xref: /linux/include/linux/resctrl.h (revision 1e525507)
1 /* SPDX-License-Identifier: GPL-2.0 */
2 #ifndef _RESCTRL_H
3 #define _RESCTRL_H
4 
5 #include <linux/kernel.h>
6 #include <linux/list.h>
7 #include <linux/pid.h>
8 
9 /* CLOSID, RMID value used by the default control group */
10 #define RESCTRL_RESERVED_CLOSID		0
11 #define RESCTRL_RESERVED_RMID		0
12 
13 #define RESCTRL_PICK_ANY_CPU		-1
14 
15 #ifdef CONFIG_PROC_CPU_RESCTRL
16 
17 int proc_resctrl_show(struct seq_file *m,
18 		      struct pid_namespace *ns,
19 		      struct pid *pid,
20 		      struct task_struct *tsk);
21 
22 #endif
23 
24 /* max value for struct rdt_domain's mbps_val */
25 #define MBA_MAX_MBPS   U32_MAX
26 
27 /**
28  * enum resctrl_conf_type - The type of configuration.
29  * @CDP_NONE:	No prioritisation, both code and data are controlled or monitored.
30  * @CDP_CODE:	Configuration applies to instruction fetches.
31  * @CDP_DATA:	Configuration applies to reads and writes.
32  */
33 enum resctrl_conf_type {
34 	CDP_NONE,
35 	CDP_CODE,
36 	CDP_DATA,
37 };
38 
39 #define CDP_NUM_TYPES	(CDP_DATA + 1)
40 
41 /*
42  * Event IDs, the values match those used to program IA32_QM_EVTSEL before
43  * reading IA32_QM_CTR on RDT systems.
44  */
45 enum resctrl_event_id {
46 	QOS_L3_OCCUP_EVENT_ID		= 0x01,
47 	QOS_L3_MBM_TOTAL_EVENT_ID	= 0x02,
48 	QOS_L3_MBM_LOCAL_EVENT_ID	= 0x03,
49 };
50 
51 /**
52  * struct resctrl_staged_config - parsed configuration to be applied
53  * @new_ctrl:		new ctrl value to be loaded
54  * @have_new_ctrl:	whether the user provided new_ctrl is valid
55  */
56 struct resctrl_staged_config {
57 	u32			new_ctrl;
58 	bool			have_new_ctrl;
59 };
60 
61 /**
62  * struct rdt_domain - group of CPUs sharing a resctrl resource
63  * @list:		all instances of this resource
64  * @id:			unique id for this instance
65  * @cpu_mask:		which CPUs share this resource
66  * @rmid_busy_llc:	bitmap of which limbo RMIDs are above threshold
67  * @mbm_total:		saved state for MBM total bandwidth
68  * @mbm_local:		saved state for MBM local bandwidth
69  * @mbm_over:		worker to periodically read MBM h/w counters
70  * @cqm_limbo:		worker to periodically read CQM h/w counters
71  * @mbm_work_cpu:	worker CPU for MBM h/w counters
72  * @cqm_work_cpu:	worker CPU for CQM h/w counters
73  * @plr:		pseudo-locked region (if any) associated with domain
74  * @staged_config:	parsed configuration to be applied
75  * @mbps_val:		When mba_sc is enabled, this holds the array of user
76  *			specified control values for mba_sc in MBps, indexed
77  *			by closid
78  */
79 struct rdt_domain {
80 	struct list_head		list;
81 	int				id;
82 	struct cpumask			cpu_mask;
83 	unsigned long			*rmid_busy_llc;
84 	struct mbm_state		*mbm_total;
85 	struct mbm_state		*mbm_local;
86 	struct delayed_work		mbm_over;
87 	struct delayed_work		cqm_limbo;
88 	int				mbm_work_cpu;
89 	int				cqm_work_cpu;
90 	struct pseudo_lock_region	*plr;
91 	struct resctrl_staged_config	staged_config[CDP_NUM_TYPES];
92 	u32				*mbps_val;
93 };
94 
95 /**
96  * struct resctrl_cache - Cache allocation related data
97  * @cbm_len:		Length of the cache bit mask
98  * @min_cbm_bits:	Minimum number of consecutive bits to be set.
99  *			The value 0 means the architecture can support
100  *			zero CBM.
101  * @shareable_bits:	Bitmask of shareable resource with other
102  *			executing entities
103  * @arch_has_sparse_bitmasks:	True if a bitmask like f00f is valid.
104  * @arch_has_per_cpu_cfg:	True if QOS_CFG register for this cache
105  *				level has CPU scope.
106  */
107 struct resctrl_cache {
108 	unsigned int	cbm_len;
109 	unsigned int	min_cbm_bits;
110 	unsigned int	shareable_bits;
111 	bool		arch_has_sparse_bitmasks;
112 	bool		arch_has_per_cpu_cfg;
113 };
114 
115 /**
116  * enum membw_throttle_mode - System's memory bandwidth throttling mode
117  * @THREAD_THROTTLE_UNDEFINED:	Not relevant to the system
118  * @THREAD_THROTTLE_MAX:	Memory bandwidth is throttled at the core
119  *				always using smallest bandwidth percentage
120  *				assigned to threads, aka "max throttling"
121  * @THREAD_THROTTLE_PER_THREAD:	Memory bandwidth is throttled at the thread
122  */
123 enum membw_throttle_mode {
124 	THREAD_THROTTLE_UNDEFINED = 0,
125 	THREAD_THROTTLE_MAX,
126 	THREAD_THROTTLE_PER_THREAD,
127 };
128 
129 /**
130  * struct resctrl_membw - Memory bandwidth allocation related data
131  * @min_bw:		Minimum memory bandwidth percentage user can request
132  * @bw_gran:		Granularity at which the memory bandwidth is allocated
133  * @delay_linear:	True if memory B/W delay is in linear scale
134  * @arch_needs_linear:	True if we can't configure non-linear resources
135  * @throttle_mode:	Bandwidth throttling mode when threads request
136  *			different memory bandwidths
137  * @mba_sc:		True if MBA software controller(mba_sc) is enabled
138  * @mb_map:		Mapping of memory B/W percentage to memory B/W delay
139  */
140 struct resctrl_membw {
141 	u32				min_bw;
142 	u32				bw_gran;
143 	u32				delay_linear;
144 	bool				arch_needs_linear;
145 	enum membw_throttle_mode	throttle_mode;
146 	bool				mba_sc;
147 	u32				*mb_map;
148 };
149 
150 struct rdt_parse_data;
151 struct resctrl_schema;
152 
153 /**
154  * struct rdt_resource - attributes of a resctrl resource
155  * @rid:		The index of the resource
156  * @alloc_capable:	Is allocation available on this machine
157  * @mon_capable:	Is monitor feature available on this machine
158  * @num_rmid:		Number of RMIDs available
159  * @cache_level:	Which cache level defines scope of this resource
160  * @cache:		Cache allocation related data
161  * @membw:		If the component has bandwidth controls, their properties.
162  * @domains:		RCU list of all domains for this resource
163  * @name:		Name to use in "schemata" file.
164  * @data_width:		Character width of data when displaying
165  * @default_ctrl:	Specifies default cache cbm or memory B/W percent.
166  * @format_str:		Per resource format string to show domain value
167  * @parse_ctrlval:	Per resource function pointer to parse control values
168  * @evt_list:		List of monitoring events
169  * @fflags:		flags to choose base and info files
170  * @cdp_capable:	Is the CDP feature available on this resource
171  */
172 struct rdt_resource {
173 	int			rid;
174 	bool			alloc_capable;
175 	bool			mon_capable;
176 	int			num_rmid;
177 	int			cache_level;
178 	struct resctrl_cache	cache;
179 	struct resctrl_membw	membw;
180 	struct list_head	domains;
181 	char			*name;
182 	int			data_width;
183 	u32			default_ctrl;
184 	const char		*format_str;
185 	int			(*parse_ctrlval)(struct rdt_parse_data *data,
186 						 struct resctrl_schema *s,
187 						 struct rdt_domain *d);
188 	struct list_head	evt_list;
189 	unsigned long		fflags;
190 	bool			cdp_capable;
191 };
192 
193 /**
194  * struct resctrl_schema - configuration abilities of a resource presented to
195  *			   user-space
196  * @list:	Member of resctrl_schema_all.
197  * @name:	The name to use in the "schemata" file.
198  * @conf_type:	Whether this schema is specific to code/data.
199  * @res:	The resource structure exported by the architecture to describe
200  *		the hardware that is configured by this schema.
201  * @num_closid:	The number of closid that can be used with this schema. When
202  *		features like CDP are enabled, this will be lower than the
203  *		hardware supports for the resource.
204  */
205 struct resctrl_schema {
206 	struct list_head		list;
207 	char				name[8];
208 	enum resctrl_conf_type		conf_type;
209 	struct rdt_resource		*res;
210 	u32				num_closid;
211 };
212 
213 /* The number of closid supported by this resource regardless of CDP */
214 u32 resctrl_arch_get_num_closid(struct rdt_resource *r);
215 int resctrl_arch_update_domains(struct rdt_resource *r, u32 closid);
216 
217 /*
218  * Update the ctrl_val and apply this config right now.
219  * Must be called on one of the domain's CPUs.
220  */
221 int resctrl_arch_update_one(struct rdt_resource *r, struct rdt_domain *d,
222 			    u32 closid, enum resctrl_conf_type t, u32 cfg_val);
223 
224 u32 resctrl_arch_get_config(struct rdt_resource *r, struct rdt_domain *d,
225 			    u32 closid, enum resctrl_conf_type type);
226 int resctrl_online_domain(struct rdt_resource *r, struct rdt_domain *d);
227 void resctrl_offline_domain(struct rdt_resource *r, struct rdt_domain *d);
228 void resctrl_online_cpu(unsigned int cpu);
229 void resctrl_offline_cpu(unsigned int cpu);
230 
231 /**
232  * resctrl_arch_rmid_read() - Read the eventid counter corresponding to rmid
233  *			      for this resource and domain.
234  * @r:			resource that the counter should be read from.
235  * @d:			domain that the counter should be read from.
236  * @closid:		closid that matches the rmid. Depending on the architecture, the
237  *			counter may match traffic of both @closid and @rmid, or @rmid
238  *			only.
239  * @rmid:		rmid of the counter to read.
240  * @eventid:		eventid to read, e.g. L3 occupancy.
241  * @val:		result of the counter read in bytes.
242  * @arch_mon_ctx:	An architecture specific value from
243  *			resctrl_arch_mon_ctx_alloc(), for MPAM this identifies
244  *			the hardware monitor allocated for this read request.
245  *
246  * Some architectures need to sleep when first programming some of the counters.
247  * (specifically: arm64's MPAM cache occupancy counters can return 'not ready'
248  *  for a short period of time). Call from a non-migrateable process context on
249  * a CPU that belongs to domain @d. e.g. use smp_call_on_cpu() or
250  * schedule_work_on(). This function can be called with interrupts masked,
251  * e.g. using smp_call_function_any(), but may consistently return an error.
252  *
253  * Return:
254  * 0 on success, or -EIO, -EINVAL etc on error.
255  */
256 int resctrl_arch_rmid_read(struct rdt_resource *r, struct rdt_domain *d,
257 			   u32 closid, u32 rmid, enum resctrl_event_id eventid,
258 			   u64 *val, void *arch_mon_ctx);
259 
260 /**
261  * resctrl_arch_rmid_read_context_check()  - warn about invalid contexts
262  *
263  * When built with CONFIG_DEBUG_ATOMIC_SLEEP generate a warning when
264  * resctrl_arch_rmid_read() is called with preemption disabled.
265  *
266  * The contract with resctrl_arch_rmid_read() is that if interrupts
267  * are unmasked, it can sleep. This allows NOHZ_FULL systems to use an
268  * IPI, (and fail if the call needed to sleep), while most of the time
269  * the work is scheduled, allowing the call to sleep.
270  */
271 static inline void resctrl_arch_rmid_read_context_check(void)
272 {
273 	if (!irqs_disabled())
274 		might_sleep();
275 }
276 
277 /**
278  * resctrl_arch_reset_rmid() - Reset any private state associated with rmid
279  *			       and eventid.
280  * @r:		The domain's resource.
281  * @d:		The rmid's domain.
282  * @closid:	closid that matches the rmid. Depending on the architecture, the
283  *		counter may match traffic of both @closid and @rmid, or @rmid only.
284  * @rmid:	The rmid whose counter values should be reset.
285  * @eventid:	The eventid whose counter values should be reset.
286  *
287  * This can be called from any CPU.
288  */
289 void resctrl_arch_reset_rmid(struct rdt_resource *r, struct rdt_domain *d,
290 			     u32 closid, u32 rmid,
291 			     enum resctrl_event_id eventid);
292 
293 /**
294  * resctrl_arch_reset_rmid_all() - Reset all private state associated with
295  *				   all rmids and eventids.
296  * @r:		The resctrl resource.
297  * @d:		The domain for which all architectural counter state will
298  *		be cleared.
299  *
300  * This can be called from any CPU.
301  */
302 void resctrl_arch_reset_rmid_all(struct rdt_resource *r, struct rdt_domain *d);
303 
304 extern unsigned int resctrl_rmid_realloc_threshold;
305 extern unsigned int resctrl_rmid_realloc_limit;
306 
307 #endif /* _RESCTRL_H */
308