1/*-
2 * Copyright (c) 2003-2008, Joseph Koshy
3 * Copyright (c) 2007 The FreeBSD Foundation
4 * All rights reserved.
5 *
6 * Portions of this software were developed by A. Joseph Koshy under
7 * sponsorship from the FreeBSD Foundation and Google, Inc.
8 *
9 * Redistribution and use in source and binary forms, with or without
10 * modification, are permitted provided that the following conditions
11 * are met:
12 * 1. Redistributions of source code must retain the above copyright
13 *    notice, this list of conditions and the following disclaimer.
14 * 2. Redistributions in binary form must reproduce the above copyright
15 *    notice, this list of conditions and the following disclaimer in the
16 *    documentation and/or other materials provided with the distribution.
17 *
18 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
19 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
20 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
21 * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
22 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
23 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
24 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
25 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28 * SUCH DAMAGE.
29 *
30 * $FreeBSD$
31 */
32
33#ifndef _SYS_PMC_H_
34#define	_SYS_PMC_H_
35
36#include <dev/hwpmc/pmc_events.h>
37
38#include <machine/pmc_mdep.h>
39#include <machine/profile.h>
40
41#define	PMC_MODULE_NAME		"hwpmc"
42#define	PMC_NAME_MAX		64 /* HW counter name size */
43#define	PMC_CLASS_MAX		8  /* max #classes of PMCs per-system */
44
45/*
46 * Kernel<->userland API version number [MMmmpppp]
47 *
48 * Major numbers are to be incremented when an incompatible change to
49 * the ABI occurs that older clients will not be able to handle.
50 *
51 * Minor numbers are incremented when a backwards compatible change
52 * occurs that allows older correct programs to run unchanged.  For
53 * example, when support for a new PMC type is added.
54 *
55 * The patch version is incremented for every bug fix.
56 */
57#define	PMC_VERSION_MAJOR	0x03
58#define	PMC_VERSION_MINOR	0x01
59#define	PMC_VERSION_PATCH	0x0000
60
61#define	PMC_VERSION		(PMC_VERSION_MAJOR << 24 |		\
62	PMC_VERSION_MINOR << 16 | PMC_VERSION_PATCH)
63
64/*
65 * Kinds of CPUs known.
66 *
67 * We keep track of CPU variants that need to be distinguished in
68 * some way for PMC operations.  CPU names are grouped by manufacturer
69 * and numbered sparsely in order to minimize changes to the ABI involved
70 * when new CPUs are added.
71 */
72
73#define	__PMC_CPUS()						\
74	__PMC_CPU(AMD_K7,	0x00,	"AMD K7")		\
75	__PMC_CPU(AMD_K8,	0x01,	"AMD K8")		\
76	__PMC_CPU(INTEL_P5,	0x80,	"Intel Pentium")	\
77	__PMC_CPU(INTEL_P6,	0x81,	"Intel Pentium Pro")	\
78	__PMC_CPU(INTEL_CL,	0x82,	"Intel Celeron")	\
79	__PMC_CPU(INTEL_PII,	0x83,	"Intel Pentium II")	\
80	__PMC_CPU(INTEL_PIII,	0x84,	"Intel Pentium III")	\
81	__PMC_CPU(INTEL_PM,	0x85,	"Intel Pentium M")	\
82	__PMC_CPU(INTEL_PIV,	0x86,	"Intel Pentium IV")	\
83	__PMC_CPU(INTEL_CORE,	0x87,	"Intel Core Solo/Duo")	\
84	__PMC_CPU(INTEL_CORE2,	0x88,	"Intel Core2")		\
85	__PMC_CPU(INTEL_CORE2EXTREME,	0x89,	"Intel Core2 Extreme")	\
86	__PMC_CPU(INTEL_ATOM,	0x8A,	"Intel Atom")		\
87	__PMC_CPU(INTEL_COREI7, 0x8B,   "Intel Core i7")	\
88	__PMC_CPU(INTEL_WESTMERE, 0x8C,   "Intel Westmere")	\
89	__PMC_CPU(INTEL_SANDYBRIDGE, 0x8D,   "Intel Sandy Bridge")	\
90	__PMC_CPU(INTEL_IVYBRIDGE, 0x8E,   "Intel Ivy Bridge")	\
91	__PMC_CPU(INTEL_SANDYBRIDGE_XEON, 0x8F,   "Intel Sandy Bridge Xeon")	\
92	__PMC_CPU(INTEL_IVYBRIDGE_XEON, 0x90,   "Intel Ivy Bridge Xeon")	\
93	__PMC_CPU(INTEL_HASWELL, 0x91,   "Intel Haswell")	\
94	__PMC_CPU(INTEL_ATOM_SILVERMONT, 0x92,	"Intel Atom Silvermont")    \
95	__PMC_CPU(INTEL_NEHALEM_EX, 0x93,   "Intel Nehalem Xeon 7500")	\
96	__PMC_CPU(INTEL_WESTMERE_EX, 0x94,   "Intel Westmere Xeon E7")	\
97	__PMC_CPU(INTEL_XSCALE,	0x100,	"Intel XScale")		\
98	__PMC_CPU(MIPS_24K,     0x200,  "MIPS 24K")		\
99	__PMC_CPU(MIPS_OCTEON,  0x201,  "Cavium Octeon")	\
100	__PMC_CPU(PPC_7450,     0x300,  "PowerPC MPC7450")	\
101	__PMC_CPU(PPC_970,      0x380,  "IBM PowerPC 970")	\
102	__PMC_CPU(GENERIC, 	0x400,  "Generic")
103
104enum pmc_cputype {
105#undef	__PMC_CPU
106#define	__PMC_CPU(S,V,D)	PMC_CPU_##S = V,
107	__PMC_CPUS()
108};
109
110#define	PMC_CPU_FIRST	PMC_CPU_AMD_K7
111#define	PMC_CPU_LAST	PMC_CPU_GENERIC
112
113/*
114 * Classes of PMCs
115 */
116
117#define	__PMC_CLASSES()							\
118	__PMC_CLASS(TSC)	/* CPU Timestamp counter */		\
119	__PMC_CLASS(K7)		/* AMD K7 performance counters */	\
120	__PMC_CLASS(K8)		/* AMD K8 performance counters */	\
121	__PMC_CLASS(P5)		/* Intel Pentium counters */		\
122	__PMC_CLASS(P6)		/* Intel Pentium Pro counters */	\
123	__PMC_CLASS(P4)		/* Intel Pentium-IV counters */		\
124	__PMC_CLASS(IAF)	/* Intel Core2/Atom, fixed function */	\
125	__PMC_CLASS(IAP)	/* Intel Core...Atom, programmable */	\
126	__PMC_CLASS(UCF)	/* Intel Uncore fixed function */	\
127	__PMC_CLASS(UCP)	/* Intel Uncore programmable */		\
128	__PMC_CLASS(XSCALE)	/* Intel XScale counters */		\
129	__PMC_CLASS(MIPS24K)	/* MIPS 24K */				\
130	__PMC_CLASS(OCTEON)	/* Cavium Octeon */			\
131	__PMC_CLASS(PPC7450)	/* Motorola MPC7450 class */		\
132	__PMC_CLASS(PPC970)	/* IBM PowerPC 970 class */		\
133	__PMC_CLASS(SOFT)	/* Software events */
134
135enum pmc_class {
136#undef  __PMC_CLASS
137#define	__PMC_CLASS(N)	PMC_CLASS_##N ,
138	__PMC_CLASSES()
139};
140
141#define	PMC_CLASS_FIRST	PMC_CLASS_TSC
142#define	PMC_CLASS_LAST	PMC_CLASS_SOFT
143
144/*
145 * A PMC can be in the following states:
146 *
147 * Hardware states:
148 *   DISABLED   -- administratively prohibited from being used.
149 *   FREE       -- HW available for use
150 * Software states:
151 *   ALLOCATED  -- allocated
152 *   STOPPED    -- allocated, but not counting events
153 *   RUNNING    -- allocated, and in operation; 'pm_runcount'
154 *                 holds the number of CPUs using this PMC at
155 *                 a given instant
156 *   DELETED    -- being destroyed
157 */
158
159#define	__PMC_HWSTATES()			\
160	__PMC_STATE(DISABLED)			\
161	__PMC_STATE(FREE)
162
163#define	__PMC_SWSTATES()			\
164	__PMC_STATE(ALLOCATED)			\
165	__PMC_STATE(STOPPED)			\
166	__PMC_STATE(RUNNING)			\
167	__PMC_STATE(DELETED)
168
169#define	__PMC_STATES()				\
170	__PMC_HWSTATES()			\
171	__PMC_SWSTATES()
172
173enum pmc_state {
174#undef	__PMC_STATE
175#define	__PMC_STATE(S)	PMC_STATE_##S,
176	__PMC_STATES()
177	__PMC_STATE(MAX)
178};
179
180#define	PMC_STATE_FIRST	PMC_STATE_DISABLED
181#define	PMC_STATE_LAST	PMC_STATE_DELETED
182
183/*
184 * An allocated PMC may used as a 'global' counter or as a
185 * 'thread-private' one.  Each such mode of use can be in either
186 * statistical sampling mode or in counting mode.  Thus a PMC in use
187 *
188 * SS i.e., SYSTEM STATISTICAL  -- system-wide statistical profiling
189 * SC i.e., SYSTEM COUNTER      -- system-wide counting mode
190 * TS i.e., THREAD STATISTICAL  -- thread virtual, statistical profiling
191 * TC i.e., THREAD COUNTER      -- thread virtual, counting mode
192 *
193 * Statistical profiling modes rely on the PMC periodically delivering
194 * a interrupt to the CPU (when the configured number of events have
195 * been measured), so the PMC must have the ability to generate
196 * interrupts.
197 *
198 * In counting modes, the PMC counts its configured events, with the
199 * value of the PMC being read whenever needed by its owner process.
200 *
201 * The thread specific modes "virtualize" the PMCs -- the PMCs appear
202 * to be thread private and count events only when the profiled thread
203 * actually executes on the CPU.
204 *
205 * The system-wide "global" modes keep the PMCs running all the time
206 * and are used to measure the behaviour of the whole system.
207 */
208
209#define	__PMC_MODES()				\
210	__PMC_MODE(SS,	0)			\
211	__PMC_MODE(SC,	1)			\
212	__PMC_MODE(TS,	2)			\
213	__PMC_MODE(TC,	3)
214
215enum pmc_mode {
216#undef	__PMC_MODE
217#define	__PMC_MODE(M,N)	PMC_MODE_##M = N,
218	__PMC_MODES()
219};
220
221#define	PMC_MODE_FIRST	PMC_MODE_SS
222#define	PMC_MODE_LAST	PMC_MODE_TC
223
224#define	PMC_IS_COUNTING_MODE(mode)				\
225	((mode) == PMC_MODE_SC || (mode) == PMC_MODE_TC)
226#define	PMC_IS_SYSTEM_MODE(mode)				\
227	((mode) == PMC_MODE_SS || (mode) == PMC_MODE_SC)
228#define	PMC_IS_SAMPLING_MODE(mode)				\
229	((mode) == PMC_MODE_SS || (mode) == PMC_MODE_TS)
230#define	PMC_IS_VIRTUAL_MODE(mode)				\
231	((mode) == PMC_MODE_TS || (mode) == PMC_MODE_TC)
232
233/*
234 * PMC row disposition
235 */
236
237#define	__PMC_DISPOSITIONS(N)					\
238	__PMC_DISP(STANDALONE)	/* global/disabled counters */	\
239	__PMC_DISP(FREE)	/* free/available */		\
240	__PMC_DISP(THREAD)	/* thread-virtual PMCs */	\
241	__PMC_DISP(UNKNOWN)	/* sentinel */
242
243enum pmc_disp {
244#undef	__PMC_DISP
245#define	__PMC_DISP(D)	PMC_DISP_##D ,
246	__PMC_DISPOSITIONS()
247};
248
249#define	PMC_DISP_FIRST	PMC_DISP_STANDALONE
250#define	PMC_DISP_LAST	PMC_DISP_THREAD
251
252/*
253 * Counter capabilities
254 *
255 * __PMC_CAPS(NAME, VALUE, DESCRIPTION)
256 */
257
258#define	__PMC_CAPS()							\
259	__PMC_CAP(INTERRUPT,	0, "generate interrupts")		\
260	__PMC_CAP(USER,		1, "count user-mode events")		\
261	__PMC_CAP(SYSTEM,	2, "count system-mode events")		\
262	__PMC_CAP(EDGE,		3, "do edge detection of events")	\
263	__PMC_CAP(THRESHOLD,	4, "ignore events below a threshold")	\
264	__PMC_CAP(READ,		5, "read PMC counter")			\
265	__PMC_CAP(WRITE,	6, "reprogram PMC counter")		\
266	__PMC_CAP(INVERT,	7, "invert comparision sense")		\
267	__PMC_CAP(QUALIFIER,	8, "further qualify monitored events")	\
268	__PMC_CAP(PRECISE,	9, "perform precise sampling")		\
269	__PMC_CAP(TAGGING,	10, "tag upstream events")		\
270	__PMC_CAP(CASCADE,	11, "cascade counters")
271
272enum pmc_caps
273{
274#undef	__PMC_CAP
275#define	__PMC_CAP(NAME, VALUE, DESCR)	PMC_CAP_##NAME = (1 << VALUE) ,
276	__PMC_CAPS()
277};
278
279#define	PMC_CAP_FIRST		PMC_CAP_INTERRUPT
280#define	PMC_CAP_LAST		PMC_CAP_CASCADE
281
282/*
283 * PMC Event Numbers
284 *
285 * These are generated from the definitions in "dev/hwpmc/pmc_events.h".
286 */
287
288enum pmc_event {
289#undef	__PMC_EV
290#undef	__PMC_EV_BLOCK
291#define	__PMC_EV_BLOCK(C,V)	PMC_EV_ ## C ## __BLOCK_START = (V) - 1 ,
292#define	__PMC_EV(C,N)		PMC_EV_ ## C ## _ ## N ,
293	__PMC_EVENTS()
294};
295
296/*
297 * PMC SYSCALL INTERFACE
298 */
299
300/*
301 * "PMC_OPS" -- these are the commands recognized by the kernel
302 * module, and are used when performing a system call from userland.
303 */
304#define	__PMC_OPS()							\
305	__PMC_OP(CONFIGURELOG, "Set log file")				\
306	__PMC_OP(FLUSHLOG, "Flush log file")				\
307	__PMC_OP(GETCPUINFO, "Get system CPU information")		\
308	__PMC_OP(GETDRIVERSTATS, "Get driver statistics")		\
309	__PMC_OP(GETMODULEVERSION, "Get module version")		\
310	__PMC_OP(GETPMCINFO, "Get per-cpu PMC information")		\
311	__PMC_OP(PMCADMIN, "Set PMC state")				\
312	__PMC_OP(PMCALLOCATE, "Allocate and configure a PMC")		\
313	__PMC_OP(PMCATTACH, "Attach a PMC to a process")		\
314	__PMC_OP(PMCDETACH, "Detach a PMC from a process")		\
315	__PMC_OP(PMCGETMSR, "Get a PMC's hardware address")		\
316	__PMC_OP(PMCRELEASE, "Release a PMC")				\
317	__PMC_OP(PMCRW, "Read/Set a PMC")				\
318	__PMC_OP(PMCSETCOUNT, "Set initial count/sampling rate")	\
319	__PMC_OP(PMCSTART, "Start a PMC")				\
320	__PMC_OP(PMCSTOP, "Stop a PMC")					\
321	__PMC_OP(WRITELOG, "Write a cookie to the log file")		\
322	__PMC_OP(CLOSELOG, "Close log file")				\
323	__PMC_OP(GETDYNEVENTINFO, "Get dynamic events list")
324
325
326enum pmc_ops {
327#undef	__PMC_OP
328#define	__PMC_OP(N, D)	PMC_OP_##N,
329	__PMC_OPS()
330};
331
332
333/*
334 * Flags used in operations on PMCs.
335 */
336
337#define	PMC_F_FORCE		0x00000001 /*OP ADMIN force operation */
338#define	PMC_F_DESCENDANTS	0x00000002 /*OP ALLOCATE track descendants */
339#define	PMC_F_LOG_PROCCSW	0x00000004 /*OP ALLOCATE track ctx switches */
340#define	PMC_F_LOG_PROCEXIT	0x00000008 /*OP ALLOCATE log proc exits */
341#define	PMC_F_NEWVALUE		0x00000010 /*OP RW write new value */
342#define	PMC_F_OLDVALUE		0x00000020 /*OP RW get old value */
343#define	PMC_F_KGMON		0x00000040 /*OP ALLOCATE kgmon(8) profiling */
344/* V2 API */
345#define	PMC_F_CALLCHAIN		0x00000080 /*OP ALLOCATE capture callchains */
346
347/* internal flags */
348#define	PMC_F_ATTACHED_TO_OWNER	0x00010000 /*attached to owner*/
349#define	PMC_F_NEEDS_LOGFILE	0x00020000 /*needs log file */
350#define	PMC_F_ATTACH_DONE	0x00040000 /*attached at least once */
351
352#define	PMC_CALLCHAIN_DEPTH_MAX	32
353
354#define	PMC_CC_F_USERSPACE	0x01	   /*userspace callchain*/
355
356/*
357 * Cookies used to denote allocated PMCs, and the values of PMCs.
358 */
359
360typedef uint32_t	pmc_id_t;
361typedef uint64_t	pmc_value_t;
362
363#define	PMC_ID_INVALID		(~ (pmc_id_t) 0)
364
365/*
366 * PMC IDs have the following format:
367 *
368 * +--------+----------+-----------+-----------+
369 * |   CPU  | PMC MODE | PMC CLASS | ROW INDEX |
370 * +--------+----------+-----------+-----------+
371 *
372 * where each field is 8 bits wide.  Field 'CPU' is set to the
373 * requested CPU for system-wide PMCs or PMC_CPU_ANY for process-mode
374 * PMCs.  Field 'PMC MODE' is the allocated PMC mode.  Field 'PMC
375 * CLASS' is the class of the PMC.  Field 'ROW INDEX' is the row index
376 * for the PMC.
377 *
378 * The 'ROW INDEX' ranges over 0..NWPMCS where NHWPMCS is the total
379 * number of hardware PMCs on this cpu.
380 */
381
382
383#define	PMC_ID_TO_ROWINDEX(ID)	((ID) & 0xFF)
384#define	PMC_ID_TO_CLASS(ID)	(((ID) & 0xFF00) >> 8)
385#define	PMC_ID_TO_MODE(ID)	(((ID) & 0xFF0000) >> 16)
386#define	PMC_ID_TO_CPU(ID)	(((ID) & 0xFF000000) >> 24)
387#define	PMC_ID_MAKE_ID(CPU,MODE,CLASS,ROWINDEX)			\
388	((((CPU) & 0xFF) << 24) | (((MODE) & 0xFF) << 16) |	\
389	(((CLASS) & 0xFF) << 8) | ((ROWINDEX) & 0xFF))
390
391/*
392 * Data structures for system calls supported by the pmc driver.
393 */
394
395/*
396 * OP PMCALLOCATE
397 *
398 * Allocate a PMC on the named CPU.
399 */
400
401#define	PMC_CPU_ANY	~0
402
403struct pmc_op_pmcallocate {
404	uint32_t	pm_caps;	/* PMC_CAP_* */
405	uint32_t	pm_cpu;		/* CPU number or PMC_CPU_ANY */
406	enum pmc_class	pm_class;	/* class of PMC desired */
407	enum pmc_event	pm_ev;		/* [enum pmc_event] desired */
408	uint32_t	pm_flags;	/* additional modifiers PMC_F_* */
409	enum pmc_mode	pm_mode;	/* desired mode */
410	pmc_id_t	pm_pmcid;	/* [return] process pmc id */
411
412	union pmc_md_op_pmcallocate pm_md; /* MD layer extensions */
413};
414
415/*
416 * OP PMCADMIN
417 *
418 * Set the administrative state (i.e., whether enabled or disabled) of
419 * a PMC 'pm_pmc' on CPU 'pm_cpu'.  Note that 'pm_pmc' specifies an
420 * absolute PMC number and need not have been first allocated by the
421 * calling process.
422 */
423
424struct pmc_op_pmcadmin {
425	int		pm_cpu;		/* CPU# */
426	uint32_t	pm_flags;	/* flags */
427	int		pm_pmc;         /* PMC# */
428	enum pmc_state  pm_state;	/* desired state */
429};
430
431/*
432 * OP PMCATTACH / OP PMCDETACH
433 *
434 * Attach/detach a PMC and a process.
435 */
436
437struct pmc_op_pmcattach {
438	pmc_id_t	pm_pmc;		/* PMC to attach to */
439	pid_t		pm_pid;		/* target process */
440};
441
442/*
443 * OP PMCSETCOUNT
444 *
445 * Set the sampling rate (i.e., the reload count) for statistical counters.
446 * 'pm_pmcid' need to have been previously allocated using PMCALLOCATE.
447 */
448
449struct pmc_op_pmcsetcount {
450	pmc_value_t	pm_count;	/* initial/sample count */
451	pmc_id_t	pm_pmcid;	/* PMC id to set */
452};
453
454
455/*
456 * OP PMCRW
457 *
458 * Read the value of a PMC named by 'pm_pmcid'.  'pm_pmcid' needs
459 * to have been previously allocated using PMCALLOCATE.
460 */
461
462
463struct pmc_op_pmcrw {
464	uint32_t	pm_flags;	/* PMC_F_{OLD,NEW}VALUE*/
465	pmc_id_t	pm_pmcid;	/* pmc id */
466	pmc_value_t	pm_value;	/* new&returned value */
467};
468
469
470/*
471 * OP GETPMCINFO
472 *
473 * retrieve PMC state for a named CPU.  The caller is expected to
474 * allocate 'npmc' * 'struct pmc_info' bytes of space for the return
475 * values.
476 */
477
478struct pmc_info {
479	char		pm_name[PMC_NAME_MAX]; /* pmc name */
480	enum pmc_class	pm_class;	/* enum pmc_class */
481	int		pm_enabled;	/* whether enabled */
482	enum pmc_disp	pm_rowdisp;	/* FREE, THREAD or STANDLONE */
483	pid_t		pm_ownerpid;	/* owner, or -1 */
484	enum pmc_mode	pm_mode;	/* current mode [enum pmc_mode] */
485	enum pmc_event	pm_event;	/* current event */
486	uint32_t	pm_flags;	/* current flags */
487	pmc_value_t	pm_reloadcount;	/* sampling counters only */
488};
489
490struct pmc_op_getpmcinfo {
491	int32_t		pm_cpu;		/* 0 <= cpu < mp_maxid */
492	struct pmc_info	pm_pmcs[];	/* space for 'npmc' structures */
493};
494
495
496/*
497 * OP GETCPUINFO
498 *
499 * Retrieve system CPU information.
500 */
501
502
503struct pmc_classinfo {
504	enum pmc_class	pm_class;	/* class id */
505	uint32_t	pm_caps;	/* counter capabilities */
506	uint32_t	pm_width;	/* width of the PMC */
507	uint32_t	pm_num;		/* number of PMCs in class */
508};
509
510struct pmc_op_getcpuinfo {
511	enum pmc_cputype pm_cputype; /* what kind of CPU */
512	uint32_t	pm_ncpu;    /* max CPU number */
513	uint32_t	pm_npmc;    /* #PMCs per CPU */
514	uint32_t	pm_nclass;  /* #classes of PMCs */
515	struct pmc_classinfo  pm_classes[PMC_CLASS_MAX];
516};
517
518/*
519 * OP CONFIGURELOG
520 *
521 * Configure a log file for writing system-wide statistics to.
522 */
523
524struct pmc_op_configurelog {
525	int		pm_flags;
526	int		pm_logfd;   /* logfile fd (or -1) */
527};
528
529/*
530 * OP GETDRIVERSTATS
531 *
532 * Retrieve pmc(4) driver-wide statistics.
533 */
534
535struct pmc_op_getdriverstats {
536	int	pm_intr_ignored;	/* #interrupts ignored */
537	int	pm_intr_processed;	/* #interrupts processed */
538	int	pm_intr_bufferfull;	/* #interrupts with ENOSPC */
539	int	pm_syscalls;		/* #syscalls */
540	int	pm_syscall_errors;	/* #syscalls with errors */
541	int	pm_buffer_requests;	/* #buffer requests */
542	int	pm_buffer_requests_failed; /* #failed buffer requests */
543	int	pm_log_sweeps;		/* #sample buffer processing passes */
544};
545
546/*
547 * OP RELEASE / OP START / OP STOP
548 *
549 * Simple operations on a PMC id.
550 */
551
552struct pmc_op_simple {
553	pmc_id_t	pm_pmcid;
554};
555
556/*
557 * OP WRITELOG
558 *
559 * Flush the current log buffer and write 4 bytes of user data to it.
560 */
561
562struct pmc_op_writelog {
563	uint32_t	pm_userdata;
564};
565
566/*
567 * OP GETMSR
568 *
569 * Retrieve the machine specific address assoicated with the allocated
570 * PMC.  This number can be used subsequently with a read-performance-counter
571 * instruction.
572 */
573
574struct pmc_op_getmsr {
575	uint32_t	pm_msr;		/* machine specific address */
576	pmc_id_t	pm_pmcid;	/* allocated pmc id */
577};
578
579/*
580 * OP GETDYNEVENTINFO
581 *
582 * Retrieve a PMC dynamic class events list.
583 */
584
585struct pmc_dyn_event_descr {
586	char		pm_ev_name[PMC_NAME_MAX];
587	enum pmc_event	pm_ev_code;
588};
589
590struct pmc_op_getdyneventinfo {
591	enum pmc_class			pm_class;
592	unsigned int			pm_nevent;
593	struct pmc_dyn_event_descr	pm_events[PMC_EV_DYN_COUNT];
594};
595
596#ifdef _KERNEL
597
598#include <sys/malloc.h>
599#include <sys/sysctl.h>
600
601#include <machine/frame.h>
602
603#define	PMC_HASH_SIZE				1024
604#define	PMC_MTXPOOL_SIZE			2048
605#define	PMC_LOG_BUFFER_SIZE			4
606#define	PMC_NLOGBUFFERS				1024
607#define	PMC_NSAMPLES				1024
608#define	PMC_CALLCHAIN_DEPTH			16
609
610#define PMC_SYSCTL_NAME_PREFIX "kern." PMC_MODULE_NAME "."
611
612/*
613 * Locking keys
614 *
615 * (b) - pmc_bufferlist_mtx (spin lock)
616 * (k) - pmc_kthread_mtx (sleep lock)
617 * (o) - po->po_mtx (spin lock)
618 */
619
620/*
621 * PMC commands
622 */
623
624struct pmc_syscall_args {
625	register_t	pmop_code;	/* one of PMC_OP_* */
626	void		*pmop_data;	/* syscall parameter */
627};
628
629/*
630 * Interface to processor specific s1tuff
631 */
632
633/*
634 * struct pmc_descr
635 *
636 * Machine independent (i.e., the common parts) of a human readable
637 * PMC description.
638 */
639
640struct pmc_descr {
641	char		pd_name[PMC_NAME_MAX]; /* name */
642	uint32_t	pd_caps;	/* capabilities */
643	enum pmc_class	pd_class;	/* class of the PMC */
644	uint32_t	pd_width;	/* width in bits */
645};
646
647/*
648 * struct pmc_target
649 *
650 * This structure records all the target processes associated with a
651 * PMC.
652 */
653
654struct pmc_target {
655	LIST_ENTRY(pmc_target)	pt_next;
656	struct pmc_process	*pt_process; /* target descriptor */
657};
658
659/*
660 * struct pmc
661 *
662 * Describes each allocated PMC.
663 *
664 * Each PMC has precisely one owner, namely the process that allocated
665 * the PMC.
666 *
667 * A PMC may be attached to multiple target processes.  The
668 * 'pm_targets' field links all the target processes being monitored
669 * by this PMC.
670 *
671 * The 'pm_savedvalue' field is protected by a mutex.
672 *
673 * On a multi-cpu machine, multiple target threads associated with a
674 * process-virtual PMC could be concurrently executing on different
675 * CPUs.  The 'pm_runcount' field is atomically incremented every time
676 * the PMC gets scheduled on a CPU and atomically decremented when it
677 * get descheduled.  Deletion of a PMC is only permitted when this
678 * field is '0'.
679 *
680 */
681
682struct pmc {
683	LIST_HEAD(,pmc_target)	pm_targets;	/* list of target processes */
684	LIST_ENTRY(pmc)		pm_next;	/* owner's list */
685
686	/*
687	 * System-wide PMCs are allocated on a CPU and are not moved
688	 * around.  For system-wide PMCs we record the CPU the PMC was
689	 * allocated on in the 'CPU' field of the pmc ID.
690	 *
691	 * Virtual PMCs run on whichever CPU is currently executing
692	 * their targets' threads.  For these PMCs we need to save
693	 * their current PMC counter values when they are taken off
694	 * CPU.
695	 */
696
697	union {
698		pmc_value_t	pm_savedvalue;	/* Virtual PMCS */
699	} pm_gv;
700
701	/*
702	 * For sampling mode PMCs, we keep track of the PMC's "reload
703	 * count", which is the counter value to be loaded in when
704	 * arming the PMC for the next counting session.  For counting
705	 * modes on PMCs that are read-only (e.g., the x86 TSC), we
706	 * keep track of the initial value at the start of
707	 * counting-mode operation.
708	 */
709
710	union {
711		pmc_value_t	pm_reloadcount;	/* sampling PMC modes */
712		pmc_value_t	pm_initial;	/* counting PMC modes */
713	} pm_sc;
714
715	uint32_t	pm_stalled;	/* marks stalled sampling PMCs */
716	uint32_t	pm_caps;	/* PMC capabilities */
717	enum pmc_event	pm_event;	/* event being measured */
718	uint32_t	pm_flags;	/* additional flags PMC_F_... */
719	struct pmc_owner *pm_owner;	/* owner thread state */
720	int		pm_runcount;	/* #cpus currently on */
721	enum pmc_state	pm_state;	/* current PMC state */
722
723	/*
724	 * The PMC ID field encodes the row-index for the PMC, its
725	 * mode, class and the CPU# associated with the PMC.
726	 */
727
728	pmc_id_t	pm_id;		/* allocated PMC id */
729
730	/* md extensions */
731	union pmc_md_pmc	pm_md;
732};
733
734/*
735 * Accessor macros for 'struct pmc'
736 */
737
738#define	PMC_TO_MODE(P)		PMC_ID_TO_MODE((P)->pm_id)
739#define	PMC_TO_CLASS(P)		PMC_ID_TO_CLASS((P)->pm_id)
740#define	PMC_TO_ROWINDEX(P)	PMC_ID_TO_ROWINDEX((P)->pm_id)
741#define	PMC_TO_CPU(P)		PMC_ID_TO_CPU((P)->pm_id)
742
743
744/*
745 * struct pmc_process
746 *
747 * Record a 'target' process being profiled.
748 *
749 * The target process being profiled could be different from the owner
750 * process which allocated the PMCs.  Each target process descriptor
751 * is associated with NHWPMC 'struct pmc *' pointers.  Each PMC at a
752 * given hardware row-index 'n' will use slot 'n' of the 'pp_pmcs[]'
753 * array.  The size of this structure is thus PMC architecture
754 * dependent.
755 *
756 */
757
758struct pmc_targetstate {
759	struct pmc	*pp_pmc;   /* target PMC */
760	pmc_value_t	pp_pmcval; /* per-process value */
761};
762
763struct pmc_process {
764	LIST_ENTRY(pmc_process) pp_next;	/* hash chain */
765	int		pp_refcnt;		/* reference count */
766	uint32_t	pp_flags;		/* flags PMC_PP_* */
767	struct proc	*pp_proc;		/* target thread */
768	struct pmc_targetstate pp_pmcs[];       /* NHWPMCs */
769};
770
771#define	PMC_PP_ENABLE_MSR_ACCESS	0x00000001
772
773/*
774 * struct pmc_owner
775 *
776 * We associate a PMC with an 'owner' process.
777 *
778 * A process can be associated with 0..NCPUS*NHWPMC PMCs during its
779 * lifetime, where NCPUS is the numbers of CPUS in the system and
780 * NHWPMC is the number of hardware PMCs per CPU.  These are
781 * maintained in the list headed by the 'po_pmcs' to save on space.
782 *
783 */
784
785struct pmc_owner  {
786	LIST_ENTRY(pmc_owner)	po_next;	/* hash chain */
787	LIST_ENTRY(pmc_owner)	po_ssnext;	/* list of SS PMC owners */
788	LIST_HEAD(, pmc)	po_pmcs;	/* owned PMC list */
789	TAILQ_HEAD(, pmclog_buffer) po_logbuffers; /* (o) logbuffer list */
790	struct mtx		po_mtx;		/* spin lock for (o) */
791	struct proc		*po_owner;	/* owner proc */
792	uint32_t		po_flags;	/* (k) flags PMC_PO_* */
793	struct proc		*po_kthread;	/* (k) helper kthread */
794	struct pmclog_buffer	*po_curbuf;	/* current log buffer */
795	struct file		*po_file;	/* file reference */
796	int			po_error;	/* recorded error */
797	short			po_sscount;	/* # SS PMCs owned */
798	short			po_logprocmaps;	/* global mappings done */
799};
800
801#define	PMC_PO_OWNS_LOGFILE		0x00000001 /* has a log file */
802#define	PMC_PO_SHUTDOWN			0x00000010 /* in the process of shutdown */
803#define	PMC_PO_INITIAL_MAPPINGS_DONE	0x00000020
804
805/*
806 * struct pmc_hw -- describe the state of the PMC hardware
807 *
808 * When in use, a HW PMC is associated with one allocated 'struct pmc'
809 * pointed to by field 'phw_pmc'.  When inactive, this field is NULL.
810 *
811 * On an SMP box, one or more HW PMC's in process virtual mode with
812 * the same 'phw_pmc' could be executing on different CPUs.  In order
813 * to handle this case correctly, we need to ensure that only
814 * incremental counts get added to the saved value in the associated
815 * 'struct pmc'.  The 'phw_save' field is used to keep the saved PMC
816 * value at the time the hardware is started during this context
817 * switch (i.e., the difference between the new (hardware) count and
818 * the saved count is atomically added to the count field in 'struct
819 * pmc' at context switch time).
820 *
821 */
822
823struct pmc_hw {
824	uint32_t	phw_state;	/* see PHW_* macros below */
825	struct pmc	*phw_pmc;	/* current thread PMC */
826};
827
828#define	PMC_PHW_RI_MASK		0x000000FF
829#define	PMC_PHW_CPU_SHIFT	8
830#define	PMC_PHW_CPU_MASK	0x0000FF00
831#define	PMC_PHW_FLAGS_SHIFT	16
832#define	PMC_PHW_FLAGS_MASK	0xFFFF0000
833
834#define	PMC_PHW_INDEX_TO_STATE(ri)	((ri) & PMC_PHW_RI_MASK)
835#define	PMC_PHW_STATE_TO_INDEX(state)	((state) & PMC_PHW_RI_MASK)
836#define	PMC_PHW_CPU_TO_STATE(cpu)	(((cpu) << PMC_PHW_CPU_SHIFT) & \
837	PMC_PHW_CPU_MASK)
838#define	PMC_PHW_STATE_TO_CPU(state)	(((state) & PMC_PHW_CPU_MASK) >> \
839	PMC_PHW_CPU_SHIFT)
840#define	PMC_PHW_FLAGS_TO_STATE(flags)	(((flags) << PMC_PHW_FLAGS_SHIFT) & \
841	PMC_PHW_FLAGS_MASK)
842#define	PMC_PHW_STATE_TO_FLAGS(state)	(((state) & PMC_PHW_FLAGS_MASK) >> \
843	PMC_PHW_FLAGS_SHIFT)
844#define	PMC_PHW_FLAG_IS_ENABLED		(PMC_PHW_FLAGS_TO_STATE(0x01))
845#define	PMC_PHW_FLAG_IS_SHAREABLE	(PMC_PHW_FLAGS_TO_STATE(0x02))
846
847/*
848 * struct pmc_sample
849 *
850 * Space for N (tunable) PC samples and associated control data.
851 */
852
853struct pmc_sample {
854	uint16_t		ps_nsamples;	/* callchain depth */
855	uint8_t			ps_cpu;		/* cpu number */
856	uint8_t			ps_flags;	/* other flags */
857	pid_t			ps_pid;		/* process PID or -1 */
858	struct thread		*ps_td;		/* which thread */
859	struct pmc		*ps_pmc;	/* interrupting PMC */
860	uintptr_t		*ps_pc;		/* (const) callchain start */
861};
862
863#define 	PMC_SAMPLE_FREE		((uint16_t) 0)
864#define 	PMC_SAMPLE_INUSE	((uint16_t) 0xFFFF)
865
866struct pmc_samplebuffer {
867	struct pmc_sample * volatile ps_read;	/* read pointer */
868	struct pmc_sample * volatile ps_write;	/* write pointer */
869	uintptr_t		*ps_callchains;	/* all saved call chains */
870	struct pmc_sample	*ps_fence;	/* one beyond ps_samples[] */
871	struct pmc_sample	ps_samples[];	/* array of sample entries */
872};
873
874
875/*
876 * struct pmc_cpustate
877 *
878 * A CPU is modelled as a collection of HW PMCs with space for additional
879 * flags.
880 */
881
882struct pmc_cpu {
883	uint32_t	pc_state;	/* physical cpu number + flags */
884	struct pmc_samplebuffer *pc_sb[2]; /* space for samples */
885	struct pmc_hw	*pc_hwpmcs[];	/* 'npmc' pointers */
886};
887
888#define	PMC_PCPU_CPU_MASK		0x000000FF
889#define	PMC_PCPU_FLAGS_MASK		0xFFFFFF00
890#define	PMC_PCPU_FLAGS_SHIFT		8
891#define	PMC_PCPU_STATE_TO_CPU(S)	((S) & PMC_PCPU_CPU_MASK)
892#define	PMC_PCPU_STATE_TO_FLAGS(S)	(((S) & PMC_PCPU_FLAGS_MASK) >> PMC_PCPU_FLAGS_SHIFT)
893#define	PMC_PCPU_FLAGS_TO_STATE(F)	(((F) << PMC_PCPU_FLAGS_SHIFT) & PMC_PCPU_FLAGS_MASK)
894#define	PMC_PCPU_CPU_TO_STATE(C)	((C) & PMC_PCPU_CPU_MASK)
895#define	PMC_PCPU_FLAG_HTT		(PMC_PCPU_FLAGS_TO_STATE(0x1))
896
897/*
898 * struct pmc_binding
899 *
900 * CPU binding information.
901 */
902
903struct pmc_binding {
904	int	pb_bound;	/* is bound? */
905	int	pb_cpu;		/* if so, to which CPU */
906};
907
908
909struct pmc_mdep;
910
911/*
912 * struct pmc_classdep
913 *
914 * PMC class-dependent operations.
915 */
916struct pmc_classdep {
917	uint32_t	pcd_caps;	/* class capabilities */
918	enum pmc_class	pcd_class;	/* class id */
919	int		pcd_num;	/* number of PMCs */
920	int		pcd_ri;		/* row index of the first PMC in class */
921	int		pcd_width;	/* width of the PMC */
922
923	/* configuring/reading/writing the hardware PMCs */
924	int (*pcd_config_pmc)(int _cpu, int _ri, struct pmc *_pm);
925	int (*pcd_get_config)(int _cpu, int _ri, struct pmc **_ppm);
926	int (*pcd_read_pmc)(int _cpu, int _ri, pmc_value_t *_value);
927	int (*pcd_write_pmc)(int _cpu, int _ri, pmc_value_t _value);
928
929	/* pmc allocation/release */
930	int (*pcd_allocate_pmc)(int _cpu, int _ri, struct pmc *_t,
931		const struct pmc_op_pmcallocate *_a);
932	int (*pcd_release_pmc)(int _cpu, int _ri, struct pmc *_pm);
933
934	/* starting and stopping PMCs */
935	int (*pcd_start_pmc)(int _cpu, int _ri);
936	int (*pcd_stop_pmc)(int _cpu, int _ri);
937
938	/* description */
939	int (*pcd_describe)(int _cpu, int _ri, struct pmc_info *_pi,
940		struct pmc **_ppmc);
941
942	/* class-dependent initialization & finalization */
943	int (*pcd_pcpu_init)(struct pmc_mdep *_md, int _cpu);
944	int (*pcd_pcpu_fini)(struct pmc_mdep *_md, int _cpu);
945
946	/* machine-specific interface */
947	int (*pcd_get_msr)(int _ri, uint32_t *_msr);
948};
949
950/*
951 * struct pmc_mdep
952 *
953 * Machine dependent bits needed per CPU type.
954 */
955
956struct pmc_mdep  {
957	uint32_t	pmd_cputype;    /* from enum pmc_cputype */
958	uint32_t	pmd_npmc;	/* number of PMCs per CPU */
959	uint32_t	pmd_nclass;	/* number of PMC classes present */
960
961	/*
962	 * Machine dependent methods.
963	 */
964
965	/* per-cpu initialization and finalization */
966	int (*pmd_pcpu_init)(struct pmc_mdep *_md, int _cpu);
967	int (*pmd_pcpu_fini)(struct pmc_mdep *_md, int _cpu);
968
969	/* thread context switch in/out */
970	int (*pmd_switch_in)(struct pmc_cpu *_p, struct pmc_process *_pp);
971	int (*pmd_switch_out)(struct pmc_cpu *_p, struct pmc_process *_pp);
972
973	/* handle a PMC interrupt */
974	int (*pmd_intr)(int _cpu, struct trapframe *_tf);
975
976	/*
977	 * PMC class dependent information.
978	 */
979	struct pmc_classdep pmd_classdep[];
980};
981
982/*
983 * Per-CPU state.  This is an array of 'mp_ncpu' pointers
984 * to struct pmc_cpu descriptors.
985 */
986
987extern struct pmc_cpu **pmc_pcpu;
988
989/* driver statistics */
990extern struct pmc_op_getdriverstats pmc_stats;
991
992#if	defined(DEBUG)
993
994/* debug flags, major flag groups */
995struct pmc_debugflags {
996	int	pdb_CPU;
997	int	pdb_CSW;
998	int	pdb_LOG;
999	int	pdb_MDP;
1000	int	pdb_MOD;
1001	int	pdb_OWN;
1002	int	pdb_PMC;
1003	int	pdb_PRC;
1004	int	pdb_SAM;
1005};
1006
1007extern struct pmc_debugflags pmc_debugflags;
1008
1009#define	PMC_DEBUG_STRSIZE		128
1010#define	PMC_DEBUG_DEFAULT_FLAGS		{ 0, 0, 0, 0, 0, 0, 0, 0 }
1011
1012#define	PMCDBG(M,N,L,F,...) do {					\
1013	if (pmc_debugflags.pdb_ ## M & (1 << PMC_DEBUG_MIN_ ## N))	\
1014		printf(#M ":" #N ":" #L  ": " F "\n", __VA_ARGS__);	\
1015} while (0)
1016
1017/* Major numbers */
1018#define	PMC_DEBUG_MAJ_CPU		0 /* cpu switches */
1019#define	PMC_DEBUG_MAJ_CSW		1 /* context switches */
1020#define	PMC_DEBUG_MAJ_LOG		2 /* logging */
1021#define	PMC_DEBUG_MAJ_MDP		3 /* machine dependent */
1022#define	PMC_DEBUG_MAJ_MOD		4 /* misc module infrastructure */
1023#define	PMC_DEBUG_MAJ_OWN		5 /* owner */
1024#define	PMC_DEBUG_MAJ_PMC		6 /* pmc management */
1025#define	PMC_DEBUG_MAJ_PRC		7 /* processes */
1026#define	PMC_DEBUG_MAJ_SAM		8 /* sampling */
1027
1028/* Minor numbers */
1029
1030/* Common (8 bits) */
1031#define	PMC_DEBUG_MIN_ALL		0 /* allocation */
1032#define	PMC_DEBUG_MIN_REL		1 /* release */
1033#define	PMC_DEBUG_MIN_OPS		2 /* ops: start, stop, ... */
1034#define	PMC_DEBUG_MIN_INI		3 /* init */
1035#define	PMC_DEBUG_MIN_FND		4 /* find */
1036
1037/* MODULE */
1038#define	PMC_DEBUG_MIN_PMH	       14 /* pmc_hook */
1039#define	PMC_DEBUG_MIN_PMS	       15 /* pmc_syscall */
1040
1041/* OWN */
1042#define	PMC_DEBUG_MIN_ORM		8 /* owner remove */
1043#define	PMC_DEBUG_MIN_OMR		9 /* owner maybe remove */
1044
1045/* PROCESSES */
1046#define	PMC_DEBUG_MIN_TLK		8 /* link target */
1047#define	PMC_DEBUG_MIN_TUL		9 /* unlink target */
1048#define	PMC_DEBUG_MIN_EXT	       10 /* process exit */
1049#define	PMC_DEBUG_MIN_EXC	       11 /* process exec */
1050#define	PMC_DEBUG_MIN_FRK	       12 /* process fork */
1051#define	PMC_DEBUG_MIN_ATT	       13 /* attach/detach */
1052#define	PMC_DEBUG_MIN_SIG	       14 /* signalling */
1053
1054/* CONTEXT SWITCHES */
1055#define	PMC_DEBUG_MIN_SWI		8 /* switch in */
1056#define	PMC_DEBUG_MIN_SWO		9 /* switch out */
1057
1058/* PMC */
1059#define	PMC_DEBUG_MIN_REG		8 /* pmc register */
1060#define	PMC_DEBUG_MIN_ALR		9 /* allocate row */
1061
1062/* MACHINE DEPENDENT LAYER */
1063#define	PMC_DEBUG_MIN_REA		8 /* read */
1064#define	PMC_DEBUG_MIN_WRI		9 /* write */
1065#define	PMC_DEBUG_MIN_CFG	       10 /* config */
1066#define	PMC_DEBUG_MIN_STA	       11 /* start */
1067#define	PMC_DEBUG_MIN_STO	       12 /* stop */
1068#define	PMC_DEBUG_MIN_INT	       13 /* interrupts */
1069
1070/* CPU */
1071#define	PMC_DEBUG_MIN_BND		8 /* bind */
1072#define	PMC_DEBUG_MIN_SEL		9 /* select */
1073
1074/* LOG */
1075#define	PMC_DEBUG_MIN_GTB		8 /* get buf */
1076#define	PMC_DEBUG_MIN_SIO		9 /* schedule i/o */
1077#define	PMC_DEBUG_MIN_FLS	       10 /* flush */
1078#define	PMC_DEBUG_MIN_SAM	       11 /* sample */
1079#define	PMC_DEBUG_MIN_CLO	       12 /* close */
1080
1081#else
1082#define	PMCDBG(M,N,L,F,...)		/* nothing */
1083#endif
1084
1085/* declare a dedicated memory pool */
1086MALLOC_DECLARE(M_PMC);
1087
1088/*
1089 * Functions
1090 */
1091
1092struct pmc_mdep *pmc_md_initialize(void);	/* MD init function */
1093void	pmc_md_finalize(struct pmc_mdep *_md);	/* MD fini function */
1094int	pmc_getrowdisp(int _ri);
1095int	pmc_process_interrupt(int _cpu, int _soft, struct pmc *_pm,
1096    struct trapframe *_tf, int _inuserspace);
1097int	pmc_save_kernel_callchain(uintptr_t *_cc, int _maxsamples,
1098    struct trapframe *_tf);
1099int	pmc_save_user_callchain(uintptr_t *_cc, int _maxsamples,
1100    struct trapframe *_tf);
1101struct pmc_mdep *pmc_mdep_alloc(int nclasses);
1102void pmc_mdep_free(struct pmc_mdep *md);
1103#endif /* _KERNEL */
1104#endif /* _SYS_PMC_H_ */
1105