 	info->pmu = (struct pmu) {
860 		.name = dev_name,
861 		.parent = dev,
862 		.module = THIS_MODULE,
863 		.event_init = cxl_pmu_event_init,
864 		.pmu_enable = cxl_pmu_enable,
865 		.pmu_disable = cxl_pmu_disable,
866 		.add = cxl_pmu_event_add,
867 		.del = cxl_pmu_event_del,
868 		.start = cxl_pmu_event_start,
869 		.stop = cxl_pmu_event_stop,
870 		.read = cxl_pmu_read,
871 		.task_ctx_nr = perf_invalid_context,
872 		.attr_groups = cxl_pmu_attr_groups,
873 		.capabilities = PERF_PMU_CAP_NO_EXCLUDE,
874 	};
875 
876 	if (info->irq <= 0)
877 		return -EINVAL;
878 
879 	rc = pci_irq_vector(pdev, info->irq);
880 	if (rc < 0)
881 		return rc;
882 	irq = rc;
883 
884 	irq_name = devm_kasprintf(dev, GFP_KERNEL, "%s_overflow\n", dev_name);
885 	if (!irq_name)
886 		return -ENOMEM;
887 
888 	rc = devm_request_irq(dev, irq, cxl_pmu_irq, IRQF_SHARED | IRQF_ONESHOT,
889 			      irq_name, info);
890 	if (rc)
891 		return rc;
892 	info->irq = irq;
893 
894 	rc = cpuhp_state_add_instance(cxl_pmu_cpuhp_state_num, &info->node);
895 	if (rc)
896 		return rc;
897 
898 	rc = devm_add_action_or_reset(dev, cxl_pmu_cpuhp_remove, info);
899 	if (rc)
900 		return rc;
901 
902 	rc = perf_pmu_register(&info->pmu, info->pmu.name, -1);
903 	if (rc)
904 		return rc;
905 
906 	rc = devm_add_action_or_reset(dev, cxl_pmu_perf_unregister, info);
907 	if (rc)
908 		return rc;
909 
910 	return 0;
911 }
912 
913 static struct cxl_driver cxl_pmu_driver = {
914 	.name = "cxl_pmu",
915 	.probe = cxl_pmu_probe,
916 	.id = CXL_DEVICE_PMU,
917 };
918 
919 static int cxl_pmu_online_cpu(unsigned int cpu, struct hlist_node *node)
920 {
921 	struct cxl_pmu_info *info = hlist_entry_safe(node, struct cxl_pmu_info, node);
922 
923 	if (info->on_cpu != -1)
924 		return 0;
925 
926 	info->on_cpu = cpu;
927 	/*
928 	 * CPU HP lock is held so we should be guaranteed that the CPU hasn't yet
929 	 * gone away again.
930 	 */
931 	WARN_ON(irq_set_affinity(info->irq, cpumask_of(cpu)));
932 
933 	return 0;
934 }
935 
936 static int cxl_pmu_offline_cpu(unsigned int cpu, struct hlist_node *node)
937 {
938 	struct cxl_pmu_info *info = hlist_entry_safe(node, struct cxl_pmu_info, node);
939 	unsigned int target;
940 
941 	if (info->on_cpu != cpu)
942 		return 0;
943 
944 	info->on_cpu = -1;
945 	target = cpumask_any_but(cpu_online_mask, cpu);
946 	if (target >= nr_cpu_ids) {
947 		dev_err(info->pmu.dev, "Unable to find a suitable CPU\n");
948 		return 0;
949 	}
950 
951 	perf_pmu_migrate_context(&info->pmu, cpu, target);
952 	info->on_cpu = target;
953 	/*
954 	 * CPU HP lock is held so we should be guaranteed that this CPU hasn't yet
955 	 * gone away.
956 	 */
957 	WARN_ON(irq_set_affinity(info->irq, cpumask_of(target)));
958 
959 	return 0;
960 }
961 
962 static __init int cxl_pmu_init(void)
963 {
964 	int rc;
965 
966 	rc = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN,
967 				     "AP_PERF_CXL_PMU_ONLINE",
968 				     cxl_pmu_online_cpu, cxl_pmu_offline_cpu);
969 	if (rc < 0)
970 		return rc;
971 	cxl_pmu_cpuhp_state_num = rc;
972 
973 	rc = cxl_driver_register(&cxl_pmu_driver);
974 	if (rc)
975 		cpuhp_remove_multi_state(cxl_pmu_cpuhp_state_num);
976 
977 	return rc;
978 }
979 
980 static __exit void cxl_pmu_exit(void)
981 {
982 	cxl_driver_unregister(&cxl_pmu_driver);
983 	cpuhp_remove_multi_state(cxl_pmu_cpuhp_state_num);
984 }
985 
986 MODULE_LICENSE("GPL");
987 MODULE_IMPORT_NS(CXL);
988 module_init(cxl_pmu_init);
989 module_exit(cxl_pmu_exit);
990 MODULE_ALIAS_CXL(CXL_DEVICE_PMU);
991