2 * Copyright (c) 2003-2008, Joseph Koshy
3 * Copyright (c) 2007 The FreeBSD Foundation
6 * Portions of this software were developed by A. Joseph Koshy under
7 * sponsorship from the FreeBSD Foundation and Google, Inc.
9 * Redistribution and use in source and binary forms, with or without
10 * modification, are permitted provided that the following conditions
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.
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
36 #include <dev/hwpmc/pmc_events.h>
38 #include <machine/pmc_mdep.h>
39 #include <machine/profile.h>
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 */
46 * Kernel<->userland API version number [MMmmpppp]
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.
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.
55 * The patch version is incremented for every bug fix.
57 #define PMC_VERSION_MAJOR 0x03
58 #define PMC_VERSION_MINOR 0x01
59 #define PMC_VERSION_PATCH 0x0000
61 #define PMC_VERSION (PMC_VERSION_MAJOR << 24 | \
62 PMC_VERSION_MINOR << 16 | PMC_VERSION_PATCH)
65 * Kinds of CPUs known.
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.
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_HASWELL_XEON, 0x95, "Intel Haswell Xeon E5 v3") \
98 __PMC_CPU(INTEL_BROADWELL, 0x96, "Intel Broadwell") \
99 __PMC_CPU(INTEL_XSCALE, 0x100, "Intel XScale") \
100 __PMC_CPU(MIPS_24K, 0x200, "MIPS 24K") \
101 __PMC_CPU(MIPS_OCTEON, 0x201, "Cavium Octeon") \
102 __PMC_CPU(MIPS_74K, 0x202, "MIPS 74K") \
103 __PMC_CPU(PPC_7450, 0x300, "PowerPC MPC7450") \
104 __PMC_CPU(PPC_E500, 0x340, "PowerPC e500 Core") \
105 __PMC_CPU(PPC_MPC85XX, 0x340, "Freescale PowerPC MPC85XX") \
106 __PMC_CPU(PPC_970, 0x380, "IBM PowerPC 970") \
107 __PMC_CPU(GENERIC, 0x400, "Generic") \
108 __PMC_CPU(ARMV7, 0x500, "ARMv7") \
109 __PMC_CPU(ARMV8_CORTEX_A53, 0x600, "ARMv8 Cortex A53") \
110 __PMC_CPU(ARMV8_CORTEX_A57, 0x601, "ARMv8 Cortex A57")
114 #define __PMC_CPU(S,V,D) PMC_CPU_##S = V,
118 #define PMC_CPU_FIRST PMC_CPU_AMD_K7
119 #define PMC_CPU_LAST PMC_CPU_GENERIC
125 #define __PMC_CLASSES() \
126 __PMC_CLASS(TSC, 0x000, "CPU Timestamp counter") \
127 __PMC_CLASS(K7, 0x100, "AMD K7 performance counters") \
128 __PMC_CLASS(K8, 0x101, "AMD K8 performance counters") \
129 __PMC_CLASS(P5, 0x102, "Intel Pentium counters") \
130 __PMC_CLASS(P6, 0x103, "Intel Pentium Pro counters") \
131 __PMC_CLASS(P4, 0x104, "Intel Pentium-IV counters") \
132 __PMC_CLASS(IAF, 0x105, "Intel Core2/Atom, fixed function") \
133 __PMC_CLASS(IAP, 0x106, "Intel Core...Atom, programmable") \
134 __PMC_CLASS(UCF, 0x107, "Intel Uncore fixed function") \
135 __PMC_CLASS(UCP, 0x108, "Intel Uncore programmable") \
136 __PMC_CLASS(XSCALE, 0x200, "Intel XScale counters") \
137 __PMC_CLASS(ARMV7, 0x201, "ARMv7") \
138 __PMC_CLASS(ARMV8, 0x202, "ARMv8") \
139 __PMC_CLASS(MIPS24K, 0x300, "MIPS 24K") \
140 __PMC_CLASS(OCTEON, 0x301, "Cavium Octeon") \
141 __PMC_CLASS(MIPS74K, 0x302, "MIPS 74K") \
142 __PMC_CLASS(PPC7450, 0x400, "Motorola MPC7450 class") \
143 __PMC_CLASS(PPC970, 0x401, "IBM PowerPC 970 class") \
144 __PMC_CLASS(E500, 0x402, "Freescale e500 class") \
145 __PMC_CLASS(SOFT, 0x8000, "Software events")
149 #define __PMC_CLASS(S,V,D) PMC_CLASS_##S = V,
153 #define PMC_CLASS_FIRST PMC_CLASS_TSC
154 #define PMC_CLASS_LAST PMC_CLASS_SOFT
157 * A PMC can be in the following states:
160 * DISABLED -- administratively prohibited from being used.
161 * FREE -- HW available for use
163 * ALLOCATED -- allocated
164 * STOPPED -- allocated, but not counting events
165 * RUNNING -- allocated, and in operation; 'pm_runcount'
166 * holds the number of CPUs using this PMC at
168 * DELETED -- being destroyed
171 #define __PMC_HWSTATES() \
172 __PMC_STATE(DISABLED) \
175 #define __PMC_SWSTATES() \
176 __PMC_STATE(ALLOCATED) \
177 __PMC_STATE(STOPPED) \
178 __PMC_STATE(RUNNING) \
181 #define __PMC_STATES() \
187 #define __PMC_STATE(S) PMC_STATE_##S,
192 #define PMC_STATE_FIRST PMC_STATE_DISABLED
193 #define PMC_STATE_LAST PMC_STATE_DELETED
196 * An allocated PMC may used as a 'global' counter or as a
197 * 'thread-private' one. Each such mode of use can be in either
198 * statistical sampling mode or in counting mode. Thus a PMC in use
200 * SS i.e., SYSTEM STATISTICAL -- system-wide statistical profiling
201 * SC i.e., SYSTEM COUNTER -- system-wide counting mode
202 * TS i.e., THREAD STATISTICAL -- thread virtual, statistical profiling
203 * TC i.e., THREAD COUNTER -- thread virtual, counting mode
205 * Statistical profiling modes rely on the PMC periodically delivering
206 * a interrupt to the CPU (when the configured number of events have
207 * been measured), so the PMC must have the ability to generate
210 * In counting modes, the PMC counts its configured events, with the
211 * value of the PMC being read whenever needed by its owner process.
213 * The thread specific modes "virtualize" the PMCs -- the PMCs appear
214 * to be thread private and count events only when the profiled thread
215 * actually executes on the CPU.
217 * The system-wide "global" modes keep the PMCs running all the time
218 * and are used to measure the behaviour of the whole system.
221 #define __PMC_MODES() \
229 #define __PMC_MODE(M,N) PMC_MODE_##M = N,
233 #define PMC_MODE_FIRST PMC_MODE_SS
234 #define PMC_MODE_LAST PMC_MODE_TC
236 #define PMC_IS_COUNTING_MODE(mode) \
237 ((mode) == PMC_MODE_SC || (mode) == PMC_MODE_TC)
238 #define PMC_IS_SYSTEM_MODE(mode) \
239 ((mode) == PMC_MODE_SS || (mode) == PMC_MODE_SC)
240 #define PMC_IS_SAMPLING_MODE(mode) \
241 ((mode) == PMC_MODE_SS || (mode) == PMC_MODE_TS)
242 #define PMC_IS_VIRTUAL_MODE(mode) \
243 ((mode) == PMC_MODE_TS || (mode) == PMC_MODE_TC)
246 * PMC row disposition
249 #define __PMC_DISPOSITIONS(N) \
250 __PMC_DISP(STANDALONE) /* global/disabled counters */ \
251 __PMC_DISP(FREE) /* free/available */ \
252 __PMC_DISP(THREAD) /* thread-virtual PMCs */ \
253 __PMC_DISP(UNKNOWN) /* sentinel */
257 #define __PMC_DISP(D) PMC_DISP_##D ,
261 #define PMC_DISP_FIRST PMC_DISP_STANDALONE
262 #define PMC_DISP_LAST PMC_DISP_THREAD
265 * Counter capabilities
267 * __PMC_CAPS(NAME, VALUE, DESCRIPTION)
270 #define __PMC_CAPS() \
271 __PMC_CAP(INTERRUPT, 0, "generate interrupts") \
272 __PMC_CAP(USER, 1, "count user-mode events") \
273 __PMC_CAP(SYSTEM, 2, "count system-mode events") \
274 __PMC_CAP(EDGE, 3, "do edge detection of events") \
275 __PMC_CAP(THRESHOLD, 4, "ignore events below a threshold") \
276 __PMC_CAP(READ, 5, "read PMC counter") \
277 __PMC_CAP(WRITE, 6, "reprogram PMC counter") \
278 __PMC_CAP(INVERT, 7, "invert comparision sense") \
279 __PMC_CAP(QUALIFIER, 8, "further qualify monitored events") \
280 __PMC_CAP(PRECISE, 9, "perform precise sampling") \
281 __PMC_CAP(TAGGING, 10, "tag upstream events") \
282 __PMC_CAP(CASCADE, 11, "cascade counters")
287 #define __PMC_CAP(NAME, VALUE, DESCR) PMC_CAP_##NAME = (1 << VALUE) ,
291 #define PMC_CAP_FIRST PMC_CAP_INTERRUPT
292 #define PMC_CAP_LAST PMC_CAP_CASCADE
297 * These are generated from the definitions in "dev/hwpmc/pmc_events.h".
302 #undef __PMC_EV_BLOCK
303 #define __PMC_EV_BLOCK(C,V) PMC_EV_ ## C ## __BLOCK_START = (V) - 1 ,
304 #define __PMC_EV(C,N) PMC_EV_ ## C ## _ ## N ,
309 * PMC SYSCALL INTERFACE
313 * "PMC_OPS" -- these are the commands recognized by the kernel
314 * module, and are used when performing a system call from userland.
316 #define __PMC_OPS() \
317 __PMC_OP(CONFIGURELOG, "Set log file") \
318 __PMC_OP(FLUSHLOG, "Flush log file") \
319 __PMC_OP(GETCPUINFO, "Get system CPU information") \
320 __PMC_OP(GETDRIVERSTATS, "Get driver statistics") \
321 __PMC_OP(GETMODULEVERSION, "Get module version") \
322 __PMC_OP(GETPMCINFO, "Get per-cpu PMC information") \
323 __PMC_OP(PMCADMIN, "Set PMC state") \
324 __PMC_OP(PMCALLOCATE, "Allocate and configure a PMC") \
325 __PMC_OP(PMCATTACH, "Attach a PMC to a process") \
326 __PMC_OP(PMCDETACH, "Detach a PMC from a process") \
327 __PMC_OP(PMCGETMSR, "Get a PMC's hardware address") \
328 __PMC_OP(PMCRELEASE, "Release a PMC") \
329 __PMC_OP(PMCRW, "Read/Set a PMC") \
330 __PMC_OP(PMCSETCOUNT, "Set initial count/sampling rate") \
331 __PMC_OP(PMCSTART, "Start a PMC") \
332 __PMC_OP(PMCSTOP, "Stop a PMC") \
333 __PMC_OP(WRITELOG, "Write a cookie to the log file") \
334 __PMC_OP(CLOSELOG, "Close log file") \
335 __PMC_OP(GETDYNEVENTINFO, "Get dynamic events list")
340 #define __PMC_OP(N, D) PMC_OP_##N,
346 * Flags used in operations on PMCs.
349 #define PMC_F_FORCE 0x00000001 /*OP ADMIN force operation */
350 #define PMC_F_DESCENDANTS 0x00000002 /*OP ALLOCATE track descendants */
351 #define PMC_F_LOG_PROCCSW 0x00000004 /*OP ALLOCATE track ctx switches */
352 #define PMC_F_LOG_PROCEXIT 0x00000008 /*OP ALLOCATE log proc exits */
353 #define PMC_F_NEWVALUE 0x00000010 /*OP RW write new value */
354 #define PMC_F_OLDVALUE 0x00000020 /*OP RW get old value */
355 #define PMC_F_KGMON 0x00000040 /*OP ALLOCATE kgmon(8) profiling */
357 #define PMC_F_CALLCHAIN 0x00000080 /*OP ALLOCATE capture callchains */
360 #define PMC_F_ATTACHED_TO_OWNER 0x00010000 /*attached to owner*/
361 #define PMC_F_NEEDS_LOGFILE 0x00020000 /*needs log file */
362 #define PMC_F_ATTACH_DONE 0x00040000 /*attached at least once */
364 #define PMC_CALLCHAIN_DEPTH_MAX 128
366 #define PMC_CC_F_USERSPACE 0x01 /*userspace callchain*/
369 * Cookies used to denote allocated PMCs, and the values of PMCs.
372 typedef uint32_t pmc_id_t;
373 typedef uint64_t pmc_value_t;
375 #define PMC_ID_INVALID (~ (pmc_id_t) 0)
378 * PMC IDs have the following format:
380 * +--------+----------+-----------+-----------+
381 * | CPU | PMC MODE | PMC CLASS | ROW INDEX |
382 * +--------+----------+-----------+-----------+
384 * where each field is 8 bits wide. Field 'CPU' is set to the
385 * requested CPU for system-wide PMCs or PMC_CPU_ANY for process-mode
386 * PMCs. Field 'PMC MODE' is the allocated PMC mode. Field 'PMC
387 * CLASS' is the class of the PMC. Field 'ROW INDEX' is the row index
390 * The 'ROW INDEX' ranges over 0..NWPMCS where NHWPMCS is the total
391 * number of hardware PMCs on this cpu.
395 #define PMC_ID_TO_ROWINDEX(ID) ((ID) & 0xFF)
396 #define PMC_ID_TO_CLASS(ID) (((ID) & 0xFF00) >> 8)
397 #define PMC_ID_TO_MODE(ID) (((ID) & 0xFF0000) >> 16)
398 #define PMC_ID_TO_CPU(ID) (((ID) & 0xFF000000) >> 24)
399 #define PMC_ID_MAKE_ID(CPU,MODE,CLASS,ROWINDEX) \
400 ((((CPU) & 0xFF) << 24) | (((MODE) & 0xFF) << 16) | \
401 (((CLASS) & 0xFF) << 8) | ((ROWINDEX) & 0xFF))
404 * Data structures for system calls supported by the pmc driver.
410 * Allocate a PMC on the named CPU.
413 #define PMC_CPU_ANY ~0
415 struct pmc_op_pmcallocate {
416 uint32_t pm_caps; /* PMC_CAP_* */
417 uint32_t pm_cpu; /* CPU number or PMC_CPU_ANY */
418 enum pmc_class pm_class; /* class of PMC desired */
419 enum pmc_event pm_ev; /* [enum pmc_event] desired */
420 uint32_t pm_flags; /* additional modifiers PMC_F_* */
421 enum pmc_mode pm_mode; /* desired mode */
422 pmc_id_t pm_pmcid; /* [return] process pmc id */
424 union pmc_md_op_pmcallocate pm_md; /* MD layer extensions */
430 * Set the administrative state (i.e., whether enabled or disabled) of
431 * a PMC 'pm_pmc' on CPU 'pm_cpu'. Note that 'pm_pmc' specifies an
432 * absolute PMC number and need not have been first allocated by the
436 struct pmc_op_pmcadmin {
437 int pm_cpu; /* CPU# */
438 uint32_t pm_flags; /* flags */
439 int pm_pmc; /* PMC# */
440 enum pmc_state pm_state; /* desired state */
444 * OP PMCATTACH / OP PMCDETACH
446 * Attach/detach a PMC and a process.
449 struct pmc_op_pmcattach {
450 pmc_id_t pm_pmc; /* PMC to attach to */
451 pid_t pm_pid; /* target process */
457 * Set the sampling rate (i.e., the reload count) for statistical counters.
458 * 'pm_pmcid' need to have been previously allocated using PMCALLOCATE.
461 struct pmc_op_pmcsetcount {
462 pmc_value_t pm_count; /* initial/sample count */
463 pmc_id_t pm_pmcid; /* PMC id to set */
470 * Read the value of a PMC named by 'pm_pmcid'. 'pm_pmcid' needs
471 * to have been previously allocated using PMCALLOCATE.
475 struct pmc_op_pmcrw {
476 uint32_t pm_flags; /* PMC_F_{OLD,NEW}VALUE*/
477 pmc_id_t pm_pmcid; /* pmc id */
478 pmc_value_t pm_value; /* new&returned value */
485 * retrieve PMC state for a named CPU. The caller is expected to
486 * allocate 'npmc' * 'struct pmc_info' bytes of space for the return
491 char pm_name[PMC_NAME_MAX]; /* pmc name */
492 enum pmc_class pm_class; /* enum pmc_class */
493 int pm_enabled; /* whether enabled */
494 enum pmc_disp pm_rowdisp; /* FREE, THREAD or STANDLONE */
495 pid_t pm_ownerpid; /* owner, or -1 */
496 enum pmc_mode pm_mode; /* current mode [enum pmc_mode] */
497 enum pmc_event pm_event; /* current event */
498 uint32_t pm_flags; /* current flags */
499 pmc_value_t pm_reloadcount; /* sampling counters only */
502 struct pmc_op_getpmcinfo {
503 int32_t pm_cpu; /* 0 <= cpu < mp_maxid */
504 struct pmc_info pm_pmcs[]; /* space for 'npmc' structures */
511 * Retrieve system CPU information.
515 struct pmc_classinfo {
516 enum pmc_class pm_class; /* class id */
517 uint32_t pm_caps; /* counter capabilities */
518 uint32_t pm_width; /* width of the PMC */
519 uint32_t pm_num; /* number of PMCs in class */
522 struct pmc_op_getcpuinfo {
523 enum pmc_cputype pm_cputype; /* what kind of CPU */
524 uint32_t pm_ncpu; /* max CPU number */
525 uint32_t pm_npmc; /* #PMCs per CPU */
526 uint32_t pm_nclass; /* #classes of PMCs */
527 struct pmc_classinfo pm_classes[PMC_CLASS_MAX];
533 * Configure a log file for writing system-wide statistics to.
536 struct pmc_op_configurelog {
538 int pm_logfd; /* logfile fd (or -1) */
544 * Retrieve pmc(4) driver-wide statistics.
547 struct pmc_op_getdriverstats {
548 int pm_intr_ignored; /* #interrupts ignored */
549 int pm_intr_processed; /* #interrupts processed */
550 int pm_intr_bufferfull; /* #interrupts with ENOSPC */
551 int pm_syscalls; /* #syscalls */
552 int pm_syscall_errors; /* #syscalls with errors */
553 int pm_buffer_requests; /* #buffer requests */
554 int pm_buffer_requests_failed; /* #failed buffer requests */
555 int pm_log_sweeps; /* #sample buffer processing passes */
559 * OP RELEASE / OP START / OP STOP
561 * Simple operations on a PMC id.
564 struct pmc_op_simple {
571 * Flush the current log buffer and write 4 bytes of user data to it.
574 struct pmc_op_writelog {
575 uint32_t pm_userdata;
581 * Retrieve the machine specific address assoicated with the allocated
582 * PMC. This number can be used subsequently with a read-performance-counter
586 struct pmc_op_getmsr {
587 uint32_t pm_msr; /* machine specific address */
588 pmc_id_t pm_pmcid; /* allocated pmc id */
594 * Retrieve a PMC dynamic class events list.
597 struct pmc_dyn_event_descr {
598 char pm_ev_name[PMC_NAME_MAX];
599 enum pmc_event pm_ev_code;
602 struct pmc_op_getdyneventinfo {
603 enum pmc_class pm_class;
604 unsigned int pm_nevent;
605 struct pmc_dyn_event_descr pm_events[PMC_EV_DYN_COUNT];
610 #include <sys/malloc.h>
611 #include <sys/sysctl.h>
613 #include <machine/frame.h>
615 #define PMC_HASH_SIZE 1024
616 #define PMC_MTXPOOL_SIZE 2048
617 #define PMC_LOG_BUFFER_SIZE 4
618 #define PMC_NLOGBUFFERS 1024
619 #define PMC_NSAMPLES 1024
620 #define PMC_CALLCHAIN_DEPTH 32
622 #define PMC_SYSCTL_NAME_PREFIX "kern." PMC_MODULE_NAME "."
627 * (b) - pmc_bufferlist_mtx (spin lock)
628 * (k) - pmc_kthread_mtx (sleep lock)
629 * (o) - po->po_mtx (spin lock)
636 struct pmc_syscall_args {
637 register_t pmop_code; /* one of PMC_OP_* */
638 void *pmop_data; /* syscall parameter */
642 * Interface to processor specific s1tuff
648 * Machine independent (i.e., the common parts) of a human readable
653 char pd_name[PMC_NAME_MAX]; /* name */
654 uint32_t pd_caps; /* capabilities */
655 enum pmc_class pd_class; /* class of the PMC */
656 uint32_t pd_width; /* width in bits */
662 * This structure records all the target processes associated with a
667 LIST_ENTRY(pmc_target) pt_next;
668 struct pmc_process *pt_process; /* target descriptor */
674 * Describes each allocated PMC.
676 * Each PMC has precisely one owner, namely the process that allocated
679 * A PMC may be attached to multiple target processes. The
680 * 'pm_targets' field links all the target processes being monitored
683 * The 'pm_savedvalue' field is protected by a mutex.
685 * On a multi-cpu machine, multiple target threads associated with a
686 * process-virtual PMC could be concurrently executing on different
687 * CPUs. The 'pm_runcount' field is atomically incremented every time
688 * the PMC gets scheduled on a CPU and atomically decremented when it
689 * get descheduled. Deletion of a PMC is only permitted when this
695 LIST_HEAD(,pmc_target) pm_targets; /* list of target processes */
696 LIST_ENTRY(pmc) pm_next; /* owner's list */
699 * System-wide PMCs are allocated on a CPU and are not moved
700 * around. For system-wide PMCs we record the CPU the PMC was
701 * allocated on in the 'CPU' field of the pmc ID.
703 * Virtual PMCs run on whichever CPU is currently executing
704 * their targets' threads. For these PMCs we need to save
705 * their current PMC counter values when they are taken off
710 pmc_value_t pm_savedvalue; /* Virtual PMCS */
714 * For sampling mode PMCs, we keep track of the PMC's "reload
715 * count", which is the counter value to be loaded in when
716 * arming the PMC for the next counting session. For counting
717 * modes on PMCs that are read-only (e.g., the x86 TSC), we
718 * keep track of the initial value at the start of
719 * counting-mode operation.
723 pmc_value_t pm_reloadcount; /* sampling PMC modes */
724 pmc_value_t pm_initial; /* counting PMC modes */
727 uint32_t pm_stalled; /* marks stalled sampling PMCs */
728 uint32_t pm_caps; /* PMC capabilities */
729 enum pmc_event pm_event; /* event being measured */
730 uint32_t pm_flags; /* additional flags PMC_F_... */
731 struct pmc_owner *pm_owner; /* owner thread state */
732 int pm_runcount; /* #cpus currently on */
733 enum pmc_state pm_state; /* current PMC state */
736 * The PMC ID field encodes the row-index for the PMC, its
737 * mode, class and the CPU# associated with the PMC.
740 pmc_id_t pm_id; /* allocated PMC id */
743 union pmc_md_pmc pm_md;
747 * Accessor macros for 'struct pmc'
750 #define PMC_TO_MODE(P) PMC_ID_TO_MODE((P)->pm_id)
751 #define PMC_TO_CLASS(P) PMC_ID_TO_CLASS((P)->pm_id)
752 #define PMC_TO_ROWINDEX(P) PMC_ID_TO_ROWINDEX((P)->pm_id)
753 #define PMC_TO_CPU(P) PMC_ID_TO_CPU((P)->pm_id)
759 * Record a 'target' process being profiled.
761 * The target process being profiled could be different from the owner
762 * process which allocated the PMCs. Each target process descriptor
763 * is associated with NHWPMC 'struct pmc *' pointers. Each PMC at a
764 * given hardware row-index 'n' will use slot 'n' of the 'pp_pmcs[]'
765 * array. The size of this structure is thus PMC architecture
770 struct pmc_targetstate {
771 struct pmc *pp_pmc; /* target PMC */
772 pmc_value_t pp_pmcval; /* per-process value */
776 LIST_ENTRY(pmc_process) pp_next; /* hash chain */
777 int pp_refcnt; /* reference count */
778 uint32_t pp_flags; /* flags PMC_PP_* */
779 struct proc *pp_proc; /* target thread */
780 struct pmc_targetstate pp_pmcs[]; /* NHWPMCs */
783 #define PMC_PP_ENABLE_MSR_ACCESS 0x00000001
788 * We associate a PMC with an 'owner' process.
790 * A process can be associated with 0..NCPUS*NHWPMC PMCs during its
791 * lifetime, where NCPUS is the numbers of CPUS in the system and
792 * NHWPMC is the number of hardware PMCs per CPU. These are
793 * maintained in the list headed by the 'po_pmcs' to save on space.
798 LIST_ENTRY(pmc_owner) po_next; /* hash chain */
799 LIST_ENTRY(pmc_owner) po_ssnext; /* list of SS PMC owners */
800 LIST_HEAD(, pmc) po_pmcs; /* owned PMC list */
801 TAILQ_HEAD(, pmclog_buffer) po_logbuffers; /* (o) logbuffer list */
802 struct mtx po_mtx; /* spin lock for (o) */
803 struct proc *po_owner; /* owner proc */
804 uint32_t po_flags; /* (k) flags PMC_PO_* */
805 struct proc *po_kthread; /* (k) helper kthread */
806 struct pmclog_buffer *po_curbuf; /* current log buffer */
807 struct file *po_file; /* file reference */
808 int po_error; /* recorded error */
809 short po_sscount; /* # SS PMCs owned */
810 short po_logprocmaps; /* global mappings done */
813 #define PMC_PO_OWNS_LOGFILE 0x00000001 /* has a log file */
814 #define PMC_PO_SHUTDOWN 0x00000010 /* in the process of shutdown */
815 #define PMC_PO_INITIAL_MAPPINGS_DONE 0x00000020
818 * struct pmc_hw -- describe the state of the PMC hardware
820 * When in use, a HW PMC is associated with one allocated 'struct pmc'
821 * pointed to by field 'phw_pmc'. When inactive, this field is NULL.
823 * On an SMP box, one or more HW PMC's in process virtual mode with
824 * the same 'phw_pmc' could be executing on different CPUs. In order
825 * to handle this case correctly, we need to ensure that only
826 * incremental counts get added to the saved value in the associated
827 * 'struct pmc'. The 'phw_save' field is used to keep the saved PMC
828 * value at the time the hardware is started during this context
829 * switch (i.e., the difference between the new (hardware) count and
830 * the saved count is atomically added to the count field in 'struct
831 * pmc' at context switch time).
836 uint32_t phw_state; /* see PHW_* macros below */
837 struct pmc *phw_pmc; /* current thread PMC */
840 #define PMC_PHW_RI_MASK 0x000000FF
841 #define PMC_PHW_CPU_SHIFT 8
842 #define PMC_PHW_CPU_MASK 0x0000FF00
843 #define PMC_PHW_FLAGS_SHIFT 16
844 #define PMC_PHW_FLAGS_MASK 0xFFFF0000
846 #define PMC_PHW_INDEX_TO_STATE(ri) ((ri) & PMC_PHW_RI_MASK)
847 #define PMC_PHW_STATE_TO_INDEX(state) ((state) & PMC_PHW_RI_MASK)
848 #define PMC_PHW_CPU_TO_STATE(cpu) (((cpu) << PMC_PHW_CPU_SHIFT) & \
850 #define PMC_PHW_STATE_TO_CPU(state) (((state) & PMC_PHW_CPU_MASK) >> \
852 #define PMC_PHW_FLAGS_TO_STATE(flags) (((flags) << PMC_PHW_FLAGS_SHIFT) & \
854 #define PMC_PHW_STATE_TO_FLAGS(state) (((state) & PMC_PHW_FLAGS_MASK) >> \
856 #define PMC_PHW_FLAG_IS_ENABLED (PMC_PHW_FLAGS_TO_STATE(0x01))
857 #define PMC_PHW_FLAG_IS_SHAREABLE (PMC_PHW_FLAGS_TO_STATE(0x02))
862 * Space for N (tunable) PC samples and associated control data.
866 uint16_t ps_nsamples; /* callchain depth */
867 uint8_t ps_cpu; /* cpu number */
868 uint8_t ps_flags; /* other flags */
869 pid_t ps_pid; /* process PID or -1 */
870 struct thread *ps_td; /* which thread */
871 struct pmc *ps_pmc; /* interrupting PMC */
872 uintptr_t *ps_pc; /* (const) callchain start */
875 #define PMC_SAMPLE_FREE ((uint16_t) 0)
876 #define PMC_SAMPLE_INUSE ((uint16_t) 0xFFFF)
878 struct pmc_samplebuffer {
879 struct pmc_sample * volatile ps_read; /* read pointer */
880 struct pmc_sample * volatile ps_write; /* write pointer */
881 uintptr_t *ps_callchains; /* all saved call chains */
882 struct pmc_sample *ps_fence; /* one beyond ps_samples[] */
883 struct pmc_sample ps_samples[]; /* array of sample entries */
888 * struct pmc_cpustate
890 * A CPU is modelled as a collection of HW PMCs with space for additional
895 uint32_t pc_state; /* physical cpu number + flags */
896 struct pmc_samplebuffer *pc_sb[2]; /* space for samples */
897 struct pmc_hw *pc_hwpmcs[]; /* 'npmc' pointers */
900 #define PMC_PCPU_CPU_MASK 0x000000FF
901 #define PMC_PCPU_FLAGS_MASK 0xFFFFFF00
902 #define PMC_PCPU_FLAGS_SHIFT 8
903 #define PMC_PCPU_STATE_TO_CPU(S) ((S) & PMC_PCPU_CPU_MASK)
904 #define PMC_PCPU_STATE_TO_FLAGS(S) (((S) & PMC_PCPU_FLAGS_MASK) >> PMC_PCPU_FLAGS_SHIFT)
905 #define PMC_PCPU_FLAGS_TO_STATE(F) (((F) << PMC_PCPU_FLAGS_SHIFT) & PMC_PCPU_FLAGS_MASK)
906 #define PMC_PCPU_CPU_TO_STATE(C) ((C) & PMC_PCPU_CPU_MASK)
907 #define PMC_PCPU_FLAG_HTT (PMC_PCPU_FLAGS_TO_STATE(0x1))
912 * CPU binding information.
916 int pb_bound; /* is bound? */
917 int pb_cpu; /* if so, to which CPU */
924 * struct pmc_classdep
926 * PMC class-dependent operations.
928 struct pmc_classdep {
929 uint32_t pcd_caps; /* class capabilities */
930 enum pmc_class pcd_class; /* class id */
931 int pcd_num; /* number of PMCs */
932 int pcd_ri; /* row index of the first PMC in class */
933 int pcd_width; /* width of the PMC */
935 /* configuring/reading/writing the hardware PMCs */
936 int (*pcd_config_pmc)(int _cpu, int _ri, struct pmc *_pm);
937 int (*pcd_get_config)(int _cpu, int _ri, struct pmc **_ppm);
938 int (*pcd_read_pmc)(int _cpu, int _ri, pmc_value_t *_value);
939 int (*pcd_write_pmc)(int _cpu, int _ri, pmc_value_t _value);
941 /* pmc allocation/release */
942 int (*pcd_allocate_pmc)(int _cpu, int _ri, struct pmc *_t,
943 const struct pmc_op_pmcallocate *_a);
944 int (*pcd_release_pmc)(int _cpu, int _ri, struct pmc *_pm);
946 /* starting and stopping PMCs */
947 int (*pcd_start_pmc)(int _cpu, int _ri);
948 int (*pcd_stop_pmc)(int _cpu, int _ri);
951 int (*pcd_describe)(int _cpu, int _ri, struct pmc_info *_pi,
954 /* class-dependent initialization & finalization */
955 int (*pcd_pcpu_init)(struct pmc_mdep *_md, int _cpu);
956 int (*pcd_pcpu_fini)(struct pmc_mdep *_md, int _cpu);
958 /* machine-specific interface */
959 int (*pcd_get_msr)(int _ri, uint32_t *_msr);
965 * Machine dependent bits needed per CPU type.
969 uint32_t pmd_cputype; /* from enum pmc_cputype */
970 uint32_t pmd_npmc; /* number of PMCs per CPU */
971 uint32_t pmd_nclass; /* number of PMC classes present */
974 * Machine dependent methods.
977 /* per-cpu initialization and finalization */
978 int (*pmd_pcpu_init)(struct pmc_mdep *_md, int _cpu);
979 int (*pmd_pcpu_fini)(struct pmc_mdep *_md, int _cpu);
981 /* thread context switch in/out */
982 int (*pmd_switch_in)(struct pmc_cpu *_p, struct pmc_process *_pp);
983 int (*pmd_switch_out)(struct pmc_cpu *_p, struct pmc_process *_pp);
985 /* handle a PMC interrupt */
986 int (*pmd_intr)(int _cpu, struct trapframe *_tf);
989 * PMC class dependent information.
991 struct pmc_classdep pmd_classdep[];
995 * Per-CPU state. This is an array of 'mp_ncpu' pointers
996 * to struct pmc_cpu descriptors.
999 extern struct pmc_cpu **pmc_pcpu;
1001 /* driver statistics */
1002 extern struct pmc_op_getdriverstats pmc_stats;
1004 #if defined(HWPMC_DEBUG)
1005 #include <sys/ktr.h>
1007 /* debug flags, major flag groups */
1008 struct pmc_debugflags {
1020 extern struct pmc_debugflags pmc_debugflags;
1022 #define KTR_PMC KTR_SUBSYS
1024 #define PMC_DEBUG_STRSIZE 128
1025 #define PMC_DEBUG_DEFAULT_FLAGS { 0, 0, 0, 0, 0, 0, 0, 0 }
1027 #define PMCDBG0(M, N, L, F) do { \
1028 if (pmc_debugflags.pdb_ ## M & (1 << PMC_DEBUG_MIN_ ## N)) \
1029 CTR0(KTR_PMC, #M ":" #N ":" #L ": " F); \
1031 #define PMCDBG1(M, N, L, F, p1) do { \
1032 if (pmc_debugflags.pdb_ ## M & (1 << PMC_DEBUG_MIN_ ## N)) \
1033 CTR1(KTR_PMC, #M ":" #N ":" #L ": " F, p1); \
1035 #define PMCDBG2(M, N, L, F, p1, p2) do { \
1036 if (pmc_debugflags.pdb_ ## M & (1 << PMC_DEBUG_MIN_ ## N)) \
1037 CTR2(KTR_PMC, #M ":" #N ":" #L ": " F, p1, p2); \
1039 #define PMCDBG3(M, N, L, F, p1, p2, p3) do { \
1040 if (pmc_debugflags.pdb_ ## M & (1 << PMC_DEBUG_MIN_ ## N)) \
1041 CTR3(KTR_PMC, #M ":" #N ":" #L ": " F, p1, p2, p3); \
1043 #define PMCDBG4(M, N, L, F, p1, p2, p3, p4) do { \
1044 if (pmc_debugflags.pdb_ ## M & (1 << PMC_DEBUG_MIN_ ## N)) \
1045 CTR4(KTR_PMC, #M ":" #N ":" #L ": " F, p1, p2, p3, p4);\
1047 #define PMCDBG5(M, N, L, F, p1, p2, p3, p4, p5) do { \
1048 if (pmc_debugflags.pdb_ ## M & (1 << PMC_DEBUG_MIN_ ## N)) \
1049 CTR5(KTR_PMC, #M ":" #N ":" #L ": " F, p1, p2, p3, p4, \
1052 #define PMCDBG6(M, N, L, F, p1, p2, p3, p4, p5, p6) do { \
1053 if (pmc_debugflags.pdb_ ## M & (1 << PMC_DEBUG_MIN_ ## N)) \
1054 CTR6(KTR_PMC, #M ":" #N ":" #L ": " F, p1, p2, p3, p4, \
1059 #define PMC_DEBUG_MAJ_CPU 0 /* cpu switches */
1060 #define PMC_DEBUG_MAJ_CSW 1 /* context switches */
1061 #define PMC_DEBUG_MAJ_LOG 2 /* logging */
1062 #define PMC_DEBUG_MAJ_MDP 3 /* machine dependent */
1063 #define PMC_DEBUG_MAJ_MOD 4 /* misc module infrastructure */
1064 #define PMC_DEBUG_MAJ_OWN 5 /* owner */
1065 #define PMC_DEBUG_MAJ_PMC 6 /* pmc management */
1066 #define PMC_DEBUG_MAJ_PRC 7 /* processes */
1067 #define PMC_DEBUG_MAJ_SAM 8 /* sampling */
1071 /* Common (8 bits) */
1072 #define PMC_DEBUG_MIN_ALL 0 /* allocation */
1073 #define PMC_DEBUG_MIN_REL 1 /* release */
1074 #define PMC_DEBUG_MIN_OPS 2 /* ops: start, stop, ... */
1075 #define PMC_DEBUG_MIN_INI 3 /* init */
1076 #define PMC_DEBUG_MIN_FND 4 /* find */
1079 #define PMC_DEBUG_MIN_PMH 14 /* pmc_hook */
1080 #define PMC_DEBUG_MIN_PMS 15 /* pmc_syscall */
1083 #define PMC_DEBUG_MIN_ORM 8 /* owner remove */
1084 #define PMC_DEBUG_MIN_OMR 9 /* owner maybe remove */
1087 #define PMC_DEBUG_MIN_TLK 8 /* link target */
1088 #define PMC_DEBUG_MIN_TUL 9 /* unlink target */
1089 #define PMC_DEBUG_MIN_EXT 10 /* process exit */
1090 #define PMC_DEBUG_MIN_EXC 11 /* process exec */
1091 #define PMC_DEBUG_MIN_FRK 12 /* process fork */
1092 #define PMC_DEBUG_MIN_ATT 13 /* attach/detach */
1093 #define PMC_DEBUG_MIN_SIG 14 /* signalling */
1095 /* CONTEXT SWITCHES */
1096 #define PMC_DEBUG_MIN_SWI 8 /* switch in */
1097 #define PMC_DEBUG_MIN_SWO 9 /* switch out */
1100 #define PMC_DEBUG_MIN_REG 8 /* pmc register */
1101 #define PMC_DEBUG_MIN_ALR 9 /* allocate row */
1103 /* MACHINE DEPENDENT LAYER */
1104 #define PMC_DEBUG_MIN_REA 8 /* read */
1105 #define PMC_DEBUG_MIN_WRI 9 /* write */
1106 #define PMC_DEBUG_MIN_CFG 10 /* config */
1107 #define PMC_DEBUG_MIN_STA 11 /* start */
1108 #define PMC_DEBUG_MIN_STO 12 /* stop */
1109 #define PMC_DEBUG_MIN_INT 13 /* interrupts */
1112 #define PMC_DEBUG_MIN_BND 8 /* bind */
1113 #define PMC_DEBUG_MIN_SEL 9 /* select */
1116 #define PMC_DEBUG_MIN_GTB 8 /* get buf */
1117 #define PMC_DEBUG_MIN_SIO 9 /* schedule i/o */
1118 #define PMC_DEBUG_MIN_FLS 10 /* flush */
1119 #define PMC_DEBUG_MIN_SAM 11 /* sample */
1120 #define PMC_DEBUG_MIN_CLO 12 /* close */
1123 #define PMCDBG0(M, N, L, F) /* nothing */
1124 #define PMCDBG1(M, N, L, F, p1)
1125 #define PMCDBG2(M, N, L, F, p1, p2)
1126 #define PMCDBG3(M, N, L, F, p1, p2, p3)
1127 #define PMCDBG4(M, N, L, F, p1, p2, p3, p4)
1128 #define PMCDBG5(M, N, L, F, p1, p2, p3, p4, p5)
1129 #define PMCDBG6(M, N, L, F, p1, p2, p3, p4, p5, p6)
1132 /* declare a dedicated memory pool */
1133 MALLOC_DECLARE(M_PMC);
1139 struct pmc_mdep *pmc_md_initialize(void); /* MD init function */
1140 void pmc_md_finalize(struct pmc_mdep *_md); /* MD fini function */
1141 int pmc_getrowdisp(int _ri);
1142 int pmc_process_interrupt(int _cpu, int _soft, struct pmc *_pm,
1143 struct trapframe *_tf, int _inuserspace);
1144 int pmc_save_kernel_callchain(uintptr_t *_cc, int _maxsamples,
1145 struct trapframe *_tf);
1146 int pmc_save_user_callchain(uintptr_t *_cc, int _maxsamples,
1147 struct trapframe *_tf);
1148 struct pmc_mdep *pmc_mdep_alloc(int nclasses);
1149 void pmc_mdep_free(struct pmc_mdep *md);
1150 #endif /* _KERNEL */
1151 #endif /* _SYS_PMC_H_ */