xref: /freebsd/sys/sys/pmc.h (revision 82d6d46d0d1dbdaa32c754f97e94a772d01cb26c)
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_XSCALE,		0x100,	"Intel XScale")			\
111     __PMC_CPU(PPC_7450,			0x300,	"PowerPC MPC7450")		\
112     __PMC_CPU(PPC_E500,			0x340,	"PowerPC e500 Core")		\
113     __PMC_CPU(PPC_970,			0x380,	"IBM PowerPC 970")		\
114     __PMC_CPU(PPC_POWER8,		0x390,	"IBM POWER8")			\
115     __PMC_CPU(GENERIC,			0x400,	"Generic")			\
116     __PMC_CPU(ARMV7_CORTEX_A5,		0x500,	"ARMv7 Cortex A5")		\
117     __PMC_CPU(ARMV7_CORTEX_A7,		0x501,	"ARMv7 Cortex A7")		\
118     __PMC_CPU(ARMV7_CORTEX_A8,		0x502,	"ARMv7 Cortex A8")		\
119     __PMC_CPU(ARMV7_CORTEX_A9,		0x503,	"ARMv7 Cortex A9")		\
120     __PMC_CPU(ARMV7_CORTEX_A15,		0x504,	"ARMv7 Cortex A15")		\
121     __PMC_CPU(ARMV7_CORTEX_A17,		0x505,	"ARMv7 Cortex A17")		\
122     __PMC_CPU(ARMV8_CORTEX_A53,		0x600,	"ARMv8 Cortex A53")		\
123     __PMC_CPU(ARMV8_CORTEX_A57,		0x601,	"ARMv8 Cortex A57")		\
124     __PMC_CPU(ARMV8_CORTEX_A76,		0x602,	"ARMv8 Cortex A76")
125 
126 enum pmc_cputype {
127 #undef	__PMC_CPU
128 #define	__PMC_CPU(S,V,D)	PMC_CPU_##S = V,
129 	__PMC_CPUS()
130 };
131 
132 #define	PMC_CPU_FIRST	PMC_CPU_AMD_K8
133 #define	PMC_CPU_LAST	PMC_CPU_ARMV8_CORTEX_A76
134 
135 /*
136  * Classes of PMCs
137  */
138 #define	__PMC_CLASSES()								\
139     __PMC_CLASS(TSC,		0x00,	"CPU Timestamp counter")		\
140     __PMC_CLASS(K8,		0x02,	"AMD K8 performance counters")		\
141     __PMC_CLASS(IAF,		0x06,	"Intel Core2/Atom, fixed function")	\
142     __PMC_CLASS(IAP,		0x07,	"Intel Core...Atom, programmable")	\
143     __PMC_CLASS(UCF,		0x08,	"Intel Uncore fixed function")		\
144     __PMC_CLASS(UCP,		0x09,	"Intel Uncore programmable")		\
145     __PMC_CLASS(XSCALE,		0x0A,	"Intel XScale counters")		\
146     __PMC_CLASS(PPC7450,	0x0D,	"Motorola MPC7450 class")		\
147     __PMC_CLASS(PPC970,		0x0E,	"IBM PowerPC 970 class")		\
148     __PMC_CLASS(SOFT,		0x0F,	"Software events")			\
149     __PMC_CLASS(ARMV7,		0x10,	"ARMv7")				\
150     __PMC_CLASS(ARMV8,		0x11,	"ARMv8")				\
151     __PMC_CLASS(E500,		0x13,	"Freescale e500 class")			\
152     __PMC_CLASS(POWER8,		0x15,	"IBM POWER8 class")			\
153     __PMC_CLASS(DMC620_PMU_CD2,	0x16,	"ARM DMC620 Memory Controller PMU CLKDIV2") \
154     __PMC_CLASS(DMC620_PMU_C,	0x17,	"ARM DMC620 Memory Controller PMU CLK")	\
155     __PMC_CLASS(CMN600_PMU,	0x18,	"Arm CoreLink CMN600 Coherent Mesh Network PMU")
156 
157 enum pmc_class {
158 #undef  __PMC_CLASS
159 #define	__PMC_CLASS(S,V,D)	PMC_CLASS_##S = V,
160 	__PMC_CLASSES()
161 };
162 
163 #define	PMC_CLASS_FIRST	PMC_CLASS_TSC
164 #define	PMC_CLASS_LAST	PMC_CLASS_CMN600_PMU
165 
166 /*
167  * A PMC can be in the following states:
168  *
169  * Hardware states:
170  *   DISABLED   -- administratively prohibited from being used.
171  *   FREE       -- HW available for use
172  * Software states:
173  *   ALLOCATED  -- allocated
174  *   STOPPED    -- allocated, but not counting events
175  *   RUNNING    -- allocated, and in operation; 'pm_runcount'
176  *                 holds the number of CPUs using this PMC at
177  *                 a given instant
178  *   DELETED    -- being destroyed
179  */
180 
181 #define	__PMC_HWSTATES()			\
182 	__PMC_STATE(DISABLED)			\
183 	__PMC_STATE(FREE)
184 
185 #define	__PMC_SWSTATES()			\
186 	__PMC_STATE(ALLOCATED)			\
187 	__PMC_STATE(STOPPED)			\
188 	__PMC_STATE(RUNNING)			\
189 	__PMC_STATE(DELETED)
190 
191 #define	__PMC_STATES()				\
192 	__PMC_HWSTATES()			\
193 	__PMC_SWSTATES()
194 
195 enum pmc_state {
196 #undef	__PMC_STATE
197 #define	__PMC_STATE(S)	PMC_STATE_##S,
198 	__PMC_STATES()
199 	__PMC_STATE(MAX)
200 };
201 
202 #define	PMC_STATE_FIRST	PMC_STATE_DISABLED
203 #define	PMC_STATE_LAST	PMC_STATE_DELETED
204 
205 /*
206  * An allocated PMC may used as a 'global' counter or as a
207  * 'thread-private' one.  Each such mode of use can be in either
208  * statistical sampling mode or in counting mode.  Thus a PMC in use
209  *
210  * SS i.e., SYSTEM STATISTICAL  -- system-wide statistical profiling
211  * SC i.e., SYSTEM COUNTER      -- system-wide counting mode
212  * TS i.e., THREAD STATISTICAL  -- thread virtual, statistical profiling
213  * TC i.e., THREAD COUNTER      -- thread virtual, counting mode
214  *
215  * Statistical profiling modes rely on the PMC periodically delivering
216  * a interrupt to the CPU (when the configured number of events have
217  * been measured), so the PMC must have the ability to generate
218  * interrupts.
219  *
220  * In counting modes, the PMC counts its configured events, with the
221  * value of the PMC being read whenever needed by its owner process.
222  *
223  * The thread specific modes "virtualize" the PMCs -- the PMCs appear
224  * to be thread private and count events only when the profiled thread
225  * actually executes on the CPU.
226  *
227  * The system-wide "global" modes keep the PMCs running all the time
228  * and are used to measure the behaviour of the whole system.
229  */
230 
231 #define	__PMC_MODES()				\
232 	__PMC_MODE(SS,	0)			\
233 	__PMC_MODE(SC,	1)			\
234 	__PMC_MODE(TS,	2)			\
235 	__PMC_MODE(TC,	3)
236 
237 enum pmc_mode {
238 #undef	__PMC_MODE
239 #define	__PMC_MODE(M,N)	PMC_MODE_##M = N,
240 	__PMC_MODES()
241 };
242 
243 #define	PMC_MODE_FIRST	PMC_MODE_SS
244 #define	PMC_MODE_LAST	PMC_MODE_TC
245 
246 #define	PMC_IS_COUNTING_MODE(mode)				\
247 	((mode) == PMC_MODE_SC || (mode) == PMC_MODE_TC)
248 #define	PMC_IS_SYSTEM_MODE(mode)				\
249 	((mode) == PMC_MODE_SS || (mode) == PMC_MODE_SC)
250 #define	PMC_IS_SAMPLING_MODE(mode)				\
251 	((mode) == PMC_MODE_SS || (mode) == PMC_MODE_TS)
252 #define	PMC_IS_VIRTUAL_MODE(mode)				\
253 	((mode) == PMC_MODE_TS || (mode) == PMC_MODE_TC)
254 
255 /*
256  * PMC row disposition
257  */
258 
259 #define	__PMC_DISPOSITIONS(N)					\
260 	__PMC_DISP(STANDALONE)	/* global/disabled counters */	\
261 	__PMC_DISP(FREE)	/* free/available */		\
262 	__PMC_DISP(THREAD)	/* thread-virtual PMCs */	\
263 	__PMC_DISP(UNKNOWN)	/* sentinel */
264 
265 enum pmc_disp {
266 #undef	__PMC_DISP
267 #define	__PMC_DISP(D)	PMC_DISP_##D ,
268 	__PMC_DISPOSITIONS()
269 };
270 
271 #define	PMC_DISP_FIRST	PMC_DISP_STANDALONE
272 #define	PMC_DISP_LAST	PMC_DISP_THREAD
273 
274 /*
275  * Counter capabilities
276  *
277  * __PMC_CAPS(NAME, VALUE, DESCRIPTION)
278  */
279 
280 #define	__PMC_CAPS()							\
281 	__PMC_CAP(INTERRUPT,	0, "generate interrupts")		\
282 	__PMC_CAP(USER,		1, "count user-mode events")		\
283 	__PMC_CAP(SYSTEM,	2, "count system-mode events")		\
284 	__PMC_CAP(EDGE,		3, "do edge detection of events")	\
285 	__PMC_CAP(THRESHOLD,	4, "ignore events below a threshold")	\
286 	__PMC_CAP(READ,		5, "read PMC counter")			\
287 	__PMC_CAP(WRITE,	6, "reprogram PMC counter")		\
288 	__PMC_CAP(INVERT,	7, "invert comparison sense")		\
289 	__PMC_CAP(QUALIFIER,	8, "further qualify monitored events")	\
290 	__PMC_CAP(PRECISE,	9, "perform precise sampling")		\
291 	__PMC_CAP(TAGGING,	10, "tag upstream events")		\
292 	__PMC_CAP(CASCADE,	11, "cascade counters")			\
293 	__PMC_CAP(SYSWIDE,	12, "system wide counter")		\
294 	__PMC_CAP(DOMWIDE,	13, "NUMA domain wide counter")
295 
296 enum pmc_caps
297 {
298 #undef	__PMC_CAP
299 #define	__PMC_CAP(NAME, VALUE, DESCR)	PMC_CAP_##NAME = (1 << VALUE) ,
300 	__PMC_CAPS()
301 };
302 
303 #define	PMC_CAP_FIRST		PMC_CAP_INTERRUPT
304 #define	PMC_CAP_LAST		PMC_CAP_DOMWIDE
305 
306 /*
307  * PMC Event Numbers
308  *
309  * These are generated from the definitions in "dev/hwpmc/pmc_events.h".
310  */
311 
312 enum pmc_event {
313 #undef	__PMC_EV
314 #undef	__PMC_EV_BLOCK
315 #define	__PMC_EV_BLOCK(C,V)	PMC_EV_ ## C ## __BLOCK_START = (V) - 1 ,
316 #define	__PMC_EV(C,N)		PMC_EV_ ## C ## _ ## N ,
317 	__PMC_EVENTS()
318 };
319 
320 /*
321  * PMC SYSCALL INTERFACE
322  */
323 
324 /*
325  * "PMC_OPS" -- these are the commands recognized by the kernel
326  * module, and are used when performing a system call from userland.
327  */
328 #define	__PMC_OPS()							\
329 	__PMC_OP(CONFIGURELOG, "Set log file")				\
330 	__PMC_OP(FLUSHLOG, "Flush log file")				\
331 	__PMC_OP(GETCPUINFO, "Get system CPU information")		\
332 	__PMC_OP(GETDRIVERSTATS, "Get driver statistics")		\
333 	__PMC_OP(GETMODULEVERSION, "Get module version")		\
334 	__PMC_OP(GETPMCINFO, "Get per-cpu PMC information")		\
335 	__PMC_OP(PMCADMIN, "Set PMC state")				\
336 	__PMC_OP(PMCALLOCATE, "Allocate and configure a PMC")		\
337 	__PMC_OP(PMCATTACH, "Attach a PMC to a process")		\
338 	__PMC_OP(PMCDETACH, "Detach a PMC from a process")		\
339 	__PMC_OP(PMCGETMSR, "Get a PMC's hardware address")		\
340 	__PMC_OP(PMCRELEASE, "Release a PMC")				\
341 	__PMC_OP(PMCRW, "Read/Set a PMC")				\
342 	__PMC_OP(PMCSETCOUNT, "Set initial count/sampling rate")	\
343 	__PMC_OP(PMCSTART, "Start a PMC")				\
344 	__PMC_OP(PMCSTOP, "Stop a PMC")					\
345 	__PMC_OP(WRITELOG, "Write a cookie to the log file")		\
346 	__PMC_OP(CLOSELOG, "Close log file")				\
347 	__PMC_OP(GETDYNEVENTINFO, "Get dynamic events list")
348 
349 enum pmc_ops {
350 #undef	__PMC_OP
351 #define	__PMC_OP(N, D)	PMC_OP_##N,
352 	__PMC_OPS()
353 };
354 
355 /*
356  * Flags used in operations on PMCs.
357  */
358 
359 #define	PMC_F_UNUSED1		0x00000001 /* unused */
360 #define	PMC_F_DESCENDANTS	0x00000002 /*OP ALLOCATE track descendants */
361 #define	PMC_F_LOG_PROCCSW	0x00000004 /*OP ALLOCATE track ctx switches */
362 #define	PMC_F_LOG_PROCEXIT	0x00000008 /*OP ALLOCATE log proc exits */
363 #define	PMC_F_NEWVALUE		0x00000010 /*OP RW write new value */
364 #define	PMC_F_OLDVALUE		0x00000020 /*OP RW get old value */
365 
366 /* V2 API */
367 #define	PMC_F_CALLCHAIN		0x00000080 /*OP ALLOCATE capture callchains */
368 #define	PMC_F_USERCALLCHAIN	0x00000100 /*OP ALLOCATE use userspace stack */
369 
370 /* V10 API */
371 #define	PMC_F_EV_PMU		0x00000200 /*
372 					    * OP ALLOCATE: pm_ev has special
373 					    * userspace meaning; counter
374 					    * configuration is communicated
375 					    * through class-dependent fields
376 					    */
377 
378 /* internal flags */
379 #define	PMC_F_ATTACHED_TO_OWNER	0x00010000 /*attached to owner*/
380 #define	PMC_F_NEEDS_LOGFILE	0x00020000 /*needs log file */
381 #define	PMC_F_ATTACH_DONE	0x00040000 /*attached at least once */
382 
383 #define	PMC_CALLCHAIN_DEPTH_MAX	512
384 
385 #define	PMC_CC_F_USERSPACE	0x01	   /*userspace callchain*/
386 
387 /*
388  * Cookies used to denote allocated PMCs, and the values of PMCs.
389  */
390 
391 typedef uint32_t	pmc_id_t;
392 typedef uint64_t	pmc_value_t;
393 
394 #define	PMC_ID_INVALID		(~ (pmc_id_t) 0)
395 
396 /*
397  * PMC IDs have the following format:
398  *
399  * +-----------------------+-------+-----------+
400  * |   CPU      | PMC MODE | CLASS | ROW INDEX |
401  * +-----------------------+-------+-----------+
402  *
403  * where CPU is 12 bits, MODE 4, CLASS 8, and ROW INDEX 8  Field 'CPU'
404  * is set to the requested CPU for system-wide PMCs or PMC_CPU_ANY for
405  * process-mode PMCs.  Field 'PMC MODE' is the allocated PMC mode.
406  * Field 'PMC CLASS' is the class of the PMC.  Field 'ROW INDEX' is the
407  * row index for the PMC.
408  *
409  * The 'ROW INDEX' ranges over 0..NWPMCS where NHWPMCS is the total
410  * number of hardware PMCs on this cpu.
411  */
412 
413 #define	PMC_ID_TO_ROWINDEX(ID)	((ID) & 0xFF)
414 #define	PMC_ID_TO_CLASS(ID)	(((ID) & 0xFF00) >> 8)
415 #define	PMC_ID_TO_MODE(ID)	(((ID) & 0xF0000) >> 16)
416 #define	PMC_ID_TO_CPU(ID)	(((ID) & 0xFFF00000) >> 20)
417 #define	PMC_ID_MAKE_ID(CPU,MODE,CLASS,ROWINDEX)			\
418 	((((CPU) & 0xFFF) << 20) | (((MODE) & 0xF) << 16) |	\
419 	(((CLASS) & 0xFF) << 8) | ((ROWINDEX) & 0xFF))
420 
421 /*
422  * Data structures for system calls supported by the pmc driver.
423  */
424 
425 /*
426  * OP PMCALLOCATE
427  *
428  * Allocate a PMC on the named CPU.
429  */
430 
431 #define	PMC_CPU_ANY	~0
432 
433 struct pmc_op_pmcallocate {
434 	uint32_t	pm_caps;	/* PMC_CAP_* */
435 	uint32_t	pm_cpu;		/* CPU number or PMC_CPU_ANY */
436 	enum pmc_class	pm_class;	/* class of PMC desired */
437 	enum pmc_event	pm_ev;		/* [enum pmc_event] desired */
438 	uint32_t	pm_flags;	/* additional modifiers PMC_F_* */
439 	enum pmc_mode	pm_mode;	/* desired mode */
440 	pmc_id_t	pm_pmcid;	/* [return] process pmc id */
441 	pmc_value_t	pm_count;	/* initial/sample count */
442 
443 	union pmc_md_op_pmcallocate pm_md; /* MD layer extensions */
444 };
445 
446 /*
447  * OP PMCADMIN
448  *
449  * Set the administrative state (i.e., whether enabled or disabled) of
450  * a PMC 'pm_pmc' on CPU 'pm_cpu'.  Note that 'pm_pmc' specifies an
451  * absolute PMC number and need not have been first allocated by the
452  * calling process.
453  */
454 
455 struct pmc_op_pmcadmin {
456 	int		pm_cpu;		/* CPU# */
457 	uint32_t	pm_flags;	/* flags */
458 	int		pm_pmc;         /* PMC# */
459 	enum pmc_state  pm_state;	/* desired state */
460 };
461 
462 /*
463  * OP PMCATTACH / OP PMCDETACH
464  *
465  * Attach/detach a PMC and a process.
466  */
467 
468 struct pmc_op_pmcattach {
469 	pmc_id_t	pm_pmc;		/* PMC to attach to */
470 	pid_t		pm_pid;		/* target process */
471 };
472 
473 /*
474  * OP PMCSETCOUNT
475  *
476  * Set the sampling rate (i.e., the reload count) for statistical counters.
477  * 'pm_pmcid' need to have been previously allocated using PMCALLOCATE.
478  */
479 
480 struct pmc_op_pmcsetcount {
481 	pmc_value_t	pm_count;	/* initial/sample count */
482 	pmc_id_t	pm_pmcid;	/* PMC id to set */
483 };
484 
485 /*
486  * OP PMCRW
487  *
488  * Read the value of a PMC named by 'pm_pmcid'.  'pm_pmcid' needs
489  * to have been previously allocated using PMCALLOCATE.
490  */
491 
492 struct pmc_op_pmcrw {
493 	uint32_t	pm_flags;	/* PMC_F_{OLD,NEW}VALUE*/
494 	pmc_id_t	pm_pmcid;	/* pmc id */
495 	pmc_value_t	pm_value;	/* new&returned value */
496 };
497 
498 /*
499  * OP GETPMCINFO
500  *
501  * retrieve PMC state for a named CPU.  The caller is expected to
502  * allocate 'npmc' * 'struct pmc_info' bytes of space for the return
503  * values.
504  */
505 
506 struct pmc_info {
507 	char		pm_name[PMC_NAME_MAX]; /* pmc name */
508 	enum pmc_class	pm_class;	/* enum pmc_class */
509 	int		pm_enabled;	/* whether enabled */
510 	enum pmc_disp	pm_rowdisp;	/* FREE, THREAD or STANDLONE */
511 	pid_t		pm_ownerpid;	/* owner, or -1 */
512 	enum pmc_mode	pm_mode;	/* current mode [enum pmc_mode] */
513 	enum pmc_event	pm_event;	/* current event */
514 	uint32_t	pm_flags;	/* current flags */
515 	pmc_value_t	pm_reloadcount;	/* sampling counters only */
516 };
517 
518 struct pmc_op_getpmcinfo {
519 	int32_t		pm_cpu;		/* 0 <= cpu < mp_maxid */
520 	struct pmc_info	pm_pmcs[];	/* space for 'npmc' structures */
521 };
522 
523 /*
524  * OP GETCPUINFO
525  *
526  * Retrieve system CPU information.
527  */
528 
529 struct pmc_classinfo {
530 	enum pmc_class	pm_class;	/* class id */
531 	uint32_t	pm_caps;	/* counter capabilities */
532 	uint32_t	pm_width;	/* width of the PMC */
533 	uint32_t	pm_num;		/* number of PMCs in class */
534 };
535 
536 struct pmc_op_getcpuinfo {
537 	enum pmc_cputype pm_cputype; /* what kind of CPU */
538 	uint32_t	pm_ncpu;    /* max CPU number */
539 	uint32_t	pm_npmc;    /* #PMCs per CPU */
540 	uint32_t	pm_nclass;  /* #classes of PMCs */
541 	struct pmc_classinfo  pm_classes[PMC_CLASS_MAX];
542 };
543 
544 /*
545  * OP CONFIGURELOG
546  *
547  * Configure a log file for writing system-wide statistics to.
548  */
549 
550 struct pmc_op_configurelog {
551 	int		pm_flags;
552 	int		pm_logfd;   /* logfile fd (or -1) */
553 };
554 
555 /*
556  * OP GETDRIVERSTATS
557  *
558  * Retrieve pmc(4) driver-wide statistics.
559  */
560 #ifdef _KERNEL
561 struct pmc_driverstats {
562 	counter_u64_t	pm_intr_ignored;	/* #interrupts ignored */
563 	counter_u64_t	pm_intr_processed;	/* #interrupts processed */
564 	counter_u64_t	pm_intr_bufferfull;	/* #interrupts with ENOSPC */
565 	counter_u64_t	pm_syscalls;		/* #syscalls */
566 	counter_u64_t	pm_syscall_errors;	/* #syscalls with errors */
567 	counter_u64_t	pm_buffer_requests;	/* #buffer requests */
568 	counter_u64_t	pm_buffer_requests_failed; /* #failed buffer requests */
569 	counter_u64_t	pm_log_sweeps;		/* #sample buffer processing
570 						   passes */
571 	counter_u64_t	pm_merges;		/* merged k+u */
572 	counter_u64_t	pm_overwrites;		/* UR overwrites */
573 };
574 #endif
575 
576 struct pmc_op_getdriverstats {
577 	unsigned int	pm_intr_ignored;	/* #interrupts ignored */
578 	unsigned int	pm_intr_processed;	/* #interrupts processed */
579 	unsigned int	pm_intr_bufferfull;	/* #interrupts with ENOSPC */
580 	unsigned int	pm_syscalls;		/* #syscalls */
581 	unsigned int	pm_syscall_errors;	/* #syscalls with errors */
582 	unsigned int	pm_buffer_requests;	/* #buffer requests */
583 	unsigned int	pm_buffer_requests_failed; /* #failed buffer requests */
584 	unsigned int	pm_log_sweeps;		/* #sample buffer processing
585 						   passes */
586 };
587 
588 /*
589  * OP RELEASE / OP START / OP STOP
590  *
591  * Simple operations on a PMC id.
592  */
593 
594 struct pmc_op_simple {
595 	pmc_id_t	pm_pmcid;
596 };
597 
598 /*
599  * OP WRITELOG
600  *
601  * Flush the current log buffer and write 4 bytes of user data to it.
602  */
603 
604 struct pmc_op_writelog {
605 	uint32_t	pm_userdata;
606 };
607 
608 /*
609  * OP GETMSR
610  *
611  * Retrieve the machine specific address associated with the allocated
612  * PMC.  This number can be used subsequently with a read-performance-counter
613  * instruction.
614  */
615 
616 struct pmc_op_getmsr {
617 	uint32_t	pm_msr;		/* machine specific address */
618 	pmc_id_t	pm_pmcid;	/* allocated pmc id */
619 };
620 
621 /*
622  * OP GETDYNEVENTINFO
623  *
624  * Retrieve a PMC dynamic class events list.
625  */
626 
627 struct pmc_dyn_event_descr {
628 	char		pm_ev_name[PMC_NAME_MAX];
629 	enum pmc_event	pm_ev_code;
630 };
631 
632 struct pmc_op_getdyneventinfo {
633 	enum pmc_class			pm_class;
634 	unsigned int			pm_nevent;
635 	struct pmc_dyn_event_descr	pm_events[PMC_EV_DYN_COUNT];
636 };
637 
638 #ifdef _KERNEL
639 
640 #include <sys/malloc.h>
641 #include <sys/sysctl.h>
642 #include <sys/_cpuset.h>
643 
644 #include <machine/frame.h>
645 
646 #define	PMC_HASH_SIZE				1024
647 #define	PMC_MTXPOOL_SIZE			2048
648 #define	PMC_LOG_BUFFER_SIZE			256
649 #define	PMC_NLOGBUFFERS_PCPU			32
650 #define	PMC_NSAMPLES				256
651 #define	PMC_CALLCHAIN_DEPTH			128
652 #define	PMC_THREADLIST_MAX			128
653 
654 #define PMC_SYSCTL_NAME_PREFIX "kern." PMC_MODULE_NAME "."
655 
656 /*
657  * Locking keys
658  *
659  * (b) - pmc_bufferlist_mtx (spin lock)
660  * (k) - pmc_kthread_mtx (sleep lock)
661  * (o) - po->po_mtx (spin lock)
662  * (g) - global_epoch_preempt (epoch)
663  * (p) - pmc_sx (sx)
664  */
665 
666 /*
667  * PMC commands
668  */
669 
670 struct pmc_syscall_args {
671 	register_t	pmop_code;	/* one of PMC_OP_* */
672 	void		*pmop_data;	/* syscall parameter */
673 };
674 
675 /*
676  * Interface to processor specific s1tuff
677  */
678 
679 /*
680  * struct pmc_descr
681  *
682  * Machine independent (i.e., the common parts) of a human readable
683  * PMC description.
684  */
685 
686 struct pmc_descr {
687 	char		pd_name[PMC_NAME_MAX]; /* name */
688 	uint32_t	pd_caps;	/* capabilities */
689 	enum pmc_class	pd_class;	/* class of the PMC */
690 	uint32_t	pd_width;	/* width in bits */
691 };
692 
693 /*
694  * struct pmc_target
695  *
696  * This structure records all the target processes associated with a
697  * PMC.
698  */
699 
700 struct pmc_target {
701 	LIST_ENTRY(pmc_target)	pt_next;
702 	struct pmc_process	*pt_process; /* target descriptor */
703 };
704 
705 /*
706  * struct pmc
707  *
708  * Describes each allocated PMC.
709  *
710  * Each PMC has precisely one owner, namely the process that allocated
711  * the PMC.
712  *
713  * A PMC may be attached to multiple target processes.  The
714  * 'pm_targets' field links all the target processes being monitored
715  * by this PMC.
716  *
717  * The 'pm_savedvalue' field is protected by a mutex.
718  *
719  * On a multi-cpu machine, multiple target threads associated with a
720  * process-virtual PMC could be concurrently executing on different
721  * CPUs.  The 'pm_runcount' field is atomically incremented every time
722  * the PMC gets scheduled on a CPU and atomically decremented when it
723  * get descheduled.  Deletion of a PMC is only permitted when this
724  * field is '0'.
725  *
726  */
727 struct pmc_pcpu_state {
728 	uint32_t pps_overflowcnt;	/* count overflow interrupts */
729 	uint8_t pps_stalled;
730 	uint8_t pps_cpustate;
731 } __aligned(CACHE_LINE_SIZE);
732 struct pmc {
733 	LIST_HEAD(,pmc_target)	pm_targets;	/* list of target processes */
734 	LIST_ENTRY(pmc)		pm_next;	/* owner's list */
735 
736 	/*
737 	 * System-wide PMCs are allocated on a CPU and are not moved
738 	 * around.  For system-wide PMCs we record the CPU the PMC was
739 	 * allocated on in the 'CPU' field of the pmc ID.
740 	 *
741 	 * Virtual PMCs run on whichever CPU is currently executing
742 	 * their targets' threads.  For these PMCs we need to save
743 	 * their current PMC counter values when they are taken off
744 	 * CPU.
745 	 */
746 
747 	union {
748 		pmc_value_t	pm_savedvalue;	/* Virtual PMCS */
749 	} pm_gv;
750 
751 	/*
752 	 * For sampling mode PMCs, we keep track of the PMC's "reload
753 	 * count", which is the counter value to be loaded in when
754 	 * arming the PMC for the next counting session.  For counting
755 	 * modes on PMCs that are read-only (e.g., the x86 TSC), we
756 	 * keep track of the initial value at the start of
757 	 * counting-mode operation.
758 	 */
759 
760 	union {
761 		pmc_value_t	pm_reloadcount;	/* sampling PMC modes */
762 		pmc_value_t	pm_initial;	/* counting PMC modes */
763 	} pm_sc;
764 
765 	struct pmc_pcpu_state *pm_pcpu_state;
766 	volatile cpuset_t pm_cpustate;	/* CPUs where PMC should be active */
767 	uint32_t	pm_caps;	/* PMC capabilities */
768 	enum pmc_event	pm_event;	/* event being measured */
769 	uint32_t	pm_flags;	/* additional flags PMC_F_... */
770 	struct pmc_owner *pm_owner;	/* owner thread state */
771 	counter_u64_t		pm_runcount;	/* #cpus currently on */
772 	enum pmc_state	pm_state;	/* current PMC state */
773 
774 	/*
775 	 * The PMC ID field encodes the row-index for the PMC, its
776 	 * mode, class and the CPU# associated with the PMC.
777 	 */
778 
779 	pmc_id_t	pm_id;		/* allocated PMC id */
780 	enum pmc_class pm_class;
781 
782 	/* md extensions */
783 	union pmc_md_pmc	pm_md;
784 };
785 
786 /*
787  * Accessor macros for 'struct pmc'
788  */
789 
790 #define	PMC_TO_MODE(P)		PMC_ID_TO_MODE((P)->pm_id)
791 #define	PMC_TO_CLASS(P)		PMC_ID_TO_CLASS((P)->pm_id)
792 #define	PMC_TO_ROWINDEX(P)	PMC_ID_TO_ROWINDEX((P)->pm_id)
793 #define	PMC_TO_CPU(P)		PMC_ID_TO_CPU((P)->pm_id)
794 
795 /*
796  * struct pmc_threadpmcstate
797  *
798  * Record per-PMC, per-thread state.
799  */
800 struct pmc_threadpmcstate {
801 	pmc_value_t	pt_pmcval;	/* per-thread reload count */
802 };
803 
804 /*
805  * struct pmc_thread
806  *
807  * Record a 'target' thread being profiled.
808  */
809 struct pmc_thread {
810 	LIST_ENTRY(pmc_thread) pt_next;		/* linked list */
811 	struct thread	*pt_td;			/* target thread */
812 	struct pmc_threadpmcstate pt_pmcs[];	/* per-PMC state */
813 };
814 
815 /*
816  * struct pmc_process
817  *
818  * Record a 'target' process being profiled.
819  *
820  * The target process being profiled could be different from the owner
821  * process which allocated the PMCs.  Each target process descriptor
822  * is associated with NHWPMC 'struct pmc *' pointers.  Each PMC at a
823  * given hardware row-index 'n' will use slot 'n' of the 'pp_pmcs[]'
824  * array.  The size of this structure is thus PMC architecture
825  * dependent.
826  *
827  */
828 
829 struct pmc_targetstate {
830 	struct pmc	*pp_pmc;   /* target PMC */
831 	pmc_value_t	pp_pmcval; /* per-process value */
832 };
833 
834 struct pmc_process {
835 	LIST_ENTRY(pmc_process) pp_next;	/* hash chain */
836 	LIST_HEAD(,pmc_thread) pp_tds;		/* list of threads */
837 	struct mtx	*pp_tdslock;		/* lock on pp_tds thread list */
838 	int		pp_refcnt;		/* reference count */
839 	uint32_t	pp_flags;		/* flags PMC_PP_* */
840 	struct proc	*pp_proc;		/* target process */
841 	struct pmc_targetstate pp_pmcs[];       /* NHWPMCs */
842 };
843 
844 #define	PMC_PP_ENABLE_MSR_ACCESS	0x00000001
845 
846 /*
847  * struct pmc_owner
848  *
849  * We associate a PMC with an 'owner' process.
850  *
851  * A process can be associated with 0..NCPUS*NHWPMC PMCs during its
852  * lifetime, where NCPUS is the numbers of CPUS in the system and
853  * NHWPMC is the number of hardware PMCs per CPU.  These are
854  * maintained in the list headed by the 'po_pmcs' to save on space.
855  *
856  */
857 
858 struct pmc_owner  {
859 	LIST_ENTRY(pmc_owner)	po_next;	/* hash chain */
860 	CK_LIST_ENTRY(pmc_owner)	po_ssnext;	/* (g/p) list of SS PMC owners */
861 	LIST_HEAD(, pmc)	po_pmcs;	/* owned PMC list */
862 	TAILQ_HEAD(, pmclog_buffer) po_logbuffers; /* (o) logbuffer list */
863 	struct mtx		po_mtx;		/* spin lock for (o) */
864 	struct proc		*po_owner;	/* owner proc */
865 	uint32_t		po_flags;	/* (k) flags PMC_PO_* */
866 	struct proc		*po_kthread;	/* (k) helper kthread */
867 	struct file		*po_file;	/* file reference */
868 	int			po_error;	/* recorded error */
869 	short			po_sscount;	/* # SS PMCs owned */
870 	short			po_logprocmaps;	/* global mappings done */
871 	struct pmclog_buffer	*po_curbuf[MAXCPU];	/* current log buffer */
872 };
873 
874 #define	PMC_PO_OWNS_LOGFILE		0x00000001 /* has a log file */
875 #define	PMC_PO_SHUTDOWN			0x00000010 /* in the process of shutdown */
876 #define	PMC_PO_INITIAL_MAPPINGS_DONE	0x00000020
877 
878 /*
879  * struct pmc_hw -- describe the state of the PMC hardware
880  *
881  * When in use, a HW PMC is associated with one allocated 'struct pmc'
882  * pointed to by field 'phw_pmc'.  When inactive, this field is NULL.
883  *
884  * On an SMP box, one or more HW PMC's in process virtual mode with
885  * the same 'phw_pmc' could be executing on different CPUs.  In order
886  * to handle this case correctly, we need to ensure that only
887  * incremental counts get added to the saved value in the associated
888  * 'struct pmc'.  The 'phw_save' field is used to keep the saved PMC
889  * value at the time the hardware is started during this context
890  * switch (i.e., the difference between the new (hardware) count and
891  * the saved count is atomically added to the count field in 'struct
892  * pmc' at context switch time).
893  *
894  */
895 
896 struct pmc_hw {
897 	uint32_t	phw_state;	/* see PHW_* macros below */
898 	struct pmc	*phw_pmc;	/* current thread PMC */
899 };
900 
901 #define	PMC_PHW_RI_MASK		0x000000FF
902 #define	PMC_PHW_CPU_SHIFT	8
903 #define	PMC_PHW_CPU_MASK	0x0000FF00
904 #define	PMC_PHW_FLAGS_SHIFT	16
905 #define	PMC_PHW_FLAGS_MASK	0xFFFF0000
906 
907 #define	PMC_PHW_INDEX_TO_STATE(ri)	((ri) & PMC_PHW_RI_MASK)
908 #define	PMC_PHW_STATE_TO_INDEX(state)	((state) & PMC_PHW_RI_MASK)
909 #define	PMC_PHW_CPU_TO_STATE(cpu)	(((cpu) << PMC_PHW_CPU_SHIFT) & \
910 	PMC_PHW_CPU_MASK)
911 #define	PMC_PHW_STATE_TO_CPU(state)	(((state) & PMC_PHW_CPU_MASK) >> \
912 	PMC_PHW_CPU_SHIFT)
913 #define	PMC_PHW_FLAGS_TO_STATE(flags)	(((flags) << PMC_PHW_FLAGS_SHIFT) & \
914 	PMC_PHW_FLAGS_MASK)
915 #define	PMC_PHW_STATE_TO_FLAGS(state)	(((state) & PMC_PHW_FLAGS_MASK) >> \
916 	PMC_PHW_FLAGS_SHIFT)
917 #define	PMC_PHW_FLAG_IS_ENABLED		(PMC_PHW_FLAGS_TO_STATE(0x01))
918 #define	PMC_PHW_FLAG_IS_SHAREABLE	(PMC_PHW_FLAGS_TO_STATE(0x02))
919 
920 /*
921  * struct pmc_sample
922  *
923  * Space for N (tunable) PC samples and associated control data.
924  */
925 
926 struct pmc_sample {
927 	uint16_t		ps_nsamples;	/* callchain depth */
928 	uint16_t		ps_nsamples_actual;
929 	uint16_t		ps_cpu;		/* cpu number */
930 	uint16_t		ps_flags;	/* other flags */
931 	lwpid_t			ps_tid;		/* thread id */
932 	pid_t			ps_pid;		/* process PID or -1 */
933 	int		ps_ticks; /* ticks at sample time */
934 	/* pad */
935 	struct thread		*ps_td;		/* which thread */
936 	struct pmc		*ps_pmc;	/* interrupting PMC */
937 	uintptr_t		*ps_pc;		/* (const) callchain start */
938 	uint64_t		ps_tsc;		/* tsc value */
939 };
940 
941 #define 	PMC_SAMPLE_FREE		((uint16_t) 0)
942 #define 	PMC_USER_CALLCHAIN_PENDING	((uint16_t) 0xFFFF)
943 
944 struct pmc_samplebuffer {
945 	volatile uint64_t		ps_prodidx; /* producer index */
946 	volatile uint64_t		ps_considx; /* consumer index */
947 	uintptr_t		*ps_callchains;	/* all saved call chains */
948 	struct pmc_sample	ps_samples[];	/* array of sample entries */
949 };
950 
951 #define PMC_CONS_SAMPLE(psb)					\
952 	(&(psb)->ps_samples[(psb)->ps_considx & pmc_sample_mask])
953 
954 #define PMC_CONS_SAMPLE_OFF(psb, off)							\
955 	(&(psb)->ps_samples[(off) & pmc_sample_mask])
956 
957 #define PMC_PROD_SAMPLE(psb)					\
958 	(&(psb)->ps_samples[(psb)->ps_prodidx & pmc_sample_mask])
959 
960 /*
961  * struct pmc_cpustate
962  *
963  * A CPU is modelled as a collection of HW PMCs with space for additional
964  * flags.
965  */
966 
967 struct pmc_cpu {
968 	uint32_t	pc_state;	/* physical cpu number + flags */
969 	struct pmc_samplebuffer *pc_sb[3]; /* space for samples */
970 	struct pmc_hw	*pc_hwpmcs[];	/* 'npmc' pointers */
971 };
972 
973 #define	PMC_PCPU_CPU_MASK		0x000000FF
974 #define	PMC_PCPU_FLAGS_MASK		0xFFFFFF00
975 #define	PMC_PCPU_FLAGS_SHIFT		8
976 #define	PMC_PCPU_STATE_TO_CPU(S)	((S) & PMC_PCPU_CPU_MASK)
977 #define	PMC_PCPU_STATE_TO_FLAGS(S)	(((S) & PMC_PCPU_FLAGS_MASK) >> PMC_PCPU_FLAGS_SHIFT)
978 #define	PMC_PCPU_FLAGS_TO_STATE(F)	(((F) << PMC_PCPU_FLAGS_SHIFT) & PMC_PCPU_FLAGS_MASK)
979 #define	PMC_PCPU_CPU_TO_STATE(C)	((C) & PMC_PCPU_CPU_MASK)
980 #define	PMC_PCPU_FLAG_HTT		(PMC_PCPU_FLAGS_TO_STATE(0x1))
981 
982 /*
983  * struct pmc_binding
984  *
985  * CPU binding information.
986  */
987 
988 struct pmc_binding {
989 	int	pb_bound;	/* is bound? */
990 	int	pb_cpu;		/* if so, to which CPU */
991 	u_char	pb_priority;	/* Thread active priority. */
992 };
993 
994 struct pmc_mdep;
995 
996 /*
997  * struct pmc_classdep
998  *
999  * PMC class-dependent operations.
1000  */
1001 struct pmc_classdep {
1002 	uint32_t	pcd_caps;	/* class capabilities */
1003 	enum pmc_class	pcd_class;	/* class id */
1004 	int		pcd_num;	/* number of PMCs */
1005 	int		pcd_ri;		/* row index of the first PMC in class */
1006 	int		pcd_width;	/* width of the PMC */
1007 
1008 	/* configuring/reading/writing the hardware PMCs */
1009 	int (*pcd_config_pmc)(int _cpu, int _ri, struct pmc *_pm);
1010 	int (*pcd_get_config)(int _cpu, int _ri, struct pmc **_ppm);
1011 	int (*pcd_read_pmc)(int _cpu, int _ri, struct pmc *_pm,
1012 	    pmc_value_t *_value);
1013 	int (*pcd_write_pmc)(int _cpu, int _ri, struct pmc *_pm,
1014 	    pmc_value_t _value);
1015 
1016 	/* pmc allocation/release */
1017 	int (*pcd_allocate_pmc)(int _cpu, int _ri, struct pmc *_t,
1018 		const struct pmc_op_pmcallocate *_a);
1019 	int (*pcd_release_pmc)(int _cpu, int _ri, struct pmc *_pm);
1020 
1021 	/* starting and stopping PMCs */
1022 	int (*pcd_start_pmc)(int _cpu, int _ri, struct pmc *_pm);
1023 	int (*pcd_stop_pmc)(int _cpu, int _ri, struct pmc *_pm);
1024 
1025 	/* description */
1026 	int (*pcd_describe)(int _cpu, int _ri, struct pmc_info *_pi,
1027 		struct pmc **_ppmc);
1028 
1029 	/* class-dependent initialization & finalization */
1030 	int (*pcd_pcpu_init)(struct pmc_mdep *_md, int _cpu);
1031 	int (*pcd_pcpu_fini)(struct pmc_mdep *_md, int _cpu);
1032 
1033 	/* machine-specific interface */
1034 	int (*pcd_get_msr)(int _ri, uint32_t *_msr);
1035 };
1036 
1037 /*
1038  * struct pmc_mdep
1039  *
1040  * Machine dependent bits needed per CPU type.
1041  */
1042 
1043 struct pmc_mdep  {
1044 	uint32_t	pmd_cputype;    /* from enum pmc_cputype */
1045 	uint32_t	pmd_npmc;	/* number of PMCs per CPU */
1046 	uint32_t	pmd_nclass;	/* number of PMC classes present */
1047 
1048 	/*
1049 	 * Machine dependent methods.
1050 	 */
1051 
1052 	/* thread context switch in/out */
1053 	int (*pmd_switch_in)(struct pmc_cpu *_p, struct pmc_process *_pp);
1054 	int (*pmd_switch_out)(struct pmc_cpu *_p, struct pmc_process *_pp);
1055 
1056 	/* handle a PMC interrupt */
1057 	int (*pmd_intr)(struct trapframe *_tf);
1058 
1059 	/*
1060 	 * PMC class dependent information.
1061 	 */
1062 	struct pmc_classdep pmd_classdep[];
1063 };
1064 
1065 /*
1066  * Per-CPU state.  This is an array of 'mp_ncpu' pointers
1067  * to struct pmc_cpu descriptors.
1068  */
1069 
1070 extern struct pmc_cpu **pmc_pcpu;
1071 
1072 /* driver statistics */
1073 extern struct pmc_driverstats pmc_stats;
1074 
1075 #if	defined(HWPMC_DEBUG)
1076 
1077 /* HWPMC_DEBUG without KTR will compile but is a no-op. */
1078 #if !defined(KTR) || !defined(KTR_COMPILE) || ((KTR_COMPILE & KTR_SUBSYS) == 0)
1079 #error "HWPMC_DEBUG requires KTR and KTR_COMPILE=KTR_SUBSYS -- see ktr(4)"
1080 #endif
1081 
1082 #include <sys/ktr.h>
1083 
1084 #define	__pmcdbg_used		/* unused variable annotation */
1085 
1086 /*
1087  * Debug flags, major flag groups.
1088  *
1089  * Please keep the DEBUGGING section of the hwpmc(4) man page in sync.
1090  */
1091 struct pmc_debugflags {
1092 	int	pdb_CPU;
1093 	int	pdb_CSW;
1094 	int	pdb_LOG;
1095 	int	pdb_MDP;
1096 	int	pdb_MOD;
1097 	int	pdb_OWN;
1098 	int	pdb_PMC;
1099 	int	pdb_PRC;
1100 	int	pdb_SAM;
1101 };
1102 
1103 extern struct pmc_debugflags pmc_debugflags;
1104 
1105 #define	KTR_PMC			KTR_SUBSYS
1106 
1107 #define	PMC_DEBUG_STRSIZE		128
1108 #define	PMC_DEBUG_DEFAULT_FLAGS		{ 0, 0, 0, 0, 0, 0, 0, 0, 0 }
1109 
1110 #define	PMCDBG0(M, N, L, F) do {					\
1111 	if (pmc_debugflags.pdb_ ## M & (1 << PMC_DEBUG_MIN_ ## N))	\
1112 		CTR0(KTR_PMC, #M ":" #N ":" #L  ": " F);		\
1113 } while (0)
1114 #define	PMCDBG1(M, N, L, F, p1) do {					\
1115 	if (pmc_debugflags.pdb_ ## M & (1 << PMC_DEBUG_MIN_ ## N))	\
1116 		CTR1(KTR_PMC, #M ":" #N ":" #L  ": " F, p1);		\
1117 } while (0)
1118 #define	PMCDBG2(M, N, L, F, p1, p2) do {				\
1119 	if (pmc_debugflags.pdb_ ## M & (1 << PMC_DEBUG_MIN_ ## N))	\
1120 		CTR2(KTR_PMC, #M ":" #N ":" #L  ": " F, p1, p2);	\
1121 } while (0)
1122 #define	PMCDBG3(M, N, L, F, p1, p2, p3) do {				\
1123 	if (pmc_debugflags.pdb_ ## M & (1 << PMC_DEBUG_MIN_ ## N))	\
1124 		CTR3(KTR_PMC, #M ":" #N ":" #L  ": " F, p1, p2, p3);	\
1125 } while (0)
1126 #define	PMCDBG4(M, N, L, F, p1, p2, p3, p4) do {			\
1127 	if (pmc_debugflags.pdb_ ## M & (1 << PMC_DEBUG_MIN_ ## N))	\
1128 		CTR4(KTR_PMC, #M ":" #N ":" #L  ": " F, p1, p2, p3, p4);\
1129 } while (0)
1130 #define	PMCDBG5(M, N, L, F, p1, p2, p3, p4, p5) do {			\
1131 	if (pmc_debugflags.pdb_ ## M & (1 << PMC_DEBUG_MIN_ ## N))	\
1132 		CTR5(KTR_PMC, #M ":" #N ":" #L  ": " F, p1, p2, p3, p4,	\
1133 		    p5);						\
1134 } while (0)
1135 #define	PMCDBG6(M, N, L, F, p1, p2, p3, p4, p5, p6) do {		\
1136 	if (pmc_debugflags.pdb_ ## M & (1 << PMC_DEBUG_MIN_ ## N))	\
1137 		CTR6(KTR_PMC, #M ":" #N ":" #L  ": " F, p1, p2, p3, p4,	\
1138 		    p5, p6);						\
1139 } while (0)
1140 
1141 /* Major numbers */
1142 #define	PMC_DEBUG_MAJ_CPU		0 /* cpu switches */
1143 #define	PMC_DEBUG_MAJ_CSW		1 /* context switches */
1144 #define	PMC_DEBUG_MAJ_LOG		2 /* logging */
1145 #define	PMC_DEBUG_MAJ_MDP		3 /* machine dependent */
1146 #define	PMC_DEBUG_MAJ_MOD		4 /* misc module infrastructure */
1147 #define	PMC_DEBUG_MAJ_OWN		5 /* owner */
1148 #define	PMC_DEBUG_MAJ_PMC		6 /* pmc management */
1149 #define	PMC_DEBUG_MAJ_PRC		7 /* processes */
1150 #define	PMC_DEBUG_MAJ_SAM		8 /* sampling */
1151 
1152 /* Minor numbers */
1153 
1154 /* Common (8 bits) */
1155 #define	PMC_DEBUG_MIN_ALL		0 /* allocation */
1156 #define	PMC_DEBUG_MIN_REL		1 /* release */
1157 #define	PMC_DEBUG_MIN_OPS		2 /* ops: start, stop, ... */
1158 #define	PMC_DEBUG_MIN_INI		3 /* init */
1159 #define	PMC_DEBUG_MIN_FND		4 /* find */
1160 
1161 /* MODULE */
1162 #define	PMC_DEBUG_MIN_PMH	       14 /* pmc_hook */
1163 #define	PMC_DEBUG_MIN_PMS	       15 /* pmc_syscall */
1164 
1165 /* OWN */
1166 #define	PMC_DEBUG_MIN_ORM		8 /* owner remove */
1167 #define	PMC_DEBUG_MIN_OMR		9 /* owner maybe remove */
1168 
1169 /* PROCESSES */
1170 #define	PMC_DEBUG_MIN_TLK		8 /* link target */
1171 #define	PMC_DEBUG_MIN_TUL		9 /* unlink target */
1172 #define	PMC_DEBUG_MIN_EXT	       10 /* process exit */
1173 #define	PMC_DEBUG_MIN_EXC	       11 /* process exec */
1174 #define	PMC_DEBUG_MIN_FRK	       12 /* process fork */
1175 #define	PMC_DEBUG_MIN_ATT	       13 /* attach/detach */
1176 #define	PMC_DEBUG_MIN_SIG	       14 /* signalling */
1177 
1178 /* CONTEXT SWITCHES */
1179 #define	PMC_DEBUG_MIN_SWI		8 /* switch in */
1180 #define	PMC_DEBUG_MIN_SWO		9 /* switch out */
1181 
1182 /* PMC */
1183 #define	PMC_DEBUG_MIN_REG		8 /* pmc register */
1184 #define	PMC_DEBUG_MIN_ALR		9 /* allocate row */
1185 
1186 /* MACHINE DEPENDENT LAYER */
1187 #define	PMC_DEBUG_MIN_REA		8 /* read */
1188 #define	PMC_DEBUG_MIN_WRI		9 /* write */
1189 #define	PMC_DEBUG_MIN_CFG	       10 /* config */
1190 #define	PMC_DEBUG_MIN_STA	       11 /* start */
1191 #define	PMC_DEBUG_MIN_STO	       12 /* stop */
1192 #define	PMC_DEBUG_MIN_INT	       13 /* interrupts */
1193 
1194 /* CPU */
1195 #define	PMC_DEBUG_MIN_BND		8 /* bind */
1196 #define	PMC_DEBUG_MIN_SEL		9 /* select */
1197 
1198 /* LOG */
1199 #define	PMC_DEBUG_MIN_GTB		8 /* get buf */
1200 #define	PMC_DEBUG_MIN_SIO		9 /* schedule i/o */
1201 #define	PMC_DEBUG_MIN_FLS	       10 /* flush */
1202 #define	PMC_DEBUG_MIN_SAM	       11 /* sample */
1203 #define	PMC_DEBUG_MIN_CLO	       12 /* close */
1204 
1205 #else
1206 #define	__pmcdbg_used			__unused
1207 #define	PMCDBG0(M, N, L, F)		/* nothing */
1208 #define	PMCDBG1(M, N, L, F, p1)
1209 #define	PMCDBG2(M, N, L, F, p1, p2)
1210 #define	PMCDBG3(M, N, L, F, p1, p2, p3)
1211 #define	PMCDBG4(M, N, L, F, p1, p2, p3, p4)
1212 #define	PMCDBG5(M, N, L, F, p1, p2, p3, p4, p5)
1213 #define	PMCDBG6(M, N, L, F, p1, p2, p3, p4, p5, p6)
1214 #endif
1215 
1216 /* declare a dedicated memory pool */
1217 MALLOC_DECLARE(M_PMC);
1218 
1219 /*
1220  * Functions
1221  */
1222 
1223 struct pmc_mdep *pmc_md_initialize(void);	/* MD init function */
1224 void	pmc_md_finalize(struct pmc_mdep *_md);	/* MD fini function */
1225 int	pmc_getrowdisp(int _ri);
1226 int	pmc_process_interrupt(int _ring, struct pmc *_pm, struct trapframe *_tf);
1227 int	pmc_save_kernel_callchain(uintptr_t *_cc, int _maxsamples,
1228     struct trapframe *_tf);
1229 int	pmc_save_user_callchain(uintptr_t *_cc, int _maxsamples,
1230     struct trapframe *_tf);
1231 void	pmc_restore_cpu_binding(struct pmc_binding *pb);
1232 void	pmc_save_cpu_binding(struct pmc_binding *pb);
1233 void	pmc_select_cpu(int cpu);
1234 struct pmc_mdep *pmc_mdep_alloc(int nclasses);
1235 void pmc_mdep_free(struct pmc_mdep *md);
1236 uint64_t pmc_rdtsc(void);
1237 #endif /* _KERNEL */
1238 #endif /* _SYS_PMC_H_ */
1239