xref: /freebsd/sys/sys/pmc.h (revision 2b833162)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
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  * $FreeBSD$
33  */
34 
35 #ifndef _SYS_PMC_H_
36 #define	_SYS_PMC_H_
37 
38 #include <dev/hwpmc/pmc_events.h>
39 #include <sys/proc.h>
40 #include <sys/counter.h>
41 #include <machine/pmc_mdep.h>
42 #include <machine/profile.h>
43 #ifdef _KERNEL
44 #include <sys/epoch.h>
45 #include <ck_queue.h>
46 #endif
47 
48 #define	PMC_MODULE_NAME		"hwpmc"
49 #define	PMC_NAME_MAX		64 /* HW counter name size */
50 #define	PMC_CLASS_MAX		8  /* max #classes of PMCs per-system */
51 
52 /*
53  * Kernel<->userland API version number [MMmmpppp]
54  *
55  * Major numbers are to be incremented when an incompatible change to
56  * the ABI occurs that older clients will not be able to handle.
57  *
58  * Minor numbers are incremented when a backwards compatible change
59  * occurs that allows older correct programs to run unchanged.  For
60  * example, when support for a new PMC type is added.
61  *
62  * The patch version is incremented for every bug fix.
63  */
64 #define	PMC_VERSION_MAJOR	0x09
65 #define	PMC_VERSION_MINOR	0x03
66 #define	PMC_VERSION_PATCH	0x0000
67 
68 #define	PMC_VERSION		(PMC_VERSION_MAJOR << 24 |		\
69 	PMC_VERSION_MINOR << 16 | PMC_VERSION_PATCH)
70 
71 #define PMC_CPUID_LEN 64
72 /* cpu model name for pmu lookup */
73 extern char pmc_cpuid[PMC_CPUID_LEN];
74 
75 /*
76  * Kinds of CPUs known.
77  *
78  * We keep track of CPU variants that need to be distinguished in
79  * some way for PMC operations.  CPU names are grouped by manufacturer
80  * and numbered sparsely in order to minimize changes to the ABI involved
81  * when new CPUs are added.
82  *
83  * Please keep the pmc(3) manual page in sync with this list.
84  */
85 
86 #define	__PMC_CPUS()						\
87 	__PMC_CPU(AMD_K7,	0x00,	"AMD K7")		\
88 	__PMC_CPU(AMD_K8,	0x01,	"AMD K8")		\
89 	__PMC_CPU(INTEL_P5,	0x80,	"Intel Pentium")	\
90 	__PMC_CPU(INTEL_P6,	0x81,	"Intel Pentium Pro")	\
91 	__PMC_CPU(INTEL_CL,	0x82,	"Intel Celeron")	\
92 	__PMC_CPU(INTEL_PII,	0x83,	"Intel Pentium II")	\
93 	__PMC_CPU(INTEL_PIII,	0x84,	"Intel Pentium III")	\
94 	__PMC_CPU(INTEL_PM,	0x85,	"Intel Pentium M")	\
95 	__PMC_CPU(INTEL_PIV,	0x86,	"Intel Pentium IV")	\
96 	__PMC_CPU(INTEL_CORE,	0x87,	"Intel Core Solo/Duo")	\
97 	__PMC_CPU(INTEL_CORE2,	0x88,	"Intel Core2")		\
98 	__PMC_CPU(INTEL_CORE2EXTREME,	0x89,	"Intel Core2 Extreme")	\
99 	__PMC_CPU(INTEL_ATOM,	0x8A,	"Intel Atom")		\
100 	__PMC_CPU(INTEL_COREI7, 0x8B,   "Intel Core i7")	\
101 	__PMC_CPU(INTEL_WESTMERE, 0x8C,   "Intel Westmere")	\
102 	__PMC_CPU(INTEL_SANDYBRIDGE, 0x8D,   "Intel Sandy Bridge")	\
103 	__PMC_CPU(INTEL_IVYBRIDGE, 0x8E,   "Intel Ivy Bridge")	\
104 	__PMC_CPU(INTEL_SANDYBRIDGE_XEON, 0x8F,   "Intel Sandy Bridge Xeon")	\
105 	__PMC_CPU(INTEL_IVYBRIDGE_XEON, 0x90,   "Intel Ivy Bridge Xeon")	\
106 	__PMC_CPU(INTEL_HASWELL, 0x91,   "Intel Haswell")	\
107 	__PMC_CPU(INTEL_ATOM_SILVERMONT, 0x92,	"Intel Atom Silvermont")    \
108 	__PMC_CPU(INTEL_NEHALEM_EX, 0x93,   "Intel Nehalem Xeon 7500")	\
109 	__PMC_CPU(INTEL_WESTMERE_EX, 0x94,   "Intel Westmere Xeon E7")	\
110 	__PMC_CPU(INTEL_HASWELL_XEON, 0x95,   "Intel Haswell Xeon E5 v3") \
111 	__PMC_CPU(INTEL_BROADWELL, 0x96,   "Intel Broadwell") \
112 	__PMC_CPU(INTEL_BROADWELL_XEON, 0x97,   "Intel Broadwell Xeon") \
113 	__PMC_CPU(INTEL_SKYLAKE, 0x98,   "Intel Skylake")		\
114 	__PMC_CPU(INTEL_SKYLAKE_XEON, 0x99,   "Intel Skylake Xeon")	\
115 	__PMC_CPU(INTEL_ATOM_GOLDMONT, 0x9A,   "Intel Atom Goldmont")	\
116 	__PMC_CPU(INTEL_ICELAKE, 0x9B,	"Intel Icelake")		\
117 	__PMC_CPU(INTEL_ICELAKE_XEON, 0x9C,	"Intel Icelake Xeon")	\
118 	__PMC_CPU(INTEL_ALDERLAKE, 0x9D,	"Intel Alderlake")	\
119 	__PMC_CPU(INTEL_ATOM_GOLDMONT_P, 0x9E,	"Intel Atom Goldmont Plus")    \
120 	__PMC_CPU(INTEL_ATOM_TREMONT, 0x9F,	"Intel Atom Tremont")    \
121 	__PMC_CPU(INTEL_XSCALE,	0x100,	"Intel XScale")		\
122 	__PMC_CPU(MIPS_24K,     0x200,  "MIPS 24K")		\
123 	__PMC_CPU(MIPS_OCTEON,  0x201,  "Cavium Octeon")	\
124 	__PMC_CPU(MIPS_74K,     0x202,  "MIPS 74K")		\
125 	__PMC_CPU(MIPS_BERI,	0x203,  "BERI")			\
126 	__PMC_CPU(PPC_7450,     0x300,  "PowerPC MPC7450")	\
127 	__PMC_CPU(PPC_E500,     0x340,  "PowerPC e500 Core")	\
128 	__PMC_CPU(PPC_970,      0x380,  "IBM PowerPC 970")	\
129 	__PMC_CPU(PPC_POWER8,   0x390,  "IBM POWER8")		\
130 	__PMC_CPU(GENERIC, 	0x400,  "Generic")		\
131 	__PMC_CPU(ARMV7_CORTEX_A5,	0x500,	"ARMv7 Cortex A5")	\
132 	__PMC_CPU(ARMV7_CORTEX_A7,	0x501,	"ARMv7 Cortex A7")	\
133 	__PMC_CPU(ARMV7_CORTEX_A8,	0x502,	"ARMv7 Cortex A8")	\
134 	__PMC_CPU(ARMV7_CORTEX_A9,	0x503,	"ARMv7 Cortex A9")	\
135 	__PMC_CPU(ARMV7_CORTEX_A15,	0x504,	"ARMv7 Cortex A15")	\
136 	__PMC_CPU(ARMV7_CORTEX_A17,	0x505,	"ARMv7 Cortex A17")	\
137 	__PMC_CPU(ARMV8_CORTEX_A53,	0x600,	"ARMv8 Cortex A53")	\
138 	__PMC_CPU(ARMV8_CORTEX_A57,	0x601,	"ARMv8 Cortex A57")	\
139 	__PMC_CPU(ARMV8_CORTEX_A76,	0x602,	"ARMv8 Cortex A76")
140 
141 enum pmc_cputype {
142 #undef	__PMC_CPU
143 #define	__PMC_CPU(S,V,D)	PMC_CPU_##S = V,
144 	__PMC_CPUS()
145 };
146 
147 #define	PMC_CPU_FIRST	PMC_CPU_AMD_K7
148 #define	PMC_CPU_LAST	PMC_CPU_ARMV8_CORTEX_A76
149 
150 /*
151  * Classes of PMCs
152  */
153 
154 #define	__PMC_CLASSES()							\
155 	__PMC_CLASS(TSC,	0x00,	"CPU Timestamp counter")	\
156 	__PMC_CLASS(K7,		0x01,	"AMD K7 performance counters")	\
157 	__PMC_CLASS(K8,		0x02,	"AMD K8 performance counters")	\
158 	__PMC_CLASS(P5,		0x03,	"Intel Pentium counters")	\
159 	__PMC_CLASS(P6,		0x04,	"Intel Pentium Pro counters")	\
160 	__PMC_CLASS(P4,		0x05,	"Intel Pentium-IV counters")	\
161 	__PMC_CLASS(IAF,	0x06,	"Intel Core2/Atom, fixed function") \
162 	__PMC_CLASS(IAP,	0x07,	"Intel Core...Atom, programmable") \
163 	__PMC_CLASS(UCF,	0x08,	"Intel Uncore fixed function")	\
164 	__PMC_CLASS(UCP,	0x09,	"Intel Uncore programmable")	\
165 	__PMC_CLASS(XSCALE,	0x0A,	"Intel XScale counters")	\
166 	__PMC_CLASS(MIPS24K,	0x0B,	"MIPS 24K")			\
167 	__PMC_CLASS(OCTEON,	0x0C,	"Cavium Octeon")		\
168 	__PMC_CLASS(PPC7450,	0x0D,	"Motorola MPC7450 class")	\
169 	__PMC_CLASS(PPC970,	0x0E,	"IBM PowerPC 970 class")	\
170 	__PMC_CLASS(SOFT,	0x0F,	"Software events")		\
171 	__PMC_CLASS(ARMV7,	0x10,	"ARMv7")			\
172 	__PMC_CLASS(ARMV8,	0x11,	"ARMv8")			\
173 	__PMC_CLASS(MIPS74K,	0x12,	"MIPS 74K")			\
174 	__PMC_CLASS(E500,	0x13,	"Freescale e500 class")		\
175 	__PMC_CLASS(BERI,	0x14,	"MIPS BERI")			\
176 	__PMC_CLASS(POWER8,	0x15,	"IBM POWER8 class")		\
177 	__PMC_CLASS(DMC620_PMU_CD2, 0x16, "ARM DMC620 Memory Controller PMU CLKDIV2") \
178 	__PMC_CLASS(DMC620_PMU_C, 0x17, "ARM DMC620 Memory Controller PMU CLK") \
179 	__PMC_CLASS(CMN600_PMU, 0x18,	"Arm CoreLink CMN600 Coherent Mesh Network PMU")
180 
181 enum pmc_class {
182 #undef  __PMC_CLASS
183 #define	__PMC_CLASS(S,V,D)	PMC_CLASS_##S = V,
184 	__PMC_CLASSES()
185 };
186 
187 #define	PMC_CLASS_FIRST	PMC_CLASS_TSC
188 #define	PMC_CLASS_LAST	PMC_CLASS_CMN600_PMU
189 
190 /*
191  * A PMC can be in the following states:
192  *
193  * Hardware states:
194  *   DISABLED   -- administratively prohibited from being used.
195  *   FREE       -- HW available for use
196  * Software states:
197  *   ALLOCATED  -- allocated
198  *   STOPPED    -- allocated, but not counting events
199  *   RUNNING    -- allocated, and in operation; 'pm_runcount'
200  *                 holds the number of CPUs using this PMC at
201  *                 a given instant
202  *   DELETED    -- being destroyed
203  */
204 
205 #define	__PMC_HWSTATES()			\
206 	__PMC_STATE(DISABLED)			\
207 	__PMC_STATE(FREE)
208 
209 #define	__PMC_SWSTATES()			\
210 	__PMC_STATE(ALLOCATED)			\
211 	__PMC_STATE(STOPPED)			\
212 	__PMC_STATE(RUNNING)			\
213 	__PMC_STATE(DELETED)
214 
215 #define	__PMC_STATES()				\
216 	__PMC_HWSTATES()			\
217 	__PMC_SWSTATES()
218 
219 enum pmc_state {
220 #undef	__PMC_STATE
221 #define	__PMC_STATE(S)	PMC_STATE_##S,
222 	__PMC_STATES()
223 	__PMC_STATE(MAX)
224 };
225 
226 #define	PMC_STATE_FIRST	PMC_STATE_DISABLED
227 #define	PMC_STATE_LAST	PMC_STATE_DELETED
228 
229 /*
230  * An allocated PMC may used as a 'global' counter or as a
231  * 'thread-private' one.  Each such mode of use can be in either
232  * statistical sampling mode or in counting mode.  Thus a PMC in use
233  *
234  * SS i.e., SYSTEM STATISTICAL  -- system-wide statistical profiling
235  * SC i.e., SYSTEM COUNTER      -- system-wide counting mode
236  * TS i.e., THREAD STATISTICAL  -- thread virtual, statistical profiling
237  * TC i.e., THREAD COUNTER      -- thread virtual, counting mode
238  *
239  * Statistical profiling modes rely on the PMC periodically delivering
240  * a interrupt to the CPU (when the configured number of events have
241  * been measured), so the PMC must have the ability to generate
242  * interrupts.
243  *
244  * In counting modes, the PMC counts its configured events, with the
245  * value of the PMC being read whenever needed by its owner process.
246  *
247  * The thread specific modes "virtualize" the PMCs -- the PMCs appear
248  * to be thread private and count events only when the profiled thread
249  * actually executes on the CPU.
250  *
251  * The system-wide "global" modes keep the PMCs running all the time
252  * and are used to measure the behaviour of the whole system.
253  */
254 
255 #define	__PMC_MODES()				\
256 	__PMC_MODE(SS,	0)			\
257 	__PMC_MODE(SC,	1)			\
258 	__PMC_MODE(TS,	2)			\
259 	__PMC_MODE(TC,	3)
260 
261 enum pmc_mode {
262 #undef	__PMC_MODE
263 #define	__PMC_MODE(M,N)	PMC_MODE_##M = N,
264 	__PMC_MODES()
265 };
266 
267 #define	PMC_MODE_FIRST	PMC_MODE_SS
268 #define	PMC_MODE_LAST	PMC_MODE_TC
269 
270 #define	PMC_IS_COUNTING_MODE(mode)				\
271 	((mode) == PMC_MODE_SC || (mode) == PMC_MODE_TC)
272 #define	PMC_IS_SYSTEM_MODE(mode)				\
273 	((mode) == PMC_MODE_SS || (mode) == PMC_MODE_SC)
274 #define	PMC_IS_SAMPLING_MODE(mode)				\
275 	((mode) == PMC_MODE_SS || (mode) == PMC_MODE_TS)
276 #define	PMC_IS_VIRTUAL_MODE(mode)				\
277 	((mode) == PMC_MODE_TS || (mode) == PMC_MODE_TC)
278 
279 /*
280  * PMC row disposition
281  */
282 
283 #define	__PMC_DISPOSITIONS(N)					\
284 	__PMC_DISP(STANDALONE)	/* global/disabled counters */	\
285 	__PMC_DISP(FREE)	/* free/available */		\
286 	__PMC_DISP(THREAD)	/* thread-virtual PMCs */	\
287 	__PMC_DISP(UNKNOWN)	/* sentinel */
288 
289 enum pmc_disp {
290 #undef	__PMC_DISP
291 #define	__PMC_DISP(D)	PMC_DISP_##D ,
292 	__PMC_DISPOSITIONS()
293 };
294 
295 #define	PMC_DISP_FIRST	PMC_DISP_STANDALONE
296 #define	PMC_DISP_LAST	PMC_DISP_THREAD
297 
298 /*
299  * Counter capabilities
300  *
301  * __PMC_CAPS(NAME, VALUE, DESCRIPTION)
302  */
303 
304 #define	__PMC_CAPS()							\
305 	__PMC_CAP(INTERRUPT,	0, "generate interrupts")		\
306 	__PMC_CAP(USER,		1, "count user-mode events")		\
307 	__PMC_CAP(SYSTEM,	2, "count system-mode events")		\
308 	__PMC_CAP(EDGE,		3, "do edge detection of events")	\
309 	__PMC_CAP(THRESHOLD,	4, "ignore events below a threshold")	\
310 	__PMC_CAP(READ,		5, "read PMC counter")			\
311 	__PMC_CAP(WRITE,	6, "reprogram PMC counter")		\
312 	__PMC_CAP(INVERT,	7, "invert comparison sense")		\
313 	__PMC_CAP(QUALIFIER,	8, "further qualify monitored events")	\
314 	__PMC_CAP(PRECISE,	9, "perform precise sampling")		\
315 	__PMC_CAP(TAGGING,	10, "tag upstream events")		\
316 	__PMC_CAP(CASCADE,	11, "cascade counters")			\
317 	__PMC_CAP(SYSWIDE,	12, "system wide counter")		\
318 	__PMC_CAP(DOMWIDE,	13, "NUMA domain wide counter")
319 
320 enum pmc_caps
321 {
322 #undef	__PMC_CAP
323 #define	__PMC_CAP(NAME, VALUE, DESCR)	PMC_CAP_##NAME = (1 << VALUE) ,
324 	__PMC_CAPS()
325 };
326 
327 #define	PMC_CAP_FIRST		PMC_CAP_INTERRUPT
328 #define	PMC_CAP_LAST		PMC_CAP_DOMWIDE
329 
330 /*
331  * PMC Event Numbers
332  *
333  * These are generated from the definitions in "dev/hwpmc/pmc_events.h".
334  */
335 
336 enum pmc_event {
337 #undef	__PMC_EV
338 #undef	__PMC_EV_BLOCK
339 #define	__PMC_EV_BLOCK(C,V)	PMC_EV_ ## C ## __BLOCK_START = (V) - 1 ,
340 #define	__PMC_EV(C,N)		PMC_EV_ ## C ## _ ## N ,
341 	__PMC_EVENTS()
342 };
343 
344 /*
345  * PMC SYSCALL INTERFACE
346  */
347 
348 /*
349  * "PMC_OPS" -- these are the commands recognized by the kernel
350  * module, and are used when performing a system call from userland.
351  */
352 #define	__PMC_OPS()							\
353 	__PMC_OP(CONFIGURELOG, "Set log file")				\
354 	__PMC_OP(FLUSHLOG, "Flush log file")				\
355 	__PMC_OP(GETCPUINFO, "Get system CPU information")		\
356 	__PMC_OP(GETDRIVERSTATS, "Get driver statistics")		\
357 	__PMC_OP(GETMODULEVERSION, "Get module version")		\
358 	__PMC_OP(GETPMCINFO, "Get per-cpu PMC information")		\
359 	__PMC_OP(PMCADMIN, "Set PMC state")				\
360 	__PMC_OP(PMCALLOCATE, "Allocate and configure a PMC")		\
361 	__PMC_OP(PMCATTACH, "Attach a PMC to a process")		\
362 	__PMC_OP(PMCDETACH, "Detach a PMC from a process")		\
363 	__PMC_OP(PMCGETMSR, "Get a PMC's hardware address")		\
364 	__PMC_OP(PMCRELEASE, "Release a PMC")				\
365 	__PMC_OP(PMCRW, "Read/Set a PMC")				\
366 	__PMC_OP(PMCSETCOUNT, "Set initial count/sampling rate")	\
367 	__PMC_OP(PMCSTART, "Start a PMC")				\
368 	__PMC_OP(PMCSTOP, "Stop a PMC")					\
369 	__PMC_OP(WRITELOG, "Write a cookie to the log file")		\
370 	__PMC_OP(CLOSELOG, "Close log file")				\
371 	__PMC_OP(GETDYNEVENTINFO, "Get dynamic events list")
372 
373 enum pmc_ops {
374 #undef	__PMC_OP
375 #define	__PMC_OP(N, D)	PMC_OP_##N,
376 	__PMC_OPS()
377 };
378 
379 /*
380  * Flags used in operations on PMCs.
381  */
382 
383 #define	PMC_F_UNUSED1		0x00000001 /* unused */
384 #define	PMC_F_DESCENDANTS	0x00000002 /*OP ALLOCATE track descendants */
385 #define	PMC_F_LOG_PROCCSW	0x00000004 /*OP ALLOCATE track ctx switches */
386 #define	PMC_F_LOG_PROCEXIT	0x00000008 /*OP ALLOCATE log proc exits */
387 #define	PMC_F_NEWVALUE		0x00000010 /*OP RW write new value */
388 #define	PMC_F_OLDVALUE		0x00000020 /*OP RW get old value */
389 
390 /* V2 API */
391 #define	PMC_F_CALLCHAIN		0x00000080 /*OP ALLOCATE capture callchains */
392 #define	PMC_F_USERCALLCHAIN	0x00000100 /*OP ALLOCATE use userspace stack */
393 
394 /* internal flags */
395 #define	PMC_F_ATTACHED_TO_OWNER	0x00010000 /*attached to owner*/
396 #define	PMC_F_NEEDS_LOGFILE	0x00020000 /*needs log file */
397 #define	PMC_F_ATTACH_DONE	0x00040000 /*attached at least once */
398 
399 #define	PMC_CALLCHAIN_DEPTH_MAX	512
400 
401 #define	PMC_CC_F_USERSPACE	0x01	   /*userspace callchain*/
402 
403 /*
404  * Cookies used to denote allocated PMCs, and the values of PMCs.
405  */
406 
407 typedef uint32_t	pmc_id_t;
408 typedef uint64_t	pmc_value_t;
409 
410 #define	PMC_ID_INVALID		(~ (pmc_id_t) 0)
411 
412 /*
413  * PMC IDs have the following format:
414  *
415  * +-----------------------+-------+-----------+
416  * |   CPU      | PMC MODE | CLASS | ROW INDEX |
417  * +-----------------------+-------+-----------+
418  *
419  * where CPU is 12 bits, MODE 4, CLASS 8, and ROW INDEX 8  Field 'CPU'
420  * is set to the requested CPU for system-wide PMCs or PMC_CPU_ANY for
421  * process-mode PMCs.  Field 'PMC MODE' is the allocated PMC mode.
422  * Field 'PMC CLASS' is the class of the PMC.  Field 'ROW INDEX' is the
423  * row index for the PMC.
424  *
425  * The 'ROW INDEX' ranges over 0..NWPMCS where NHWPMCS is the total
426  * number of hardware PMCs on this cpu.
427  */
428 
429 #define	PMC_ID_TO_ROWINDEX(ID)	((ID) & 0xFF)
430 #define	PMC_ID_TO_CLASS(ID)	(((ID) & 0xFF00) >> 8)
431 #define	PMC_ID_TO_MODE(ID)	(((ID) & 0xF0000) >> 16)
432 #define	PMC_ID_TO_CPU(ID)	(((ID) & 0xFFF00000) >> 20)
433 #define	PMC_ID_MAKE_ID(CPU,MODE,CLASS,ROWINDEX)			\
434 	((((CPU) & 0xFFF) << 20) | (((MODE) & 0xF) << 16) |	\
435 	(((CLASS) & 0xFF) << 8) | ((ROWINDEX) & 0xFF))
436 
437 /*
438  * Data structures for system calls supported by the pmc driver.
439  */
440 
441 /*
442  * OP PMCALLOCATE
443  *
444  * Allocate a PMC on the named CPU.
445  */
446 
447 #define	PMC_CPU_ANY	~0
448 
449 struct pmc_op_pmcallocate {
450 	uint32_t	pm_caps;	/* PMC_CAP_* */
451 	uint32_t	pm_cpu;		/* CPU number or PMC_CPU_ANY */
452 	enum pmc_class	pm_class;	/* class of PMC desired */
453 	enum pmc_event	pm_ev;		/* [enum pmc_event] desired */
454 	uint32_t	pm_flags;	/* additional modifiers PMC_F_* */
455 	enum pmc_mode	pm_mode;	/* desired mode */
456 	pmc_id_t	pm_pmcid;	/* [return] process pmc id */
457 	pmc_value_t	pm_count;	/* initial/sample count */
458 
459 	union pmc_md_op_pmcallocate pm_md; /* MD layer extensions */
460 };
461 
462 /*
463  * OP PMCADMIN
464  *
465  * Set the administrative state (i.e., whether enabled or disabled) of
466  * a PMC 'pm_pmc' on CPU 'pm_cpu'.  Note that 'pm_pmc' specifies an
467  * absolute PMC number and need not have been first allocated by the
468  * calling process.
469  */
470 
471 struct pmc_op_pmcadmin {
472 	int		pm_cpu;		/* CPU# */
473 	uint32_t	pm_flags;	/* flags */
474 	int		pm_pmc;         /* PMC# */
475 	enum pmc_state  pm_state;	/* desired state */
476 };
477 
478 /*
479  * OP PMCATTACH / OP PMCDETACH
480  *
481  * Attach/detach a PMC and a process.
482  */
483 
484 struct pmc_op_pmcattach {
485 	pmc_id_t	pm_pmc;		/* PMC to attach to */
486 	pid_t		pm_pid;		/* target process */
487 };
488 
489 /*
490  * OP PMCSETCOUNT
491  *
492  * Set the sampling rate (i.e., the reload count) for statistical counters.
493  * 'pm_pmcid' need to have been previously allocated using PMCALLOCATE.
494  */
495 
496 struct pmc_op_pmcsetcount {
497 	pmc_value_t	pm_count;	/* initial/sample count */
498 	pmc_id_t	pm_pmcid;	/* PMC id to set */
499 };
500 
501 /*
502  * OP PMCRW
503  *
504  * Read the value of a PMC named by 'pm_pmcid'.  'pm_pmcid' needs
505  * to have been previously allocated using PMCALLOCATE.
506  */
507 
508 struct pmc_op_pmcrw {
509 	uint32_t	pm_flags;	/* PMC_F_{OLD,NEW}VALUE*/
510 	pmc_id_t	pm_pmcid;	/* pmc id */
511 	pmc_value_t	pm_value;	/* new&returned value */
512 };
513 
514 /*
515  * OP GETPMCINFO
516  *
517  * retrieve PMC state for a named CPU.  The caller is expected to
518  * allocate 'npmc' * 'struct pmc_info' bytes of space for the return
519  * values.
520  */
521 
522 struct pmc_info {
523 	char		pm_name[PMC_NAME_MAX]; /* pmc name */
524 	enum pmc_class	pm_class;	/* enum pmc_class */
525 	int		pm_enabled;	/* whether enabled */
526 	enum pmc_disp	pm_rowdisp;	/* FREE, THREAD or STANDLONE */
527 	pid_t		pm_ownerpid;	/* owner, or -1 */
528 	enum pmc_mode	pm_mode;	/* current mode [enum pmc_mode] */
529 	enum pmc_event	pm_event;	/* current event */
530 	uint32_t	pm_flags;	/* current flags */
531 	pmc_value_t	pm_reloadcount;	/* sampling counters only */
532 };
533 
534 struct pmc_op_getpmcinfo {
535 	int32_t		pm_cpu;		/* 0 <= cpu < mp_maxid */
536 	struct pmc_info	pm_pmcs[];	/* space for 'npmc' structures */
537 };
538 
539 /*
540  * OP GETCPUINFO
541  *
542  * Retrieve system CPU information.
543  */
544 
545 struct pmc_classinfo {
546 	enum pmc_class	pm_class;	/* class id */
547 	uint32_t	pm_caps;	/* counter capabilities */
548 	uint32_t	pm_width;	/* width of the PMC */
549 	uint32_t	pm_num;		/* number of PMCs in class */
550 };
551 
552 struct pmc_op_getcpuinfo {
553 	enum pmc_cputype pm_cputype; /* what kind of CPU */
554 	uint32_t	pm_ncpu;    /* max CPU number */
555 	uint32_t	pm_npmc;    /* #PMCs per CPU */
556 	uint32_t	pm_nclass;  /* #classes of PMCs */
557 	struct pmc_classinfo  pm_classes[PMC_CLASS_MAX];
558 };
559 
560 /*
561  * OP CONFIGURELOG
562  *
563  * Configure a log file for writing system-wide statistics to.
564  */
565 
566 struct pmc_op_configurelog {
567 	int		pm_flags;
568 	int		pm_logfd;   /* logfile fd (or -1) */
569 };
570 
571 /*
572  * OP GETDRIVERSTATS
573  *
574  * Retrieve pmc(4) driver-wide statistics.
575  */
576 #ifdef _KERNEL
577 struct pmc_driverstats {
578 	counter_u64_t	pm_intr_ignored;	/* #interrupts ignored */
579 	counter_u64_t	pm_intr_processed;	/* #interrupts processed */
580 	counter_u64_t	pm_intr_bufferfull;	/* #interrupts with ENOSPC */
581 	counter_u64_t	pm_syscalls;		/* #syscalls */
582 	counter_u64_t	pm_syscall_errors;	/* #syscalls with errors */
583 	counter_u64_t	pm_buffer_requests;	/* #buffer requests */
584 	counter_u64_t	pm_buffer_requests_failed; /* #failed buffer requests */
585 	counter_u64_t	pm_log_sweeps;		/* #sample buffer processing
586 						   passes */
587 	counter_u64_t	pm_merges;		/* merged k+u */
588 	counter_u64_t	pm_overwrites;		/* UR overwrites */
589 };
590 #endif
591 
592 struct pmc_op_getdriverstats {
593 	unsigned int	pm_intr_ignored;	/* #interrupts ignored */
594 	unsigned int	pm_intr_processed;	/* #interrupts processed */
595 	unsigned int	pm_intr_bufferfull;	/* #interrupts with ENOSPC */
596 	unsigned int	pm_syscalls;		/* #syscalls */
597 	unsigned int	pm_syscall_errors;	/* #syscalls with errors */
598 	unsigned int	pm_buffer_requests;	/* #buffer requests */
599 	unsigned int	pm_buffer_requests_failed; /* #failed buffer requests */
600 	unsigned int	pm_log_sweeps;		/* #sample buffer processing
601 						   passes */
602 };
603 
604 /*
605  * OP RELEASE / OP START / OP STOP
606  *
607  * Simple operations on a PMC id.
608  */
609 
610 struct pmc_op_simple {
611 	pmc_id_t	pm_pmcid;
612 };
613 
614 /*
615  * OP WRITELOG
616  *
617  * Flush the current log buffer and write 4 bytes of user data to it.
618  */
619 
620 struct pmc_op_writelog {
621 	uint32_t	pm_userdata;
622 };
623 
624 /*
625  * OP GETMSR
626  *
627  * Retrieve the machine specific address associated with the allocated
628  * PMC.  This number can be used subsequently with a read-performance-counter
629  * instruction.
630  */
631 
632 struct pmc_op_getmsr {
633 	uint32_t	pm_msr;		/* machine specific address */
634 	pmc_id_t	pm_pmcid;	/* allocated pmc id */
635 };
636 
637 /*
638  * OP GETDYNEVENTINFO
639  *
640  * Retrieve a PMC dynamic class events list.
641  */
642 
643 struct pmc_dyn_event_descr {
644 	char		pm_ev_name[PMC_NAME_MAX];
645 	enum pmc_event	pm_ev_code;
646 };
647 
648 struct pmc_op_getdyneventinfo {
649 	enum pmc_class			pm_class;
650 	unsigned int			pm_nevent;
651 	struct pmc_dyn_event_descr	pm_events[PMC_EV_DYN_COUNT];
652 };
653 
654 #ifdef _KERNEL
655 
656 #include <sys/malloc.h>
657 #include <sys/sysctl.h>
658 #include <sys/_cpuset.h>
659 
660 #include <machine/frame.h>
661 
662 #define	PMC_HASH_SIZE				1024
663 #define	PMC_MTXPOOL_SIZE			2048
664 #define	PMC_LOG_BUFFER_SIZE			256
665 #define	PMC_NLOGBUFFERS_PCPU			32
666 #define	PMC_NSAMPLES				256
667 #define	PMC_CALLCHAIN_DEPTH			128
668 #define	PMC_THREADLIST_MAX			128
669 
670 #define PMC_SYSCTL_NAME_PREFIX "kern." PMC_MODULE_NAME "."
671 
672 /*
673  * Locking keys
674  *
675  * (b) - pmc_bufferlist_mtx (spin lock)
676  * (k) - pmc_kthread_mtx (sleep lock)
677  * (o) - po->po_mtx (spin lock)
678  * (g) - global_epoch_preempt (epoch)
679  * (p) - pmc_sx (sx)
680  */
681 
682 /*
683  * PMC commands
684  */
685 
686 struct pmc_syscall_args {
687 	register_t	pmop_code;	/* one of PMC_OP_* */
688 	void		*pmop_data;	/* syscall parameter */
689 };
690 
691 /*
692  * Interface to processor specific s1tuff
693  */
694 
695 /*
696  * struct pmc_descr
697  *
698  * Machine independent (i.e., the common parts) of a human readable
699  * PMC description.
700  */
701 
702 struct pmc_descr {
703 	char		pd_name[PMC_NAME_MAX]; /* name */
704 	uint32_t	pd_caps;	/* capabilities */
705 	enum pmc_class	pd_class;	/* class of the PMC */
706 	uint32_t	pd_width;	/* width in bits */
707 };
708 
709 /*
710  * struct pmc_target
711  *
712  * This structure records all the target processes associated with a
713  * PMC.
714  */
715 
716 struct pmc_target {
717 	LIST_ENTRY(pmc_target)	pt_next;
718 	struct pmc_process	*pt_process; /* target descriptor */
719 };
720 
721 /*
722  * struct pmc
723  *
724  * Describes each allocated PMC.
725  *
726  * Each PMC has precisely one owner, namely the process that allocated
727  * the PMC.
728  *
729  * A PMC may be attached to multiple target processes.  The
730  * 'pm_targets' field links all the target processes being monitored
731  * by this PMC.
732  *
733  * The 'pm_savedvalue' field is protected by a mutex.
734  *
735  * On a multi-cpu machine, multiple target threads associated with a
736  * process-virtual PMC could be concurrently executing on different
737  * CPUs.  The 'pm_runcount' field is atomically incremented every time
738  * the PMC gets scheduled on a CPU and atomically decremented when it
739  * get descheduled.  Deletion of a PMC is only permitted when this
740  * field is '0'.
741  *
742  */
743 struct pmc_pcpu_state {
744 	uint32_t pps_overflowcnt;	/* count overflow interrupts */
745 	uint8_t pps_stalled;
746 	uint8_t pps_cpustate;
747 } __aligned(CACHE_LINE_SIZE);
748 struct pmc {
749 	LIST_HEAD(,pmc_target)	pm_targets;	/* list of target processes */
750 	LIST_ENTRY(pmc)		pm_next;	/* owner's list */
751 
752 	/*
753 	 * System-wide PMCs are allocated on a CPU and are not moved
754 	 * around.  For system-wide PMCs we record the CPU the PMC was
755 	 * allocated on in the 'CPU' field of the pmc ID.
756 	 *
757 	 * Virtual PMCs run on whichever CPU is currently executing
758 	 * their targets' threads.  For these PMCs we need to save
759 	 * their current PMC counter values when they are taken off
760 	 * CPU.
761 	 */
762 
763 	union {
764 		pmc_value_t	pm_savedvalue;	/* Virtual PMCS */
765 	} pm_gv;
766 
767 	/*
768 	 * For sampling mode PMCs, we keep track of the PMC's "reload
769 	 * count", which is the counter value to be loaded in when
770 	 * arming the PMC for the next counting session.  For counting
771 	 * modes on PMCs that are read-only (e.g., the x86 TSC), we
772 	 * keep track of the initial value at the start of
773 	 * counting-mode operation.
774 	 */
775 
776 	union {
777 		pmc_value_t	pm_reloadcount;	/* sampling PMC modes */
778 		pmc_value_t	pm_initial;	/* counting PMC modes */
779 	} pm_sc;
780 
781 	struct pmc_pcpu_state *pm_pcpu_state;
782 	volatile cpuset_t pm_cpustate;	/* CPUs where PMC should be active */
783 	uint32_t	pm_caps;	/* PMC capabilities */
784 	enum pmc_event	pm_event;	/* event being measured */
785 	uint32_t	pm_flags;	/* additional flags PMC_F_... */
786 	struct pmc_owner *pm_owner;	/* owner thread state */
787 	counter_u64_t		pm_runcount;	/* #cpus currently on */
788 	enum pmc_state	pm_state;	/* current PMC state */
789 
790 	/*
791 	 * The PMC ID field encodes the row-index for the PMC, its
792 	 * mode, class and the CPU# associated with the PMC.
793 	 */
794 
795 	pmc_id_t	pm_id;		/* allocated PMC id */
796 	enum pmc_class pm_class;
797 
798 	/* md extensions */
799 	union pmc_md_pmc	pm_md;
800 };
801 
802 /*
803  * Accessor macros for 'struct pmc'
804  */
805 
806 #define	PMC_TO_MODE(P)		PMC_ID_TO_MODE((P)->pm_id)
807 #define	PMC_TO_CLASS(P)		PMC_ID_TO_CLASS((P)->pm_id)
808 #define	PMC_TO_ROWINDEX(P)	PMC_ID_TO_ROWINDEX((P)->pm_id)
809 #define	PMC_TO_CPU(P)		PMC_ID_TO_CPU((P)->pm_id)
810 
811 /*
812  * struct pmc_threadpmcstate
813  *
814  * Record per-PMC, per-thread state.
815  */
816 struct pmc_threadpmcstate {
817 	pmc_value_t	pt_pmcval;	/* per-thread reload count */
818 };
819 
820 /*
821  * struct pmc_thread
822  *
823  * Record a 'target' thread being profiled.
824  */
825 struct pmc_thread {
826 	LIST_ENTRY(pmc_thread) pt_next;		/* linked list */
827 	struct thread	*pt_td;			/* target thread */
828 	struct pmc_threadpmcstate pt_pmcs[];	/* per-PMC state */
829 };
830 
831 /*
832  * struct pmc_process
833  *
834  * Record a 'target' process being profiled.
835  *
836  * The target process being profiled could be different from the owner
837  * process which allocated the PMCs.  Each target process descriptor
838  * is associated with NHWPMC 'struct pmc *' pointers.  Each PMC at a
839  * given hardware row-index 'n' will use slot 'n' of the 'pp_pmcs[]'
840  * array.  The size of this structure is thus PMC architecture
841  * dependent.
842  *
843  */
844 
845 struct pmc_targetstate {
846 	struct pmc	*pp_pmc;   /* target PMC */
847 	pmc_value_t	pp_pmcval; /* per-process value */
848 };
849 
850 struct pmc_process {
851 	LIST_ENTRY(pmc_process) pp_next;	/* hash chain */
852 	LIST_HEAD(,pmc_thread) pp_tds;		/* list of threads */
853 	struct mtx	*pp_tdslock;		/* lock on pp_tds thread list */
854 	int		pp_refcnt;		/* reference count */
855 	uint32_t	pp_flags;		/* flags PMC_PP_* */
856 	struct proc	*pp_proc;		/* target process */
857 	struct pmc_targetstate pp_pmcs[];       /* NHWPMCs */
858 };
859 
860 #define	PMC_PP_ENABLE_MSR_ACCESS	0x00000001
861 
862 /*
863  * struct pmc_owner
864  *
865  * We associate a PMC with an 'owner' process.
866  *
867  * A process can be associated with 0..NCPUS*NHWPMC PMCs during its
868  * lifetime, where NCPUS is the numbers of CPUS in the system and
869  * NHWPMC is the number of hardware PMCs per CPU.  These are
870  * maintained in the list headed by the 'po_pmcs' to save on space.
871  *
872  */
873 
874 struct pmc_owner  {
875 	LIST_ENTRY(pmc_owner)	po_next;	/* hash chain */
876 	CK_LIST_ENTRY(pmc_owner)	po_ssnext;	/* (g/p) list of SS PMC owners */
877 	LIST_HEAD(, pmc)	po_pmcs;	/* owned PMC list */
878 	TAILQ_HEAD(, pmclog_buffer) po_logbuffers; /* (o) logbuffer list */
879 	struct mtx		po_mtx;		/* spin lock for (o) */
880 	struct proc		*po_owner;	/* owner proc */
881 	uint32_t		po_flags;	/* (k) flags PMC_PO_* */
882 	struct proc		*po_kthread;	/* (k) helper kthread */
883 	struct file		*po_file;	/* file reference */
884 	int			po_error;	/* recorded error */
885 	short			po_sscount;	/* # SS PMCs owned */
886 	short			po_logprocmaps;	/* global mappings done */
887 	struct pmclog_buffer	*po_curbuf[MAXCPU];	/* current log buffer */
888 };
889 
890 #define	PMC_PO_OWNS_LOGFILE		0x00000001 /* has a log file */
891 #define	PMC_PO_SHUTDOWN			0x00000010 /* in the process of shutdown */
892 #define	PMC_PO_INITIAL_MAPPINGS_DONE	0x00000020
893 
894 /*
895  * struct pmc_hw -- describe the state of the PMC hardware
896  *
897  * When in use, a HW PMC is associated with one allocated 'struct pmc'
898  * pointed to by field 'phw_pmc'.  When inactive, this field is NULL.
899  *
900  * On an SMP box, one or more HW PMC's in process virtual mode with
901  * the same 'phw_pmc' could be executing on different CPUs.  In order
902  * to handle this case correctly, we need to ensure that only
903  * incremental counts get added to the saved value in the associated
904  * 'struct pmc'.  The 'phw_save' field is used to keep the saved PMC
905  * value at the time the hardware is started during this context
906  * switch (i.e., the difference between the new (hardware) count and
907  * the saved count is atomically added to the count field in 'struct
908  * pmc' at context switch time).
909  *
910  */
911 
912 struct pmc_hw {
913 	uint32_t	phw_state;	/* see PHW_* macros below */
914 	struct pmc	*phw_pmc;	/* current thread PMC */
915 };
916 
917 #define	PMC_PHW_RI_MASK		0x000000FF
918 #define	PMC_PHW_CPU_SHIFT	8
919 #define	PMC_PHW_CPU_MASK	0x0000FF00
920 #define	PMC_PHW_FLAGS_SHIFT	16
921 #define	PMC_PHW_FLAGS_MASK	0xFFFF0000
922 
923 #define	PMC_PHW_INDEX_TO_STATE(ri)	((ri) & PMC_PHW_RI_MASK)
924 #define	PMC_PHW_STATE_TO_INDEX(state)	((state) & PMC_PHW_RI_MASK)
925 #define	PMC_PHW_CPU_TO_STATE(cpu)	(((cpu) << PMC_PHW_CPU_SHIFT) & \
926 	PMC_PHW_CPU_MASK)
927 #define	PMC_PHW_STATE_TO_CPU(state)	(((state) & PMC_PHW_CPU_MASK) >> \
928 	PMC_PHW_CPU_SHIFT)
929 #define	PMC_PHW_FLAGS_TO_STATE(flags)	(((flags) << PMC_PHW_FLAGS_SHIFT) & \
930 	PMC_PHW_FLAGS_MASK)
931 #define	PMC_PHW_STATE_TO_FLAGS(state)	(((state) & PMC_PHW_FLAGS_MASK) >> \
932 	PMC_PHW_FLAGS_SHIFT)
933 #define	PMC_PHW_FLAG_IS_ENABLED		(PMC_PHW_FLAGS_TO_STATE(0x01))
934 #define	PMC_PHW_FLAG_IS_SHAREABLE	(PMC_PHW_FLAGS_TO_STATE(0x02))
935 
936 /*
937  * struct pmc_sample
938  *
939  * Space for N (tunable) PC samples and associated control data.
940  */
941 
942 struct pmc_sample {
943 	uint16_t		ps_nsamples;	/* callchain depth */
944 	uint16_t		ps_nsamples_actual;
945 	uint16_t		ps_cpu;		/* cpu number */
946 	uint16_t		ps_flags;	/* other flags */
947 	lwpid_t			ps_tid;		/* thread id */
948 	pid_t			ps_pid;		/* process PID or -1 */
949 	int		ps_ticks; /* ticks at sample time */
950 	/* pad */
951 	struct thread		*ps_td;		/* which thread */
952 	struct pmc		*ps_pmc;	/* interrupting PMC */
953 	uintptr_t		*ps_pc;		/* (const) callchain start */
954 	uint64_t		ps_tsc;		/* tsc value */
955 };
956 
957 #define 	PMC_SAMPLE_FREE		((uint16_t) 0)
958 #define 	PMC_USER_CALLCHAIN_PENDING	((uint16_t) 0xFFFF)
959 
960 struct pmc_samplebuffer {
961 	volatile uint64_t		ps_prodidx; /* producer index */
962 	volatile uint64_t		ps_considx; /* consumer index */
963 	uintptr_t		*ps_callchains;	/* all saved call chains */
964 	struct pmc_sample	ps_samples[];	/* array of sample entries */
965 };
966 
967 #define PMC_CONS_SAMPLE(psb)					\
968 	(&(psb)->ps_samples[(psb)->ps_considx & pmc_sample_mask])
969 
970 #define PMC_CONS_SAMPLE_OFF(psb, off)							\
971 	(&(psb)->ps_samples[(off) & pmc_sample_mask])
972 
973 #define PMC_PROD_SAMPLE(psb)					\
974 	(&(psb)->ps_samples[(psb)->ps_prodidx & pmc_sample_mask])
975 
976 /*
977  * struct pmc_cpustate
978  *
979  * A CPU is modelled as a collection of HW PMCs with space for additional
980  * flags.
981  */
982 
983 struct pmc_cpu {
984 	uint32_t	pc_state;	/* physical cpu number + flags */
985 	struct pmc_samplebuffer *pc_sb[3]; /* space for samples */
986 	struct pmc_hw	*pc_hwpmcs[];	/* 'npmc' pointers */
987 };
988 
989 #define	PMC_PCPU_CPU_MASK		0x000000FF
990 #define	PMC_PCPU_FLAGS_MASK		0xFFFFFF00
991 #define	PMC_PCPU_FLAGS_SHIFT		8
992 #define	PMC_PCPU_STATE_TO_CPU(S)	((S) & PMC_PCPU_CPU_MASK)
993 #define	PMC_PCPU_STATE_TO_FLAGS(S)	(((S) & PMC_PCPU_FLAGS_MASK) >> PMC_PCPU_FLAGS_SHIFT)
994 #define	PMC_PCPU_FLAGS_TO_STATE(F)	(((F) << PMC_PCPU_FLAGS_SHIFT) & PMC_PCPU_FLAGS_MASK)
995 #define	PMC_PCPU_CPU_TO_STATE(C)	((C) & PMC_PCPU_CPU_MASK)
996 #define	PMC_PCPU_FLAG_HTT		(PMC_PCPU_FLAGS_TO_STATE(0x1))
997 
998 /*
999  * struct pmc_binding
1000  *
1001  * CPU binding information.
1002  */
1003 
1004 struct pmc_binding {
1005 	int	pb_bound;	/* is bound? */
1006 	int	pb_cpu;		/* if so, to which CPU */
1007 	u_char	pb_priority;	/* Thread active priority. */
1008 };
1009 
1010 struct pmc_mdep;
1011 
1012 /*
1013  * struct pmc_classdep
1014  *
1015  * PMC class-dependent operations.
1016  */
1017 struct pmc_classdep {
1018 	uint32_t	pcd_caps;	/* class capabilities */
1019 	enum pmc_class	pcd_class;	/* class id */
1020 	int		pcd_num;	/* number of PMCs */
1021 	int		pcd_ri;		/* row index of the first PMC in class */
1022 	int		pcd_width;	/* width of the PMC */
1023 
1024 	/* configuring/reading/writing the hardware PMCs */
1025 	int (*pcd_config_pmc)(int _cpu, int _ri, struct pmc *_pm);
1026 	int (*pcd_get_config)(int _cpu, int _ri, struct pmc **_ppm);
1027 	int (*pcd_read_pmc)(int _cpu, int _ri, pmc_value_t *_value);
1028 	int (*pcd_write_pmc)(int _cpu, int _ri, pmc_value_t _value);
1029 
1030 	/* pmc allocation/release */
1031 	int (*pcd_allocate_pmc)(int _cpu, int _ri, struct pmc *_t,
1032 		const struct pmc_op_pmcallocate *_a);
1033 	int (*pcd_release_pmc)(int _cpu, int _ri, struct pmc *_pm);
1034 
1035 	/* starting and stopping PMCs */
1036 	int (*pcd_start_pmc)(int _cpu, int _ri);
1037 	int (*pcd_stop_pmc)(int _cpu, int _ri);
1038 
1039 	/* description */
1040 	int (*pcd_describe)(int _cpu, int _ri, struct pmc_info *_pi,
1041 		struct pmc **_ppmc);
1042 
1043 	/* class-dependent initialization & finalization */
1044 	int (*pcd_pcpu_init)(struct pmc_mdep *_md, int _cpu);
1045 	int (*pcd_pcpu_fini)(struct pmc_mdep *_md, int _cpu);
1046 
1047 	/* machine-specific interface */
1048 	int (*pcd_get_msr)(int _ri, uint32_t *_msr);
1049 };
1050 
1051 /*
1052  * struct pmc_mdep
1053  *
1054  * Machine dependent bits needed per CPU type.
1055  */
1056 
1057 struct pmc_mdep  {
1058 	uint32_t	pmd_cputype;    /* from enum pmc_cputype */
1059 	uint32_t	pmd_npmc;	/* number of PMCs per CPU */
1060 	uint32_t	pmd_nclass;	/* number of PMC classes present */
1061 
1062 	/*
1063 	 * Machine dependent methods.
1064 	 */
1065 
1066 	/* per-cpu initialization and finalization */
1067 	int (*pmd_pcpu_init)(struct pmc_mdep *_md, int _cpu);
1068 	int (*pmd_pcpu_fini)(struct pmc_mdep *_md, int _cpu);
1069 
1070 	/* thread context switch in/out */
1071 	int (*pmd_switch_in)(struct pmc_cpu *_p, struct pmc_process *_pp);
1072 	int (*pmd_switch_out)(struct pmc_cpu *_p, struct pmc_process *_pp);
1073 
1074 	/* handle a PMC interrupt */
1075 	int (*pmd_intr)(struct trapframe *_tf);
1076 
1077 	/*
1078 	 * PMC class dependent information.
1079 	 */
1080 	struct pmc_classdep pmd_classdep[];
1081 };
1082 
1083 /*
1084  * Per-CPU state.  This is an array of 'mp_ncpu' pointers
1085  * to struct pmc_cpu descriptors.
1086  */
1087 
1088 extern struct pmc_cpu **pmc_pcpu;
1089 
1090 /* driver statistics */
1091 extern struct pmc_driverstats pmc_stats;
1092 
1093 #if	defined(HWPMC_DEBUG)
1094 #include <sys/ktr.h>
1095 
1096 /* debug flags, major flag groups */
1097 struct pmc_debugflags {
1098 	int	pdb_CPU;
1099 	int	pdb_CSW;
1100 	int	pdb_LOG;
1101 	int	pdb_MDP;
1102 	int	pdb_MOD;
1103 	int	pdb_OWN;
1104 	int	pdb_PMC;
1105 	int	pdb_PRC;
1106 	int	pdb_SAM;
1107 };
1108 
1109 extern struct pmc_debugflags pmc_debugflags;
1110 
1111 #define	KTR_PMC			KTR_SUBSYS
1112 
1113 #define	PMC_DEBUG_STRSIZE		128
1114 #define	PMC_DEBUG_DEFAULT_FLAGS		{ 0, 0, 0, 0, 0, 0, 0, 0, 0 }
1115 
1116 #define	PMCDBG0(M, N, L, F) do {					\
1117 	if (pmc_debugflags.pdb_ ## M & (1 << PMC_DEBUG_MIN_ ## N))	\
1118 		CTR0(KTR_PMC, #M ":" #N ":" #L  ": " F);		\
1119 } while (0)
1120 #define	PMCDBG1(M, N, L, F, p1) do {					\
1121 	if (pmc_debugflags.pdb_ ## M & (1 << PMC_DEBUG_MIN_ ## N))	\
1122 		CTR1(KTR_PMC, #M ":" #N ":" #L  ": " F, p1);		\
1123 } while (0)
1124 #define	PMCDBG2(M, N, L, F, p1, p2) do {				\
1125 	if (pmc_debugflags.pdb_ ## M & (1 << PMC_DEBUG_MIN_ ## N))	\
1126 		CTR2(KTR_PMC, #M ":" #N ":" #L  ": " F, p1, p2);	\
1127 } while (0)
1128 #define	PMCDBG3(M, N, L, F, p1, p2, p3) do {				\
1129 	if (pmc_debugflags.pdb_ ## M & (1 << PMC_DEBUG_MIN_ ## N))	\
1130 		CTR3(KTR_PMC, #M ":" #N ":" #L  ": " F, p1, p2, p3);	\
1131 } while (0)
1132 #define	PMCDBG4(M, N, L, F, p1, p2, p3, p4) do {			\
1133 	if (pmc_debugflags.pdb_ ## M & (1 << PMC_DEBUG_MIN_ ## N))	\
1134 		CTR4(KTR_PMC, #M ":" #N ":" #L  ": " F, p1, p2, p3, p4);\
1135 } while (0)
1136 #define	PMCDBG5(M, N, L, F, p1, p2, p3, p4, p5) do {			\
1137 	if (pmc_debugflags.pdb_ ## M & (1 << PMC_DEBUG_MIN_ ## N))	\
1138 		CTR5(KTR_PMC, #M ":" #N ":" #L  ": " F, p1, p2, p3, p4,	\
1139 		    p5);						\
1140 } while (0)
1141 #define	PMCDBG6(M, N, L, F, p1, p2, p3, p4, p5, p6) do {		\
1142 	if (pmc_debugflags.pdb_ ## M & (1 << PMC_DEBUG_MIN_ ## N))	\
1143 		CTR6(KTR_PMC, #M ":" #N ":" #L  ": " F, p1, p2, p3, p4,	\
1144 		    p5, p6);						\
1145 } while (0)
1146 
1147 /* Major numbers */
1148 #define	PMC_DEBUG_MAJ_CPU		0 /* cpu switches */
1149 #define	PMC_DEBUG_MAJ_CSW		1 /* context switches */
1150 #define	PMC_DEBUG_MAJ_LOG		2 /* logging */
1151 #define	PMC_DEBUG_MAJ_MDP		3 /* machine dependent */
1152 #define	PMC_DEBUG_MAJ_MOD		4 /* misc module infrastructure */
1153 #define	PMC_DEBUG_MAJ_OWN		5 /* owner */
1154 #define	PMC_DEBUG_MAJ_PMC		6 /* pmc management */
1155 #define	PMC_DEBUG_MAJ_PRC		7 /* processes */
1156 #define	PMC_DEBUG_MAJ_SAM		8 /* sampling */
1157 
1158 /* Minor numbers */
1159 
1160 /* Common (8 bits) */
1161 #define	PMC_DEBUG_MIN_ALL		0 /* allocation */
1162 #define	PMC_DEBUG_MIN_REL		1 /* release */
1163 #define	PMC_DEBUG_MIN_OPS		2 /* ops: start, stop, ... */
1164 #define	PMC_DEBUG_MIN_INI		3 /* init */
1165 #define	PMC_DEBUG_MIN_FND		4 /* find */
1166 
1167 /* MODULE */
1168 #define	PMC_DEBUG_MIN_PMH	       14 /* pmc_hook */
1169 #define	PMC_DEBUG_MIN_PMS	       15 /* pmc_syscall */
1170 
1171 /* OWN */
1172 #define	PMC_DEBUG_MIN_ORM		8 /* owner remove */
1173 #define	PMC_DEBUG_MIN_OMR		9 /* owner maybe remove */
1174 
1175 /* PROCESSES */
1176 #define	PMC_DEBUG_MIN_TLK		8 /* link target */
1177 #define	PMC_DEBUG_MIN_TUL		9 /* unlink target */
1178 #define	PMC_DEBUG_MIN_EXT	       10 /* process exit */
1179 #define	PMC_DEBUG_MIN_EXC	       11 /* process exec */
1180 #define	PMC_DEBUG_MIN_FRK	       12 /* process fork */
1181 #define	PMC_DEBUG_MIN_ATT	       13 /* attach/detach */
1182 #define	PMC_DEBUG_MIN_SIG	       14 /* signalling */
1183 
1184 /* CONTEXT SWITCHES */
1185 #define	PMC_DEBUG_MIN_SWI		8 /* switch in */
1186 #define	PMC_DEBUG_MIN_SWO		9 /* switch out */
1187 
1188 /* PMC */
1189 #define	PMC_DEBUG_MIN_REG		8 /* pmc register */
1190 #define	PMC_DEBUG_MIN_ALR		9 /* allocate row */
1191 
1192 /* MACHINE DEPENDENT LAYER */
1193 #define	PMC_DEBUG_MIN_REA		8 /* read */
1194 #define	PMC_DEBUG_MIN_WRI		9 /* write */
1195 #define	PMC_DEBUG_MIN_CFG	       10 /* config */
1196 #define	PMC_DEBUG_MIN_STA	       11 /* start */
1197 #define	PMC_DEBUG_MIN_STO	       12 /* stop */
1198 #define	PMC_DEBUG_MIN_INT	       13 /* interrupts */
1199 
1200 /* CPU */
1201 #define	PMC_DEBUG_MIN_BND		8 /* bind */
1202 #define	PMC_DEBUG_MIN_SEL		9 /* select */
1203 
1204 /* LOG */
1205 #define	PMC_DEBUG_MIN_GTB		8 /* get buf */
1206 #define	PMC_DEBUG_MIN_SIO		9 /* schedule i/o */
1207 #define	PMC_DEBUG_MIN_FLS	       10 /* flush */
1208 #define	PMC_DEBUG_MIN_SAM	       11 /* sample */
1209 #define	PMC_DEBUG_MIN_CLO	       12 /* close */
1210 
1211 #else
1212 #define	PMCDBG0(M, N, L, F)		/* nothing */
1213 #define	PMCDBG1(M, N, L, F, p1)
1214 #define	PMCDBG2(M, N, L, F, p1, p2)
1215 #define	PMCDBG3(M, N, L, F, p1, p2, p3)
1216 #define	PMCDBG4(M, N, L, F, p1, p2, p3, p4)
1217 #define	PMCDBG5(M, N, L, F, p1, p2, p3, p4, p5)
1218 #define	PMCDBG6(M, N, L, F, p1, p2, p3, p4, p5, p6)
1219 #endif
1220 
1221 /* declare a dedicated memory pool */
1222 MALLOC_DECLARE(M_PMC);
1223 
1224 /*
1225  * Functions
1226  */
1227 
1228 struct pmc_mdep *pmc_md_initialize(void);	/* MD init function */
1229 void	pmc_md_finalize(struct pmc_mdep *_md);	/* MD fini function */
1230 int	pmc_getrowdisp(int _ri);
1231 int	pmc_process_interrupt(int _ring, struct pmc *_pm, struct trapframe *_tf);
1232 int	pmc_save_kernel_callchain(uintptr_t *_cc, int _maxsamples,
1233     struct trapframe *_tf);
1234 int	pmc_save_user_callchain(uintptr_t *_cc, int _maxsamples,
1235     struct trapframe *_tf);
1236 void	pmc_restore_cpu_binding(struct pmc_binding *pb);
1237 void	pmc_save_cpu_binding(struct pmc_binding *pb);
1238 void	pmc_select_cpu(int cpu);
1239 struct pmc_mdep *pmc_mdep_alloc(int nclasses);
1240 void pmc_mdep_free(struct pmc_mdep *md);
1241 uint64_t pmc_rdtsc(void);
1242 #endif /* _KERNEL */
1243 #endif /* _SYS_PMC_H_ */
1244