xref: /freebsd/sys/sys/pmc.h (revision a3ff85e11925ba1e25465c8ba633356f405f9214)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2003-2008, Joseph Koshy
5  * Copyright (c) 2007 The FreeBSD Foundation
6  * All rights reserved.
7  *
8  * Portions of this software were developed by A. Joseph Koshy under
9  * sponsorship from the FreeBSD Foundation and Google, Inc.
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  * 1. Redistributions of source code must retain the above copyright
15  *    notice, this list of conditions and the following disclaimer.
16  * 2. Redistributions in binary form must reproduce the above copyright
17  *    notice, this list of conditions and the following disclaimer in the
18  *    documentation and/or other materials provided with the distribution.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
21  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
24  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30  * SUCH DAMAGE.
31  */
32 
33 #ifndef _SYS_PMC_H_
34 #define	_SYS_PMC_H_
35 
36 #include <dev/hwpmc/pmc_events.h>
37 #include <sys/proc.h>
38 #include <sys/counter.h>
39 #include <machine/pmc_mdep.h>
40 #include <machine/profile.h>
41 #ifdef _KERNEL
42 #include <sys/epoch.h>
43 #include <ck_queue.h>
44 #endif
45 
46 #define	PMC_MODULE_NAME		"hwpmc"
47 #define	PMC_NAME_MAX		64 /* HW counter name size */
48 #define	PMC_CLASS_MAX		8  /* max #classes of PMCs per-system */
49 
50 /*
51  * Kernel<->userland API version number [MMmmpppp]
52  *
53  * Major numbers are to be incremented when an incompatible change to
54  * the ABI occurs that older clients will not be able to handle.
55  *
56  * Minor numbers are incremented when a backwards compatible change
57  * occurs that allows older correct programs to run unchanged.  For
58  * example, when support for a new PMC type is added.
59  *
60  * The patch version is incremented for every bug fix.
61  */
62 #define	PMC_VERSION_MAJOR	0x0A
63 #define	PMC_VERSION_MINOR	0x00
64 #define	PMC_VERSION_PATCH	0x0000
65 
66 #define	PMC_VERSION		(PMC_VERSION_MAJOR << 24 |		\
67 	PMC_VERSION_MINOR << 16 | PMC_VERSION_PATCH)
68 
69 #define PMC_CPUID_LEN 64
70 /* cpu model name for pmu lookup */
71 extern char pmc_cpuid[PMC_CPUID_LEN];
72 
73 /*
74  * Kinds of CPUs known.
75  *
76  * We keep track of CPU variants that need to be distinguished in
77  * some way for PMC operations.  CPU names are grouped by manufacturer
78  * and numbered sparsely in order to minimize changes to the ABI involved
79  * when new CPUs are added.
80  *
81  * Please keep the pmc(3) manual page in sync with this list.
82  */
83 #define	__PMC_CPUS()								\
84     __PMC_CPU(AMD_K8,			0x01,	"AMD K8")			\
85     __PMC_CPU(INTEL_CORE,		0x87,	"Intel Core Solo/Duo")		\
86     __PMC_CPU(INTEL_CORE2,		0x88,	"Intel Core2")			\
87     __PMC_CPU(INTEL_CORE2EXTREME,	0x89,	"Intel Core2 Extreme")		\
88     __PMC_CPU(INTEL_ATOM,		0x8A,	"Intel Atom")			\
89     __PMC_CPU(INTEL_COREI7,		0x8B,	"Intel Core i7")		\
90     __PMC_CPU(INTEL_WESTMERE,		0x8C,	"Intel Westmere")		\
91     __PMC_CPU(INTEL_SANDYBRIDGE,	0x8D,	"Intel Sandy Bridge")		\
92     __PMC_CPU(INTEL_IVYBRIDGE,		0x8E,	"Intel Ivy Bridge")		\
93     __PMC_CPU(INTEL_SANDYBRIDGE_XEON,	0x8F,	"Intel Sandy Bridge Xeon")	\
94     __PMC_CPU(INTEL_IVYBRIDGE_XEON,	0x90,	"Intel Ivy Bridge Xeon")	\
95     __PMC_CPU(INTEL_HASWELL,		0x91,	"Intel Haswell")		\
96     __PMC_CPU(INTEL_ATOM_SILVERMONT,	0x92,	"Intel Atom Silvermont")	\
97     __PMC_CPU(INTEL_NEHALEM_EX,		0x93,	"Intel Nehalem Xeon 7500")	\
98     __PMC_CPU(INTEL_WESTMERE_EX,	0x94,	"Intel Westmere Xeon E7")	\
99     __PMC_CPU(INTEL_HASWELL_XEON,	0x95,	"Intel Haswell Xeon E5 v3")	\
100     __PMC_CPU(INTEL_BROADWELL,		0x96,	"Intel Broadwell")		\
101     __PMC_CPU(INTEL_BROADWELL_XEON,	0x97,	"Intel Broadwell Xeon")		\
102     __PMC_CPU(INTEL_SKYLAKE,		0x98,	"Intel Skylake")		\
103     __PMC_CPU(INTEL_SKYLAKE_XEON,	0x99,	"Intel Skylake Xeon")		\
104     __PMC_CPU(INTEL_ATOM_GOLDMONT,	0x9A,	"Intel Atom Goldmont")		\
105     __PMC_CPU(INTEL_ICELAKE,		0x9B,	"Intel Icelake")		\
106     __PMC_CPU(INTEL_ICELAKE_XEON,	0x9C,	"Intel Icelake Xeon")		\
107     __PMC_CPU(INTEL_ALDERLAKE,		0x9D,	"Intel Alderlake")		\
108     __PMC_CPU(INTEL_ATOM_GOLDMONT_P,	0x9E,	"Intel Atom Goldmont Plus")	\
109     __PMC_CPU(INTEL_ATOM_TREMONT,	0x9F,	"Intel Atom Tremont")		\
110     __PMC_CPU(INTEL_EMERALD_RAPIDS,	0xA0,	"Intel Emerald Rapids")		\
111     __PMC_CPU(INTEL_ALDERLAKEN,		0xA1,	"Intel AlderlakeN")		\
112     __PMC_CPU(INTEL_GRANITE_RAPIDS,	0xA2,	"Intel Granite Rapids")		\
113     __PMC_CPU(INTEL_XSCALE,		0x100,	"Intel XScale")			\
114     __PMC_CPU(PPC_7450,			0x300,	"PowerPC MPC7450")		\
115     __PMC_CPU(PPC_E500,			0x340,	"PowerPC e500 Core")		\
116     __PMC_CPU(PPC_970,			0x380,	"IBM PowerPC 970")		\
117     __PMC_CPU(PPC_POWER8,		0x390,	"IBM POWER8")			\
118     __PMC_CPU(GENERIC,			0x400,	"Generic")			\
119     __PMC_CPU(ARMV7_CORTEX_A5,		0x500,	"ARMv7 Cortex A5")		\
120     __PMC_CPU(ARMV7_CORTEX_A7,		0x501,	"ARMv7 Cortex A7")		\
121     __PMC_CPU(ARMV7_CORTEX_A8,		0x502,	"ARMv7 Cortex A8")		\
122     __PMC_CPU(ARMV7_CORTEX_A9,		0x503,	"ARMv7 Cortex A9")		\
123     __PMC_CPU(ARMV7_CORTEX_A15,		0x504,	"ARMv7 Cortex A15")		\
124     __PMC_CPU(ARMV7_CORTEX_A17,		0x505,	"ARMv7 Cortex A17")		\
125     __PMC_CPU(ARMV8_CORTEX_A53,		0x600,	"ARMv8 Cortex A53")		\
126     __PMC_CPU(ARMV8_CORTEX_A57,		0x601,	"ARMv8 Cortex A57")		\
127     __PMC_CPU(ARMV8_CORTEX_A76,		0x602,	"ARMv8 Cortex A76")
128 
129 enum pmc_cputype {
130 #undef	__PMC_CPU
131 #define	__PMC_CPU(S,V,D)	PMC_CPU_##S = V,
132 	__PMC_CPUS()
133 };
134 
135 #define	PMC_CPU_FIRST	PMC_CPU_AMD_K8
136 #define	PMC_CPU_LAST	PMC_CPU_ARMV8_CORTEX_A76
137 
138 /*
139  * Classes of PMCs
140  */
141 #define	__PMC_CLASSES()								\
142     __PMC_CLASS(TSC,		0x00,	"CPU Timestamp counter")		\
143     __PMC_CLASS(K8,		0x02,	"AMD K8 performance counters")		\
144     __PMC_CLASS(IAF,		0x06,	"Intel Core2/Atom, fixed function")	\
145     __PMC_CLASS(IAP,		0x07,	"Intel Core...Atom, programmable")	\
146     __PMC_CLASS(UCF,		0x08,	"Intel Uncore fixed function")		\
147     __PMC_CLASS(UCP,		0x09,	"Intel Uncore programmable")		\
148     __PMC_CLASS(XSCALE,		0x0A,	"Intel XScale counters")		\
149     __PMC_CLASS(PPC7450,	0x0D,	"Motorola MPC7450 class")		\
150     __PMC_CLASS(PPC970,		0x0E,	"IBM PowerPC 970 class")		\
151     __PMC_CLASS(SOFT,		0x0F,	"Software events")			\
152     __PMC_CLASS(ARMV7,		0x10,	"ARMv7")				\
153     __PMC_CLASS(ARMV8,		0x11,	"ARMv8")				\
154     __PMC_CLASS(E500,		0x13,	"Freescale e500 class")			\
155     __PMC_CLASS(POWER8,		0x15,	"IBM POWER8 class")			\
156     __PMC_CLASS(DMC620_PMU_CD2,	0x16,	"ARM DMC620 Memory Controller PMU CLKDIV2") \
157     __PMC_CLASS(DMC620_PMU_C,	0x17,	"ARM DMC620 Memory Controller PMU CLK")	\
158     __PMC_CLASS(CMN600_PMU,	0x18,	"Arm CoreLink CMN600 Coherent Mesh Network PMU")
159 
160 enum pmc_class {
161 #undef  __PMC_CLASS
162 #define	__PMC_CLASS(S,V,D)	PMC_CLASS_##S = V,
163 	__PMC_CLASSES()
164 };
165 
166 #define	PMC_CLASS_FIRST	PMC_CLASS_TSC
167 #define	PMC_CLASS_LAST	PMC_CLASS_CMN600_PMU
168 
169 /*
170  * A PMC can be in the following states:
171  *
172  * Hardware states:
173  *   DISABLED   -- administratively prohibited from being used.
174  *   FREE       -- HW available for use
175  * Software states:
176  *   ALLOCATED  -- allocated
177  *   STOPPED    -- allocated, but not counting events
178  *   RUNNING    -- allocated, and in operation; 'pm_runcount'
179  *                 holds the number of CPUs using this PMC at
180  *                 a given instant
181  *   DELETED    -- being destroyed
182  */
183 
184 #define	__PMC_HWSTATES()			\
185 	__PMC_STATE(DISABLED)			\
186 	__PMC_STATE(FREE)
187 
188 #define	__PMC_SWSTATES()			\
189 	__PMC_STATE(ALLOCATED)			\
190 	__PMC_STATE(STOPPED)			\
191 	__PMC_STATE(RUNNING)			\
192 	__PMC_STATE(DELETED)
193 
194 #define	__PMC_STATES()				\
195 	__PMC_HWSTATES()			\
196 	__PMC_SWSTATES()
197 
198 enum pmc_state {
199 #undef	__PMC_STATE
200 #define	__PMC_STATE(S)	PMC_STATE_##S,
201 	__PMC_STATES()
202 	__PMC_STATE(MAX)
203 };
204 
205 #define	PMC_STATE_FIRST	PMC_STATE_DISABLED
206 #define	PMC_STATE_LAST	PMC_STATE_DELETED
207 
208 /*
209  * An allocated PMC may used as a 'global' counter or as a
210  * 'thread-private' one.  Each such mode of use can be in either
211  * statistical sampling mode or in counting mode.  Thus a PMC in use
212  *
213  * SS i.e., SYSTEM STATISTICAL  -- system-wide statistical profiling
214  * SC i.e., SYSTEM COUNTER      -- system-wide counting mode
215  * TS i.e., THREAD STATISTICAL  -- thread virtual, statistical profiling
216  * TC i.e., THREAD COUNTER      -- thread virtual, counting mode
217  *
218  * Statistical profiling modes rely on the PMC periodically delivering
219  * a interrupt to the CPU (when the configured number of events have
220  * been measured), so the PMC must have the ability to generate
221  * interrupts.
222  *
223  * In counting modes, the PMC counts its configured events, with the
224  * value of the PMC being read whenever needed by its owner process.
225  *
226  * The thread specific modes "virtualize" the PMCs -- the PMCs appear
227  * to be thread private and count events only when the profiled thread
228  * actually executes on the CPU.
229  *
230  * The system-wide "global" modes keep the PMCs running all the time
231  * and are used to measure the behaviour of the whole system.
232  */
233 
234 #define	__PMC_MODES()				\
235 	__PMC_MODE(SS,	0)			\
236 	__PMC_MODE(SC,	1)			\
237 	__PMC_MODE(TS,	2)			\
238 	__PMC_MODE(TC,	3)
239 
240 enum pmc_mode {
241 #undef	__PMC_MODE
242 #define	__PMC_MODE(M,N)	PMC_MODE_##M = N,
243 	__PMC_MODES()
244 };
245 
246 #define	PMC_MODE_FIRST	PMC_MODE_SS
247 #define	PMC_MODE_LAST	PMC_MODE_TC
248 
249 #define	PMC_IS_COUNTING_MODE(mode)				\
250 	((mode) == PMC_MODE_SC || (mode) == PMC_MODE_TC)
251 #define	PMC_IS_SYSTEM_MODE(mode)				\
252 	((mode) == PMC_MODE_SS || (mode) == PMC_MODE_SC)
253 #define	PMC_IS_SAMPLING_MODE(mode)				\
254 	((mode) == PMC_MODE_SS || (mode) == PMC_MODE_TS)
255 #define	PMC_IS_VIRTUAL_MODE(mode)				\
256 	((mode) == PMC_MODE_TS || (mode) == PMC_MODE_TC)
257 
258 /*
259  * PMC row disposition
260  */
261 
262 #define	__PMC_DISPOSITIONS(N)					\
263 	__PMC_DISP(STANDALONE)	/* global/disabled counters */	\
264 	__PMC_DISP(FREE)	/* free/available */		\
265 	__PMC_DISP(THREAD)	/* thread-virtual PMCs */	\
266 	__PMC_DISP(UNKNOWN)	/* sentinel */
267 
268 enum pmc_disp {
269 #undef	__PMC_DISP
270 #define	__PMC_DISP(D)	PMC_DISP_##D ,
271 	__PMC_DISPOSITIONS()
272 };
273 
274 #define	PMC_DISP_FIRST	PMC_DISP_STANDALONE
275 #define	PMC_DISP_LAST	PMC_DISP_THREAD
276 
277 /*
278  * Counter capabilities
279  *
280  * __PMC_CAPS(NAME, VALUE, DESCRIPTION)
281  */
282 
283 #define	__PMC_CAPS()							\
284 	__PMC_CAP(INTERRUPT,	0, "generate interrupts")		\
285 	__PMC_CAP(USER,		1, "count user-mode events")		\
286 	__PMC_CAP(SYSTEM,	2, "count system-mode events")		\
287 	__PMC_CAP(EDGE,		3, "do edge detection of events")	\
288 	__PMC_CAP(THRESHOLD,	4, "ignore events below a threshold")	\
289 	__PMC_CAP(READ,		5, "read PMC counter")			\
290 	__PMC_CAP(WRITE,	6, "reprogram PMC counter")		\
291 	__PMC_CAP(INVERT,	7, "invert comparison sense")		\
292 	__PMC_CAP(QUALIFIER,	8, "further qualify monitored events")	\
293 	__PMC_CAP(PRECISE,	9, "perform precise sampling")		\
294 	__PMC_CAP(TAGGING,	10, "tag upstream events")		\
295 	__PMC_CAP(CASCADE,	11, "cascade counters")			\
296 	__PMC_CAP(SYSWIDE,	12, "system wide counter")		\
297 	__PMC_CAP(DOMWIDE,	13, "NUMA domain wide counter")
298 
299 enum pmc_caps
300 {
301 #undef	__PMC_CAP
302 #define	__PMC_CAP(NAME, VALUE, DESCR)	PMC_CAP_##NAME = (1 << VALUE) ,
303 	__PMC_CAPS()
304 };
305 
306 #define	PMC_CAP_FIRST		PMC_CAP_INTERRUPT
307 #define	PMC_CAP_LAST		PMC_CAP_DOMWIDE
308 
309 /*
310  * PMC Event Numbers
311  *
312  * These are generated from the definitions in "dev/hwpmc/pmc_events.h".
313  */
314 
315 enum pmc_event {
316 #undef	__PMC_EV
317 #undef	__PMC_EV_BLOCK
318 #define	__PMC_EV_BLOCK(C,V)	PMC_EV_ ## C ## __BLOCK_START = (V) - 1 ,
319 #define	__PMC_EV(C,N)		PMC_EV_ ## C ## _ ## N ,
320 	__PMC_EVENTS()
321 };
322 
323 /*
324  * PMC SYSCALL INTERFACE
325  */
326 
327 /*
328  * "PMC_OPS" -- these are the commands recognized by the kernel
329  * module, and are used when performing a system call from userland.
330  */
331 #define	__PMC_OPS()							\
332 	__PMC_OP(CONFIGURELOG, "Set log file")				\
333 	__PMC_OP(FLUSHLOG, "Flush log file")				\
334 	__PMC_OP(GETCPUINFO, "Get system CPU information")		\
335 	__PMC_OP(GETDRIVERSTATS, "Get driver statistics")		\
336 	__PMC_OP(GETMODULEVERSION, "Get module version")		\
337 	__PMC_OP(GETPMCINFO, "Get per-cpu PMC information")		\
338 	__PMC_OP(PMCADMIN, "Set PMC state")				\
339 	__PMC_OP(PMCALLOCATE, "Allocate and configure a PMC")		\
340 	__PMC_OP(PMCATTACH, "Attach a PMC to a process")		\
341 	__PMC_OP(PMCDETACH, "Detach a PMC from a process")		\
342 	__PMC_OP(PMCGETMSR, "Get a PMC's hardware address")		\
343 	__PMC_OP(PMCRELEASE, "Release a PMC")				\
344 	__PMC_OP(PMCRW, "Read/Set a PMC")				\
345 	__PMC_OP(PMCSETCOUNT, "Set initial count/sampling rate")	\
346 	__PMC_OP(PMCSTART, "Start a PMC")				\
347 	__PMC_OP(PMCSTOP, "Stop a PMC")					\
348 	__PMC_OP(WRITELOG, "Write a cookie to the log file")		\
349 	__PMC_OP(CLOSELOG, "Close log file")				\
350 	__PMC_OP(GETDYNEVENTINFO, "Get dynamic events list")
351 
352 enum pmc_ops {
353 #undef	__PMC_OP
354 #define	__PMC_OP(N, D)	PMC_OP_##N,
355 	__PMC_OPS()
356 };
357 
358 /*
359  * Flags used in operations on PMCs.
360  */
361 
362 #define	PMC_F_UNUSED1		0x00000001 /* unused */
363 #define	PMC_F_DESCENDANTS	0x00000002 /*OP ALLOCATE track descendants */
364 #define	PMC_F_LOG_PROCCSW	0x00000004 /*OP ALLOCATE track ctx switches */
365 #define	PMC_F_LOG_PROCEXIT	0x00000008 /*OP ALLOCATE log proc exits */
366 #define	PMC_F_NEWVALUE		0x00000010 /*OP RW write new value */
367 #define	PMC_F_OLDVALUE		0x00000020 /*OP RW get old value */
368 
369 /* V2 API */
370 #define	PMC_F_CALLCHAIN		0x00000080 /*OP ALLOCATE capture callchains */
371 #define	PMC_F_USERCALLCHAIN	0x00000100 /*OP ALLOCATE use userspace stack */
372 
373 /* V10 API */
374 #define	PMC_F_EV_PMU		0x00000200 /*
375 					    * OP ALLOCATE: pm_ev has special
376 					    * userspace meaning; counter
377 					    * configuration is communicated
378 					    * through class-dependent fields
379 					    */
380 
381 /* internal flags */
382 #define	PMC_F_ATTACHED_TO_OWNER	0x00010000 /*attached to owner*/
383 #define	PMC_F_NEEDS_LOGFILE	0x00020000 /*needs log file */
384 #define	PMC_F_ATTACH_DONE	0x00040000 /*attached at least once */
385 
386 #define	PMC_CALLCHAIN_DEPTH_MAX	512
387 
388 #define	PMC_CC_F_USERSPACE	0x01	   /*userspace callchain*/
389 
390 /*
391  * Cookies used to denote allocated PMCs, and the values of PMCs.
392  */
393 
394 typedef uint32_t	pmc_id_t;
395 typedef uint64_t	pmc_value_t;
396 
397 #define	PMC_ID_INVALID		(~ (pmc_id_t) 0)
398 
399 /*
400  * PMC IDs have the following format:
401  *
402  * +-----------------------+-------+-----------+
403  * |   CPU      | PMC MODE | CLASS | ROW INDEX |
404  * +-----------------------+-------+-----------+
405  *
406  * where CPU is 12 bits, MODE 4, CLASS 8, and ROW INDEX 8  Field 'CPU'
407  * is set to the requested CPU for system-wide PMCs or PMC_CPU_ANY for
408  * process-mode PMCs.  Field 'PMC MODE' is the allocated PMC mode.
409  * Field 'PMC CLASS' is the class of the PMC.  Field 'ROW INDEX' is the
410  * row index for the PMC.
411  *
412  * The 'ROW INDEX' ranges over 0..NWPMCS where NHWPMCS is the total
413  * number of hardware PMCs on this cpu.
414  */
415 
416 #define	PMC_ID_TO_ROWINDEX(ID)	((ID) & 0xFF)
417 #define	PMC_ID_TO_CLASS(ID)	(((ID) & 0xFF00) >> 8)
418 #define	PMC_ID_TO_MODE(ID)	(((ID) & 0xF0000) >> 16)
419 #define	PMC_ID_TO_CPU(ID)	(((ID) & 0xFFF00000) >> 20)
420 #define	PMC_ID_MAKE_ID(CPU,MODE,CLASS,ROWINDEX)			\
421 	((((CPU) & 0xFFF) << 20) | (((MODE) & 0xF) << 16) |	\
422 	(((CLASS) & 0xFF) << 8) | ((ROWINDEX) & 0xFF))
423 
424 /*
425  * Data structures for system calls supported by the pmc driver.
426  */
427 
428 /*
429  * OP PMCALLOCATE
430  *
431  * Allocate a PMC on the named CPU.
432  */
433 
434 #define	PMC_CPU_ANY	~0
435 
436 struct pmc_op_pmcallocate {
437 	uint32_t	pm_caps;	/* PMC_CAP_* */
438 	uint32_t	pm_cpu;		/* CPU number or PMC_CPU_ANY */
439 	enum pmc_class	pm_class;	/* class of PMC desired */
440 	enum pmc_event	pm_ev;		/* [enum pmc_event] desired */
441 	uint32_t	pm_flags;	/* additional modifiers PMC_F_* */
442 	enum pmc_mode	pm_mode;	/* desired mode */
443 	pmc_id_t	pm_pmcid;	/* [return] process pmc id */
444 	pmc_value_t	pm_count;	/* initial/sample count */
445 
446 	union pmc_md_op_pmcallocate pm_md; /* MD layer extensions */
447 };
448 
449 /*
450  * OP PMCADMIN
451  *
452  * Set the administrative state (i.e., whether enabled or disabled) of
453  * a PMC 'pm_pmc' on CPU 'pm_cpu'.  Note that 'pm_pmc' specifies an
454  * absolute PMC number and need not have been first allocated by the
455  * calling process.
456  */
457 
458 struct pmc_op_pmcadmin {
459 	int		pm_cpu;		/* CPU# */
460 	uint32_t	pm_flags;	/* flags */
461 	int		pm_pmc;         /* PMC# */
462 	enum pmc_state  pm_state;	/* desired state */
463 };
464 
465 /*
466  * OP PMCATTACH / OP PMCDETACH
467  *
468  * Attach/detach a PMC and a process.
469  */
470 
471 struct pmc_op_pmcattach {
472 	pmc_id_t	pm_pmc;		/* PMC to attach to */
473 	pid_t		pm_pid;		/* target process */
474 };
475 
476 /*
477  * OP PMCSETCOUNT
478  *
479  * Set the sampling rate (i.e., the reload count) for statistical counters.
480  * 'pm_pmcid' need to have been previously allocated using PMCALLOCATE.
481  */
482 
483 struct pmc_op_pmcsetcount {
484 	pmc_value_t	pm_count;	/* initial/sample count */
485 	pmc_id_t	pm_pmcid;	/* PMC id to set */
486 };
487 
488 /*
489  * OP PMCRW
490  *
491  * Read the value of a PMC named by 'pm_pmcid'.  'pm_pmcid' needs
492  * to have been previously allocated using PMCALLOCATE.
493  */
494 
495 struct pmc_op_pmcrw {
496 	uint32_t	pm_flags;	/* PMC_F_{OLD,NEW}VALUE*/
497 	pmc_id_t	pm_pmcid;	/* pmc id */
498 	pmc_value_t	pm_value;	/* new&returned value */
499 };
500 
501 /*
502  * OP GETPMCINFO
503  *
504  * retrieve PMC state for a named CPU.  The caller is expected to
505  * allocate 'npmc' * 'struct pmc_info' bytes of space for the return
506  * values.
507  */
508 
509 struct pmc_info {
510 	char		pm_name[PMC_NAME_MAX]; /* pmc name */
511 	enum pmc_class	pm_class;	/* enum pmc_class */
512 	int		pm_enabled;	/* whether enabled */
513 	enum pmc_disp	pm_rowdisp;	/* FREE, THREAD or STANDLONE */
514 	pid_t		pm_ownerpid;	/* owner, or -1 */
515 	enum pmc_mode	pm_mode;	/* current mode [enum pmc_mode] */
516 	enum pmc_event	pm_event;	/* current event */
517 	uint32_t	pm_flags;	/* current flags */
518 	pmc_value_t	pm_reloadcount;	/* sampling counters only */
519 };
520 
521 struct pmc_op_getpmcinfo {
522 	int32_t		pm_cpu;		/* 0 <= cpu < mp_maxid */
523 	struct pmc_info	pm_pmcs[];	/* space for 'npmc' structures */
524 };
525 
526 /*
527  * OP GETCPUINFO
528  *
529  * Retrieve system CPU information.
530  */
531 
532 struct pmc_classinfo {
533 	enum pmc_class	pm_class;	/* class id */
534 	uint32_t	pm_caps;	/* counter capabilities */
535 	uint32_t	pm_width;	/* width of the PMC */
536 	uint32_t	pm_num;		/* number of PMCs in class */
537 };
538 
539 struct pmc_op_getcpuinfo {
540 	enum pmc_cputype pm_cputype; /* what kind of CPU */
541 	uint32_t	pm_ncpu;    /* max CPU number */
542 	uint32_t	pm_npmc;    /* #PMCs per CPU */
543 	uint32_t	pm_nclass;  /* #classes of PMCs */
544 	struct pmc_classinfo  pm_classes[PMC_CLASS_MAX];
545 };
546 
547 /*
548  * OP CONFIGURELOG
549  *
550  * Configure a log file for writing system-wide statistics to.
551  */
552 
553 struct pmc_op_configurelog {
554 	int		pm_flags;
555 	int		pm_logfd;   /* logfile fd (or -1) */
556 };
557 
558 /*
559  * OP GETDRIVERSTATS
560  *
561  * Retrieve pmc(4) driver-wide statistics.
562  */
563 #ifdef _KERNEL
564 struct pmc_driverstats {
565 	counter_u64_t	pm_intr_ignored;	/* #interrupts ignored */
566 	counter_u64_t	pm_intr_processed;	/* #interrupts processed */
567 	counter_u64_t	pm_intr_bufferfull;	/* #interrupts with ENOSPC */
568 	counter_u64_t	pm_syscalls;		/* #syscalls */
569 	counter_u64_t	pm_syscall_errors;	/* #syscalls with errors */
570 	counter_u64_t	pm_buffer_requests;	/* #buffer requests */
571 	counter_u64_t	pm_buffer_requests_failed; /* #failed buffer requests */
572 	counter_u64_t	pm_log_sweeps;		/* #sample buffer processing
573 						   passes */
574 	counter_u64_t	pm_merges;		/* merged k+u */
575 	counter_u64_t	pm_overwrites;		/* UR overwrites */
576 };
577 #endif
578 
579 struct pmc_op_getdriverstats {
580 	unsigned int	pm_intr_ignored;	/* #interrupts ignored */
581 	unsigned int	pm_intr_processed;	/* #interrupts processed */
582 	unsigned int	pm_intr_bufferfull;	/* #interrupts with ENOSPC */
583 	unsigned int	pm_syscalls;		/* #syscalls */
584 	unsigned int	pm_syscall_errors;	/* #syscalls with errors */
585 	unsigned int	pm_buffer_requests;	/* #buffer requests */
586 	unsigned int	pm_buffer_requests_failed; /* #failed buffer requests */
587 	unsigned int	pm_log_sweeps;		/* #sample buffer processing
588 						   passes */
589 };
590 
591 /*
592  * OP RELEASE / OP START / OP STOP
593  *
594  * Simple operations on a PMC id.
595  */
596 
597 struct pmc_op_simple {
598 	pmc_id_t	pm_pmcid;
599 };
600 
601 /*
602  * OP WRITELOG
603  *
604  * Flush the current log buffer and write 4 bytes of user data to it.
605  */
606 
607 struct pmc_op_writelog {
608 	uint32_t	pm_userdata;
609 };
610 
611 /*
612  * OP GETMSR
613  *
614  * Retrieve the machine specific address associated with the allocated
615  * PMC.  This number can be used subsequently with a read-performance-counter
616  * instruction.
617  */
618 
619 struct pmc_op_getmsr {
620 	uint32_t	pm_msr;		/* machine specific address */
621 	pmc_id_t	pm_pmcid;	/* allocated pmc id */
622 };
623 
624 /*
625  * OP GETDYNEVENTINFO
626  *
627  * Retrieve a PMC dynamic class events list.
628  */
629 
630 struct pmc_dyn_event_descr {
631 	char		pm_ev_name[PMC_NAME_MAX];
632 	enum pmc_event	pm_ev_code;
633 };
634 
635 struct pmc_op_getdyneventinfo {
636 	enum pmc_class			pm_class;
637 	unsigned int			pm_nevent;
638 	struct pmc_dyn_event_descr	pm_events[PMC_EV_DYN_COUNT];
639 };
640 
641 #ifdef _KERNEL
642 
643 #include <sys/malloc.h>
644 #include <sys/sysctl.h>
645 #include <sys/_cpuset.h>
646 
647 #include <machine/frame.h>
648 
649 #define	PMC_HASH_SIZE				1024
650 #define	PMC_MTXPOOL_SIZE			2048
651 #define	PMC_LOG_BUFFER_SIZE			256
652 #define	PMC_NLOGBUFFERS_PCPU			32
653 #define	PMC_NSAMPLES				256
654 #define	PMC_CALLCHAIN_DEPTH			128
655 #define	PMC_THREADLIST_MAX			128
656 
657 #define PMC_SYSCTL_NAME_PREFIX "kern." PMC_MODULE_NAME "."
658 
659 /*
660  * Locking keys
661  *
662  * (b) - pmc_bufferlist_mtx (spin lock)
663  * (k) - pmc_kthread_mtx (sleep lock)
664  * (o) - po->po_mtx (spin lock)
665  * (g) - global_epoch_preempt (epoch)
666  * (p) - pmc_sx (sx)
667  */
668 
669 /*
670  * PMC commands
671  */
672 
673 struct pmc_syscall_args {
674 	register_t	pmop_code;	/* one of PMC_OP_* */
675 	void		*pmop_data;	/* syscall parameter */
676 };
677 
678 /*
679  * Interface to processor specific s1tuff
680  */
681 
682 /*
683  * struct pmc_descr
684  *
685  * Machine independent (i.e., the common parts) of a human readable
686  * PMC description.
687  */
688 
689 struct pmc_descr {
690 	char		pd_name[PMC_NAME_MAX]; /* name */
691 	uint32_t	pd_caps;	/* capabilities */
692 	enum pmc_class	pd_class;	/* class of the PMC */
693 	uint32_t	pd_width;	/* width in bits */
694 };
695 
696 /*
697  * struct pmc_target
698  *
699  * This structure records all the target processes associated with a
700  * PMC.
701  */
702 
703 struct pmc_target {
704 	LIST_ENTRY(pmc_target)	pt_next;
705 	struct pmc_process	*pt_process; /* target descriptor */
706 };
707 
708 /*
709  * struct pmc
710  *
711  * Describes each allocated PMC.
712  *
713  * Each PMC has precisely one owner, namely the process that allocated
714  * the PMC.
715  *
716  * A PMC may be attached to multiple target processes.  The
717  * 'pm_targets' field links all the target processes being monitored
718  * by this PMC.
719  *
720  * The 'pm_savedvalue' field is protected by a mutex.
721  *
722  * On a multi-cpu machine, multiple target threads associated with a
723  * process-virtual PMC could be concurrently executing on different
724  * CPUs.  The 'pm_runcount' field is atomically incremented every time
725  * the PMC gets scheduled on a CPU and atomically decremented when it
726  * get descheduled.  Deletion of a PMC is only permitted when this
727  * field is '0'.
728  *
729  */
730 struct pmc_pcpu_state {
731 	uint32_t pps_overflowcnt;	/* count overflow interrupts */
732 	uint8_t pps_stalled;
733 	uint8_t pps_cpustate;
734 } __aligned(CACHE_LINE_SIZE);
735 struct pmc {
736 	LIST_HEAD(,pmc_target)	pm_targets;	/* list of target processes */
737 	LIST_ENTRY(pmc)		pm_next;	/* owner's list */
738 
739 	/*
740 	 * System-wide PMCs are allocated on a CPU and are not moved
741 	 * around.  For system-wide PMCs we record the CPU the PMC was
742 	 * allocated on in the 'CPU' field of the pmc ID.
743 	 *
744 	 * Virtual PMCs run on whichever CPU is currently executing
745 	 * their targets' threads.  For these PMCs we need to save
746 	 * their current PMC counter values when they are taken off
747 	 * CPU.
748 	 */
749 
750 	union {
751 		pmc_value_t	pm_savedvalue;	/* Virtual PMCS */
752 	} pm_gv;
753 
754 	/*
755 	 * For sampling mode PMCs, we keep track of the PMC's "reload
756 	 * count", which is the counter value to be loaded in when
757 	 * arming the PMC for the next counting session.  For counting
758 	 * modes on PMCs that are read-only (e.g., the x86 TSC), we
759 	 * keep track of the initial value at the start of
760 	 * counting-mode operation.
761 	 */
762 
763 	union {
764 		pmc_value_t	pm_reloadcount;	/* sampling PMC modes */
765 		pmc_value_t	pm_initial;	/* counting PMC modes */
766 	} pm_sc;
767 
768 	struct pmc_pcpu_state *pm_pcpu_state;
769 	volatile cpuset_t pm_cpustate;	/* CPUs where PMC should be active */
770 	uint32_t	pm_caps;	/* PMC capabilities */
771 	enum pmc_event	pm_event;	/* event being measured */
772 	uint32_t	pm_flags;	/* additional flags PMC_F_... */
773 	struct pmc_owner *pm_owner;	/* owner thread state */
774 	counter_u64_t		pm_runcount;	/* #cpus currently on */
775 	enum pmc_state	pm_state;	/* current PMC state */
776 
777 	/*
778 	 * The PMC ID field encodes the row-index for the PMC, its
779 	 * mode, class and the CPU# associated with the PMC.
780 	 */
781 
782 	pmc_id_t	pm_id;		/* allocated PMC id */
783 	enum pmc_class pm_class;
784 
785 	/* md extensions */
786 	union pmc_md_pmc	pm_md;
787 };
788 
789 /*
790  * Accessor macros for 'struct pmc'
791  */
792 
793 #define	PMC_TO_MODE(P)		PMC_ID_TO_MODE((P)->pm_id)
794 #define	PMC_TO_CLASS(P)		PMC_ID_TO_CLASS((P)->pm_id)
795 #define	PMC_TO_ROWINDEX(P)	PMC_ID_TO_ROWINDEX((P)->pm_id)
796 #define	PMC_TO_CPU(P)		PMC_ID_TO_CPU((P)->pm_id)
797 
798 /*
799  * struct pmc_threadpmcstate
800  *
801  * Record per-PMC, per-thread state.
802  */
803 struct pmc_threadpmcstate {
804 	pmc_value_t	pt_pmcval;	/* per-thread reload count */
805 };
806 
807 /*
808  * struct pmc_thread
809  *
810  * Record a 'target' thread being profiled.
811  */
812 struct pmc_thread {
813 	LIST_ENTRY(pmc_thread) pt_next;		/* linked list */
814 	struct thread	*pt_td;			/* target thread */
815 	struct pmc_threadpmcstate pt_pmcs[];	/* per-PMC state */
816 };
817 
818 /*
819  * struct pmc_process
820  *
821  * Record a 'target' process being profiled.
822  *
823  * The target process being profiled could be different from the owner
824  * process which allocated the PMCs.  Each target process descriptor
825  * is associated with NHWPMC 'struct pmc *' pointers.  Each PMC at a
826  * given hardware row-index 'n' will use slot 'n' of the 'pp_pmcs[]'
827  * array.  The size of this structure is thus PMC architecture
828  * dependent.
829  *
830  */
831 
832 struct pmc_targetstate {
833 	struct pmc	*pp_pmc;   /* target PMC */
834 	pmc_value_t	pp_pmcval; /* per-process value */
835 };
836 
837 struct pmc_process {
838 	LIST_ENTRY(pmc_process) pp_next;	/* hash chain */
839 	LIST_HEAD(,pmc_thread) pp_tds;		/* list of threads */
840 	struct mtx	*pp_tdslock;		/* lock on pp_tds thread list */
841 	int		pp_refcnt;		/* reference count */
842 	uint32_t	pp_flags;		/* flags PMC_PP_* */
843 	struct proc	*pp_proc;		/* target process */
844 	struct pmc_targetstate pp_pmcs[];       /* NHWPMCs */
845 };
846 
847 #define	PMC_PP_ENABLE_MSR_ACCESS	0x00000001
848 
849 /*
850  * struct pmc_owner
851  *
852  * We associate a PMC with an 'owner' process.
853  *
854  * A process can be associated with 0..NCPUS*NHWPMC PMCs during its
855  * lifetime, where NCPUS is the numbers of CPUS in the system and
856  * NHWPMC is the number of hardware PMCs per CPU.  These are
857  * maintained in the list headed by the 'po_pmcs' to save on space.
858  *
859  */
860 
861 struct pmc_owner  {
862 	LIST_ENTRY(pmc_owner)	po_next;	/* hash chain */
863 	CK_LIST_ENTRY(pmc_owner)	po_ssnext;	/* (g/p) list of SS PMC owners */
864 	LIST_HEAD(, pmc)	po_pmcs;	/* owned PMC list */
865 	TAILQ_HEAD(, pmclog_buffer) po_logbuffers; /* (o) logbuffer list */
866 	struct mtx		po_mtx;		/* spin lock for (o) */
867 	struct proc		*po_owner;	/* owner proc */
868 	uint32_t		po_flags;	/* (k) flags PMC_PO_* */
869 	struct proc		*po_kthread;	/* (k) helper kthread */
870 	struct file		*po_file;	/* file reference */
871 	int			po_error;	/* recorded error */
872 	short			po_sscount;	/* # SS PMCs owned */
873 	short			po_logprocmaps;	/* global mappings done */
874 	struct pmclog_buffer	*po_curbuf[MAXCPU];	/* current log buffer */
875 };
876 
877 #define	PMC_PO_OWNS_LOGFILE		0x00000001 /* has a log file */
878 #define	PMC_PO_SHUTDOWN			0x00000010 /* in the process of shutdown */
879 #define	PMC_PO_INITIAL_MAPPINGS_DONE	0x00000020
880 
881 /*
882  * struct pmc_hw -- describe the state of the PMC hardware
883  *
884  * When in use, a HW PMC is associated with one allocated 'struct pmc'
885  * pointed to by field 'phw_pmc'.  When inactive, this field is NULL.
886  *
887  * On an SMP box, one or more HW PMC's in process virtual mode with
888  * the same 'phw_pmc' could be executing on different CPUs.  In order
889  * to handle this case correctly, we need to ensure that only
890  * incremental counts get added to the saved value in the associated
891  * 'struct pmc'.  The 'phw_save' field is used to keep the saved PMC
892  * value at the time the hardware is started during this context
893  * switch (i.e., the difference between the new (hardware) count and
894  * the saved count is atomically added to the count field in 'struct
895  * pmc' at context switch time).
896  *
897  */
898 
899 struct pmc_hw {
900 	uint32_t	phw_state;	/* see PHW_* macros below */
901 	struct pmc	*phw_pmc;	/* current thread PMC */
902 };
903 
904 #define	PMC_PHW_RI_MASK		0x000000FF
905 #define	PMC_PHW_CPU_SHIFT	8
906 #define	PMC_PHW_CPU_MASK	0x0000FF00
907 #define	PMC_PHW_FLAGS_SHIFT	16
908 #define	PMC_PHW_FLAGS_MASK	0xFFFF0000
909 
910 #define	PMC_PHW_INDEX_TO_STATE(ri)	((ri) & PMC_PHW_RI_MASK)
911 #define	PMC_PHW_STATE_TO_INDEX(state)	((state) & PMC_PHW_RI_MASK)
912 #define	PMC_PHW_CPU_TO_STATE(cpu)	(((cpu) << PMC_PHW_CPU_SHIFT) & \
913 	PMC_PHW_CPU_MASK)
914 #define	PMC_PHW_STATE_TO_CPU(state)	(((state) & PMC_PHW_CPU_MASK) >> \
915 	PMC_PHW_CPU_SHIFT)
916 #define	PMC_PHW_FLAGS_TO_STATE(flags)	(((flags) << PMC_PHW_FLAGS_SHIFT) & \
917 	PMC_PHW_FLAGS_MASK)
918 #define	PMC_PHW_STATE_TO_FLAGS(state)	(((state) & PMC_PHW_FLAGS_MASK) >> \
919 	PMC_PHW_FLAGS_SHIFT)
920 #define	PMC_PHW_FLAG_IS_ENABLED		(PMC_PHW_FLAGS_TO_STATE(0x01))
921 #define	PMC_PHW_FLAG_IS_SHAREABLE	(PMC_PHW_FLAGS_TO_STATE(0x02))
922 
923 /*
924  * struct pmc_sample
925  *
926  * Space for N (tunable) PC samples and associated control data.
927  */
928 
929 struct pmc_sample {
930 	uint16_t		ps_nsamples;	/* callchain depth */
931 	uint16_t		ps_nsamples_actual;
932 	uint16_t		ps_cpu;		/* cpu number */
933 	uint16_t		ps_flags;	/* other flags */
934 	lwpid_t			ps_tid;		/* thread id */
935 	pid_t			ps_pid;		/* process PID or -1 */
936 	int		ps_ticks; /* ticks at sample time */
937 	/* pad */
938 	struct thread		*ps_td;		/* which thread */
939 	struct pmc		*ps_pmc;	/* interrupting PMC */
940 	uintptr_t		*ps_pc;		/* (const) callchain start */
941 	uint64_t		ps_tsc;		/* tsc value */
942 };
943 
944 #define 	PMC_SAMPLE_FREE		((uint16_t) 0)
945 #define 	PMC_USER_CALLCHAIN_PENDING	((uint16_t) 0xFFFF)
946 
947 struct pmc_samplebuffer {
948 	volatile uint64_t		ps_prodidx; /* producer index */
949 	volatile uint64_t		ps_considx; /* consumer index */
950 	uintptr_t		*ps_callchains;	/* all saved call chains */
951 	struct pmc_sample	ps_samples[];	/* array of sample entries */
952 };
953 
954 #define PMC_CONS_SAMPLE(psb)					\
955 	(&(psb)->ps_samples[(psb)->ps_considx & pmc_sample_mask])
956 
957 #define PMC_CONS_SAMPLE_OFF(psb, off)							\
958 	(&(psb)->ps_samples[(off) & pmc_sample_mask])
959 
960 #define PMC_PROD_SAMPLE(psb)					\
961 	(&(psb)->ps_samples[(psb)->ps_prodidx & pmc_sample_mask])
962 
963 /*
964  * struct pmc_cpustate
965  *
966  * A CPU is modelled as a collection of HW PMCs with space for additional
967  * flags.
968  */
969 
970 struct pmc_cpu {
971 	uint32_t	pc_state;	/* physical cpu number + flags */
972 	struct pmc_samplebuffer *pc_sb[3]; /* space for samples */
973 	struct pmc_hw	*pc_hwpmcs[];	/* 'npmc' pointers */
974 };
975 
976 #define	PMC_PCPU_CPU_MASK		0x000000FF
977 #define	PMC_PCPU_FLAGS_MASK		0xFFFFFF00
978 #define	PMC_PCPU_FLAGS_SHIFT		8
979 #define	PMC_PCPU_STATE_TO_CPU(S)	((S) & PMC_PCPU_CPU_MASK)
980 #define	PMC_PCPU_STATE_TO_FLAGS(S)	(((S) & PMC_PCPU_FLAGS_MASK) >> PMC_PCPU_FLAGS_SHIFT)
981 #define	PMC_PCPU_FLAGS_TO_STATE(F)	(((F) << PMC_PCPU_FLAGS_SHIFT) & PMC_PCPU_FLAGS_MASK)
982 #define	PMC_PCPU_CPU_TO_STATE(C)	((C) & PMC_PCPU_CPU_MASK)
983 #define	PMC_PCPU_FLAG_HTT		(PMC_PCPU_FLAGS_TO_STATE(0x1))
984 
985 /*
986  * struct pmc_binding
987  *
988  * CPU binding information.
989  */
990 
991 struct pmc_binding {
992 	int	pb_bound;	/* is bound? */
993 	int	pb_cpu;		/* if so, to which CPU */
994 	u_char	pb_priority;	/* Thread active priority. */
995 };
996 
997 struct pmc_mdep;
998 
999 /*
1000  * struct pmc_classdep
1001  *
1002  * PMC class-dependent operations.
1003  */
1004 struct pmc_classdep {
1005 	uint32_t	pcd_caps;	/* class capabilities */
1006 	enum pmc_class	pcd_class;	/* class id */
1007 	int		pcd_num;	/* number of PMCs */
1008 	int		pcd_ri;		/* row index of the first PMC in class */
1009 	int		pcd_width;	/* width of the PMC */
1010 
1011 	/* configuring/reading/writing the hardware PMCs */
1012 	int (*pcd_config_pmc)(int _cpu, int _ri, struct pmc *_pm);
1013 	int (*pcd_get_config)(int _cpu, int _ri, struct pmc **_ppm);
1014 	int (*pcd_read_pmc)(int _cpu, int _ri, struct pmc *_pm,
1015 	    pmc_value_t *_value);
1016 	int (*pcd_write_pmc)(int _cpu, int _ri, struct pmc *_pm,
1017 	    pmc_value_t _value);
1018 
1019 	/* pmc allocation/release */
1020 	int (*pcd_allocate_pmc)(int _cpu, int _ri, struct pmc *_t,
1021 		const struct pmc_op_pmcallocate *_a);
1022 	int (*pcd_release_pmc)(int _cpu, int _ri, struct pmc *_pm);
1023 
1024 	/* starting and stopping PMCs */
1025 	int (*pcd_start_pmc)(int _cpu, int _ri, struct pmc *_pm);
1026 	int (*pcd_stop_pmc)(int _cpu, int _ri, struct pmc *_pm);
1027 
1028 	/* description */
1029 	int (*pcd_describe)(int _cpu, int _ri, struct pmc_info *_pi,
1030 		struct pmc **_ppmc);
1031 
1032 	/* class-dependent initialization & finalization */
1033 	int (*pcd_pcpu_init)(struct pmc_mdep *_md, int _cpu);
1034 	int (*pcd_pcpu_fini)(struct pmc_mdep *_md, int _cpu);
1035 
1036 	/* machine-specific interface */
1037 	int (*pcd_get_msr)(int _ri, uint32_t *_msr);
1038 };
1039 
1040 /*
1041  * struct pmc_mdep
1042  *
1043  * Machine dependent bits needed per CPU type.
1044  */
1045 
1046 struct pmc_mdep  {
1047 	uint32_t	pmd_cputype;    /* from enum pmc_cputype */
1048 	uint32_t	pmd_npmc;	/* number of PMCs per CPU */
1049 	uint32_t	pmd_nclass;	/* number of PMC classes present */
1050 
1051 	/*
1052 	 * Machine dependent methods.
1053 	 */
1054 
1055 	/* thread context switch in/out */
1056 	int (*pmd_switch_in)(struct pmc_cpu *_p, struct pmc_process *_pp);
1057 	int (*pmd_switch_out)(struct pmc_cpu *_p, struct pmc_process *_pp);
1058 
1059 	/* handle a PMC interrupt */
1060 	int (*pmd_intr)(struct trapframe *_tf);
1061 
1062 	/*
1063 	 * PMC class dependent information.
1064 	 */
1065 	struct pmc_classdep pmd_classdep[];
1066 };
1067 
1068 /*
1069  * Per-CPU state.  This is an array of 'mp_ncpu' pointers
1070  * to struct pmc_cpu descriptors.
1071  */
1072 
1073 extern struct pmc_cpu **pmc_pcpu;
1074 
1075 /* driver statistics */
1076 extern struct pmc_driverstats pmc_stats;
1077 
1078 #if	defined(HWPMC_DEBUG)
1079 
1080 /* HWPMC_DEBUG without KTR will compile but is a no-op. */
1081 #if !defined(KTR) || !defined(KTR_COMPILE) || ((KTR_COMPILE & KTR_SUBSYS) == 0)
1082 #error "HWPMC_DEBUG requires KTR and KTR_COMPILE=KTR_SUBSYS -- see ktr(4)"
1083 #endif
1084 
1085 #include <sys/ktr.h>
1086 
1087 #define	__pmcdbg_used		/* unused variable annotation */
1088 
1089 /*
1090  * Debug flags, major flag groups.
1091  *
1092  * Please keep the DEBUGGING section of the hwpmc(4) man page in sync.
1093  */
1094 struct pmc_debugflags {
1095 	int	pdb_CPU;
1096 	int	pdb_CSW;
1097 	int	pdb_LOG;
1098 	int	pdb_MDP;
1099 	int	pdb_MOD;
1100 	int	pdb_OWN;
1101 	int	pdb_PMC;
1102 	int	pdb_PRC;
1103 	int	pdb_SAM;
1104 };
1105 
1106 extern struct pmc_debugflags pmc_debugflags;
1107 
1108 #define	KTR_PMC			KTR_SUBSYS
1109 
1110 #define	PMC_DEBUG_STRSIZE		128
1111 #define	PMC_DEBUG_DEFAULT_FLAGS		{ 0, 0, 0, 0, 0, 0, 0, 0, 0 }
1112 
1113 #define	PMCDBG0(M, N, L, F) do {					\
1114 	if (pmc_debugflags.pdb_ ## M & (1 << PMC_DEBUG_MIN_ ## N))	\
1115 		CTR0(KTR_PMC, #M ":" #N ":" #L  ": " F);		\
1116 } while (0)
1117 #define	PMCDBG1(M, N, L, F, p1) do {					\
1118 	if (pmc_debugflags.pdb_ ## M & (1 << PMC_DEBUG_MIN_ ## N))	\
1119 		CTR1(KTR_PMC, #M ":" #N ":" #L  ": " F, p1);		\
1120 } while (0)
1121 #define	PMCDBG2(M, N, L, F, p1, p2) do {				\
1122 	if (pmc_debugflags.pdb_ ## M & (1 << PMC_DEBUG_MIN_ ## N))	\
1123 		CTR2(KTR_PMC, #M ":" #N ":" #L  ": " F, p1, p2);	\
1124 } while (0)
1125 #define	PMCDBG3(M, N, L, F, p1, p2, p3) do {				\
1126 	if (pmc_debugflags.pdb_ ## M & (1 << PMC_DEBUG_MIN_ ## N))	\
1127 		CTR3(KTR_PMC, #M ":" #N ":" #L  ": " F, p1, p2, p3);	\
1128 } while (0)
1129 #define	PMCDBG4(M, N, L, F, p1, p2, p3, p4) do {			\
1130 	if (pmc_debugflags.pdb_ ## M & (1 << PMC_DEBUG_MIN_ ## N))	\
1131 		CTR4(KTR_PMC, #M ":" #N ":" #L  ": " F, p1, p2, p3, p4);\
1132 } while (0)
1133 #define	PMCDBG5(M, N, L, F, p1, p2, p3, p4, p5) do {			\
1134 	if (pmc_debugflags.pdb_ ## M & (1 << PMC_DEBUG_MIN_ ## N))	\
1135 		CTR5(KTR_PMC, #M ":" #N ":" #L  ": " F, p1, p2, p3, p4,	\
1136 		    p5);						\
1137 } while (0)
1138 #define	PMCDBG6(M, N, L, F, p1, p2, p3, p4, p5, p6) do {		\
1139 	if (pmc_debugflags.pdb_ ## M & (1 << PMC_DEBUG_MIN_ ## N))	\
1140 		CTR6(KTR_PMC, #M ":" #N ":" #L  ": " F, p1, p2, p3, p4,	\
1141 		    p5, p6);						\
1142 } while (0)
1143 
1144 /* Major numbers */
1145 #define	PMC_DEBUG_MAJ_CPU		0 /* cpu switches */
1146 #define	PMC_DEBUG_MAJ_CSW		1 /* context switches */
1147 #define	PMC_DEBUG_MAJ_LOG		2 /* logging */
1148 #define	PMC_DEBUG_MAJ_MDP		3 /* machine dependent */
1149 #define	PMC_DEBUG_MAJ_MOD		4 /* misc module infrastructure */
1150 #define	PMC_DEBUG_MAJ_OWN		5 /* owner */
1151 #define	PMC_DEBUG_MAJ_PMC		6 /* pmc management */
1152 #define	PMC_DEBUG_MAJ_PRC		7 /* processes */
1153 #define	PMC_DEBUG_MAJ_SAM		8 /* sampling */
1154 
1155 /* Minor numbers */
1156 
1157 /* Common (8 bits) */
1158 #define	PMC_DEBUG_MIN_ALL		0 /* allocation */
1159 #define	PMC_DEBUG_MIN_REL		1 /* release */
1160 #define	PMC_DEBUG_MIN_OPS		2 /* ops: start, stop, ... */
1161 #define	PMC_DEBUG_MIN_INI		3 /* init */
1162 #define	PMC_DEBUG_MIN_FND		4 /* find */
1163 
1164 /* MODULE */
1165 #define	PMC_DEBUG_MIN_PMH	       14 /* pmc_hook */
1166 #define	PMC_DEBUG_MIN_PMS	       15 /* pmc_syscall */
1167 
1168 /* OWN */
1169 #define	PMC_DEBUG_MIN_ORM		8 /* owner remove */
1170 #define	PMC_DEBUG_MIN_OMR		9 /* owner maybe remove */
1171 
1172 /* PROCESSES */
1173 #define	PMC_DEBUG_MIN_TLK		8 /* link target */
1174 #define	PMC_DEBUG_MIN_TUL		9 /* unlink target */
1175 #define	PMC_DEBUG_MIN_EXT	       10 /* process exit */
1176 #define	PMC_DEBUG_MIN_EXC	       11 /* process exec */
1177 #define	PMC_DEBUG_MIN_FRK	       12 /* process fork */
1178 #define	PMC_DEBUG_MIN_ATT	       13 /* attach/detach */
1179 #define	PMC_DEBUG_MIN_SIG	       14 /* signalling */
1180 
1181 /* CONTEXT SWITCHES */
1182 #define	PMC_DEBUG_MIN_SWI		8 /* switch in */
1183 #define	PMC_DEBUG_MIN_SWO		9 /* switch out */
1184 
1185 /* PMC */
1186 #define	PMC_DEBUG_MIN_REG		8 /* pmc register */
1187 #define	PMC_DEBUG_MIN_ALR		9 /* allocate row */
1188 
1189 /* MACHINE DEPENDENT LAYER */
1190 #define	PMC_DEBUG_MIN_REA		8 /* read */
1191 #define	PMC_DEBUG_MIN_WRI		9 /* write */
1192 #define	PMC_DEBUG_MIN_CFG	       10 /* config */
1193 #define	PMC_DEBUG_MIN_STA	       11 /* start */
1194 #define	PMC_DEBUG_MIN_STO	       12 /* stop */
1195 #define	PMC_DEBUG_MIN_INT	       13 /* interrupts */
1196 
1197 /* CPU */
1198 #define	PMC_DEBUG_MIN_BND		8 /* bind */
1199 #define	PMC_DEBUG_MIN_SEL		9 /* select */
1200 
1201 /* LOG */
1202 #define	PMC_DEBUG_MIN_GTB		8 /* get buf */
1203 #define	PMC_DEBUG_MIN_SIO		9 /* schedule i/o */
1204 #define	PMC_DEBUG_MIN_FLS	       10 /* flush */
1205 #define	PMC_DEBUG_MIN_SAM	       11 /* sample */
1206 #define	PMC_DEBUG_MIN_CLO	       12 /* close */
1207 
1208 #else
1209 #define	__pmcdbg_used			__unused
1210 #define	PMCDBG0(M, N, L, F)		/* nothing */
1211 #define	PMCDBG1(M, N, L, F, p1)
1212 #define	PMCDBG2(M, N, L, F, p1, p2)
1213 #define	PMCDBG3(M, N, L, F, p1, p2, p3)
1214 #define	PMCDBG4(M, N, L, F, p1, p2, p3, p4)
1215 #define	PMCDBG5(M, N, L, F, p1, p2, p3, p4, p5)
1216 #define	PMCDBG6(M, N, L, F, p1, p2, p3, p4, p5, p6)
1217 #endif
1218 
1219 /* declare a dedicated memory pool */
1220 MALLOC_DECLARE(M_PMC);
1221 
1222 /*
1223  * Functions
1224  */
1225 
1226 struct pmc_mdep *pmc_md_initialize(void);	/* MD init function */
1227 void	pmc_md_finalize(struct pmc_mdep *_md);	/* MD fini function */
1228 int	pmc_getrowdisp(int _ri);
1229 int	pmc_process_interrupt(int _ring, struct pmc *_pm, struct trapframe *_tf);
1230 int	pmc_save_kernel_callchain(uintptr_t *_cc, int _maxsamples,
1231     struct trapframe *_tf);
1232 int	pmc_save_user_callchain(uintptr_t *_cc, int _maxsamples,
1233     struct trapframe *_tf);
1234 void	pmc_restore_cpu_binding(struct pmc_binding *pb);
1235 void	pmc_save_cpu_binding(struct pmc_binding *pb);
1236 void	pmc_select_cpu(int cpu);
1237 struct pmc_mdep *pmc_mdep_alloc(int nclasses);
1238 void pmc_mdep_free(struct pmc_mdep *md);
1239 uint64_t pmc_rdtsc(void);
1240 #endif /* _KERNEL */
1241 #endif /* _SYS_PMC_H_ */
1242