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 16 /* HW counter name size */
43 #define PMC_CLASS_MAX 6 /* 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(MIPS_24K, 0x200, "MIPS 24K")
93 #define __PMC_CPU(S,V,D) PMC_CPU_##S = V,
97 #define PMC_CPU_FIRST PMC_CPU_AMD_K7
98 #define PMC_CPU_LAST PMC_CPU_MIPS_24K
104 #define __PMC_CLASSES() \
105 __PMC_CLASS(TSC) /* CPU Timestamp counter */ \
106 __PMC_CLASS(K7) /* AMD K7 performance counters */ \
107 __PMC_CLASS(K8) /* AMD K8 performance counters */ \
108 __PMC_CLASS(P5) /* Intel Pentium counters */ \
109 __PMC_CLASS(P6) /* Intel Pentium Pro counters */ \
110 __PMC_CLASS(P4) /* Intel Pentium-IV counters */ \
111 __PMC_CLASS(IAF) /* Intel Core2/Atom, fixed function */ \
112 __PMC_CLASS(IAP) /* Intel Core...Atom, programmable */ \
113 __PMC_CLASS(UCF) /* Intel Uncore fixed function */ \
114 __PMC_CLASS(UCP) /* Intel Uncore programmable */ \
115 __PMC_CLASS(MIPS24K) /* MIPS 24K */
119 #define __PMC_CLASS(N) PMC_CLASS_##N ,
123 #define PMC_CLASS_FIRST PMC_CLASS_TSC
124 #define PMC_CLASS_LAST PMC_CLASS_MIPS24K
127 * A PMC can be in the following states:
130 * DISABLED -- administratively prohibited from being used.
131 * FREE -- HW available for use
133 * ALLOCATED -- allocated
134 * STOPPED -- allocated, but not counting events
135 * RUNNING -- allocated, and in operation; 'pm_runcount'
136 * holds the number of CPUs using this PMC at
138 * DELETED -- being destroyed
141 #define __PMC_HWSTATES() \
142 __PMC_STATE(DISABLED) \
145 #define __PMC_SWSTATES() \
146 __PMC_STATE(ALLOCATED) \
147 __PMC_STATE(STOPPED) \
148 __PMC_STATE(RUNNING) \
151 #define __PMC_STATES() \
157 #define __PMC_STATE(S) PMC_STATE_##S,
162 #define PMC_STATE_FIRST PMC_STATE_DISABLED
163 #define PMC_STATE_LAST PMC_STATE_DELETED
166 * An allocated PMC may used as a 'global' counter or as a
167 * 'thread-private' one. Each such mode of use can be in either
168 * statistical sampling mode or in counting mode. Thus a PMC in use
170 * SS i.e., SYSTEM STATISTICAL -- system-wide statistical profiling
171 * SC i.e., SYSTEM COUNTER -- system-wide counting mode
172 * TS i.e., THREAD STATISTICAL -- thread virtual, statistical profiling
173 * TC i.e., THREAD COUNTER -- thread virtual, counting mode
175 * Statistical profiling modes rely on the PMC periodically delivering
176 * a interrupt to the CPU (when the configured number of events have
177 * been measured), so the PMC must have the ability to generate
180 * In counting modes, the PMC counts its configured events, with the
181 * value of the PMC being read whenever needed by its owner process.
183 * The thread specific modes "virtualize" the PMCs -- the PMCs appear
184 * to be thread private and count events only when the profiled thread
185 * actually executes on the CPU.
187 * The system-wide "global" modes keep the PMCs running all the time
188 * and are used to measure the behaviour of the whole system.
191 #define __PMC_MODES() \
199 #define __PMC_MODE(M,N) PMC_MODE_##M = N,
203 #define PMC_MODE_FIRST PMC_MODE_SS
204 #define PMC_MODE_LAST PMC_MODE_TC
206 #define PMC_IS_COUNTING_MODE(mode) \
207 ((mode) == PMC_MODE_SC || (mode) == PMC_MODE_TC)
208 #define PMC_IS_SYSTEM_MODE(mode) \
209 ((mode) == PMC_MODE_SS || (mode) == PMC_MODE_SC)
210 #define PMC_IS_SAMPLING_MODE(mode) \
211 ((mode) == PMC_MODE_SS || (mode) == PMC_MODE_TS)
212 #define PMC_IS_VIRTUAL_MODE(mode) \
213 ((mode) == PMC_MODE_TS || (mode) == PMC_MODE_TC)
216 * PMC row disposition
219 #define __PMC_DISPOSITIONS(N) \
220 __PMC_DISP(STANDALONE) /* global/disabled counters */ \
221 __PMC_DISP(FREE) /* free/available */ \
222 __PMC_DISP(THREAD) /* thread-virtual PMCs */ \
223 __PMC_DISP(UNKNOWN) /* sentinel */
227 #define __PMC_DISP(D) PMC_DISP_##D ,
231 #define PMC_DISP_FIRST PMC_DISP_STANDALONE
232 #define PMC_DISP_LAST PMC_DISP_THREAD
235 * Counter capabilities
237 * __PMC_CAPS(NAME, VALUE, DESCRIPTION)
240 #define __PMC_CAPS() \
241 __PMC_CAP(INTERRUPT, 0, "generate interrupts") \
242 __PMC_CAP(USER, 1, "count user-mode events") \
243 __PMC_CAP(SYSTEM, 2, "count system-mode events") \
244 __PMC_CAP(EDGE, 3, "do edge detection of events") \
245 __PMC_CAP(THRESHOLD, 4, "ignore events below a threshold") \
246 __PMC_CAP(READ, 5, "read PMC counter") \
247 __PMC_CAP(WRITE, 6, "reprogram PMC counter") \
248 __PMC_CAP(INVERT, 7, "invert comparision sense") \
249 __PMC_CAP(QUALIFIER, 8, "further qualify monitored events") \
250 __PMC_CAP(PRECISE, 9, "perform precise sampling") \
251 __PMC_CAP(TAGGING, 10, "tag upstream events") \
252 __PMC_CAP(CASCADE, 11, "cascade counters")
257 #define __PMC_CAP(NAME, VALUE, DESCR) PMC_CAP_##NAME = (1 << VALUE) ,
261 #define PMC_CAP_FIRST PMC_CAP_INTERRUPT
262 #define PMC_CAP_LAST PMC_CAP_CASCADE
267 * These are generated from the definitions in "dev/hwpmc/pmc_events.h".
272 #undef __PMC_EV_BLOCK
273 #define __PMC_EV_BLOCK(C,V) PMC_EV_ ## C ## __BLOCK_START = (V) - 1 ,
274 #define __PMC_EV(C,N) PMC_EV_ ## C ## _ ## N ,
279 * PMC SYSCALL INTERFACE
283 * "PMC_OPS" -- these are the commands recognized by the kernel
284 * module, and are used when performing a system call from userland.
286 #define __PMC_OPS() \
287 __PMC_OP(CONFIGURELOG, "Set log file") \
288 __PMC_OP(FLUSHLOG, "Flush log file") \
289 __PMC_OP(GETCPUINFO, "Get system CPU information") \
290 __PMC_OP(GETDRIVERSTATS, "Get driver statistics") \
291 __PMC_OP(GETMODULEVERSION, "Get module version") \
292 __PMC_OP(GETPMCINFO, "Get per-cpu PMC information") \
293 __PMC_OP(PMCADMIN, "Set PMC state") \
294 __PMC_OP(PMCALLOCATE, "Allocate and configure a PMC") \
295 __PMC_OP(PMCATTACH, "Attach a PMC to a process") \
296 __PMC_OP(PMCDETACH, "Detach a PMC from a process") \
297 __PMC_OP(PMCGETMSR, "Get a PMC's hardware address") \
298 __PMC_OP(PMCRELEASE, "Release a PMC") \
299 __PMC_OP(PMCRW, "Read/Set a PMC") \
300 __PMC_OP(PMCSETCOUNT, "Set initial count/sampling rate") \
301 __PMC_OP(PMCSTART, "Start a PMC") \
302 __PMC_OP(PMCSTOP, "Start a PMC") \
303 __PMC_OP(WRITELOG, "Write a cookie to the log file") \
304 __PMC_OP(CLOSELOG, "Close log file")
309 #define __PMC_OP(N, D) PMC_OP_##N,
315 * Flags used in operations on PMCs.
318 #define PMC_F_FORCE 0x00000001 /*OP ADMIN force operation */
319 #define PMC_F_DESCENDANTS 0x00000002 /*OP ALLOCATE track descendants */
320 #define PMC_F_LOG_PROCCSW 0x00000004 /*OP ALLOCATE track ctx switches */
321 #define PMC_F_LOG_PROCEXIT 0x00000008 /*OP ALLOCATE log proc exits */
322 #define PMC_F_NEWVALUE 0x00000010 /*OP RW write new value */
323 #define PMC_F_OLDVALUE 0x00000020 /*OP RW get old value */
324 #define PMC_F_KGMON 0x00000040 /*OP ALLOCATE kgmon(8) profiling */
326 #define PMC_F_CALLCHAIN 0x00000080 /*OP ALLOCATE capture callchains */
329 #define PMC_F_ATTACHED_TO_OWNER 0x00010000 /*attached to owner*/
330 #define PMC_F_NEEDS_LOGFILE 0x00020000 /*needs log file */
331 #define PMC_F_ATTACH_DONE 0x00040000 /*attached at least once */
333 #define PMC_CALLCHAIN_DEPTH_MAX 32
334 #define PMC_CC_F_USERSPACE 0x01 /*userspace callchain*/
337 * Cookies used to denote allocated PMCs, and the values of PMCs.
340 typedef uint32_t pmc_id_t;
341 typedef uint64_t pmc_value_t;
343 #define PMC_ID_INVALID (~ (pmc_id_t) 0)
346 * PMC IDs have the following format:
348 * +--------+----------+-----------+-----------+
349 * | CPU | PMC MODE | PMC CLASS | ROW INDEX |
350 * +--------+----------+-----------+-----------+
352 * where each field is 8 bits wide. Field 'CPU' is set to the
353 * requested CPU for system-wide PMCs or PMC_CPU_ANY for process-mode
354 * PMCs. Field 'PMC MODE' is the allocated PMC mode. Field 'PMC
355 * CLASS' is the class of the PMC. Field 'ROW INDEX' is the row index
358 * The 'ROW INDEX' ranges over 0..NWPMCS where NHWPMCS is the total
359 * number of hardware PMCs on this cpu.
363 #define PMC_ID_TO_ROWINDEX(ID) ((ID) & 0xFF)
364 #define PMC_ID_TO_CLASS(ID) (((ID) & 0xFF00) >> 8)
365 #define PMC_ID_TO_MODE(ID) (((ID) & 0xFF0000) >> 16)
366 #define PMC_ID_TO_CPU(ID) (((ID) & 0xFF000000) >> 24)
367 #define PMC_ID_MAKE_ID(CPU,MODE,CLASS,ROWINDEX) \
368 ((((CPU) & 0xFF) << 24) | (((MODE) & 0xFF) << 16) | \
369 (((CLASS) & 0xFF) << 8) | ((ROWINDEX) & 0xFF))
372 * Data structures for system calls supported by the pmc driver.
378 * Allocate a PMC on the named CPU.
381 #define PMC_CPU_ANY ~0
383 struct pmc_op_pmcallocate {
384 uint32_t pm_caps; /* PMC_CAP_* */
385 uint32_t pm_cpu; /* CPU number or PMC_CPU_ANY */
386 enum pmc_class pm_class; /* class of PMC desired */
387 enum pmc_event pm_ev; /* [enum pmc_event] desired */
388 uint32_t pm_flags; /* additional modifiers PMC_F_* */
389 enum pmc_mode pm_mode; /* desired mode */
390 pmc_id_t pm_pmcid; /* [return] process pmc id */
392 union pmc_md_op_pmcallocate pm_md; /* MD layer extensions */
398 * Set the administrative state (i.e., whether enabled or disabled) of
399 * a PMC 'pm_pmc' on CPU 'pm_cpu'. Note that 'pm_pmc' specifies an
400 * absolute PMC number and need not have been first allocated by the
404 struct pmc_op_pmcadmin {
405 int pm_cpu; /* CPU# */
406 uint32_t pm_flags; /* flags */
407 int pm_pmc; /* PMC# */
408 enum pmc_state pm_state; /* desired state */
412 * OP PMCATTACH / OP PMCDETACH
414 * Attach/detach a PMC and a process.
417 struct pmc_op_pmcattach {
418 pmc_id_t pm_pmc; /* PMC to attach to */
419 pid_t pm_pid; /* target process */
425 * Set the sampling rate (i.e., the reload count) for statistical counters.
426 * 'pm_pmcid' need to have been previously allocated using PMCALLOCATE.
429 struct pmc_op_pmcsetcount {
430 pmc_value_t pm_count; /* initial/sample count */
431 pmc_id_t pm_pmcid; /* PMC id to set */
438 * Read the value of a PMC named by 'pm_pmcid'. 'pm_pmcid' needs
439 * to have been previously allocated using PMCALLOCATE.
443 struct pmc_op_pmcrw {
444 uint32_t pm_flags; /* PMC_F_{OLD,NEW}VALUE*/
445 pmc_id_t pm_pmcid; /* pmc id */
446 pmc_value_t pm_value; /* new&returned value */
453 * retrieve PMC state for a named CPU. The caller is expected to
454 * allocate 'npmc' * 'struct pmc_info' bytes of space for the return
459 char pm_name[PMC_NAME_MAX]; /* pmc name */
460 enum pmc_class pm_class; /* enum pmc_class */
461 int pm_enabled; /* whether enabled */
462 enum pmc_disp pm_rowdisp; /* FREE, THREAD or STANDLONE */
463 pid_t pm_ownerpid; /* owner, or -1 */
464 enum pmc_mode pm_mode; /* current mode [enum pmc_mode] */
465 enum pmc_event pm_event; /* current event */
466 uint32_t pm_flags; /* current flags */
467 pmc_value_t pm_reloadcount; /* sampling counters only */
470 struct pmc_op_getpmcinfo {
471 int32_t pm_cpu; /* 0 <= cpu < mp_maxid */
472 struct pmc_info pm_pmcs[]; /* space for 'npmc' structures */
479 * Retrieve system CPU information.
482 struct pmc_classinfo {
483 enum pmc_class pm_class; /* class id */
484 uint32_t pm_caps; /* counter capabilities */
485 uint32_t pm_width; /* width of the PMC */
486 uint32_t pm_num; /* number of PMCs in class */
489 struct pmc_op_getcpuinfo {
490 enum pmc_cputype pm_cputype; /* what kind of CPU */
491 uint32_t pm_ncpu; /* max CPU number */
492 uint32_t pm_npmc; /* #PMCs per CPU */
493 uint32_t pm_nclass; /* #classes of PMCs */
494 struct pmc_classinfo pm_classes[PMC_CLASS_MAX];
500 * Configure a log file for writing system-wide statistics to.
503 struct pmc_op_configurelog {
505 int pm_logfd; /* logfile fd (or -1) */
511 * Retrieve pmc(4) driver-wide statistics.
514 struct pmc_op_getdriverstats {
515 int pm_intr_ignored; /* #interrupts ignored */
516 int pm_intr_processed; /* #interrupts processed */
517 int pm_intr_bufferfull; /* #interrupts with ENOSPC */
518 int pm_syscalls; /* #syscalls */
519 int pm_syscall_errors; /* #syscalls with errors */
520 int pm_buffer_requests; /* #buffer requests */
521 int pm_buffer_requests_failed; /* #failed buffer requests */
522 int pm_log_sweeps; /* #sample buffer processing passes */
526 * OP RELEASE / OP START / OP STOP
528 * Simple operations on a PMC id.
531 struct pmc_op_simple {
538 * Flush the current log buffer and write 4 bytes of user data to it.
541 struct pmc_op_writelog {
542 uint32_t pm_userdata;
548 * Retrieve the machine specific address assoicated with the allocated
549 * PMC. This number can be used subsequently with a read-performance-counter
553 struct pmc_op_getmsr {
554 uint32_t pm_msr; /* machine specific address */
555 pmc_id_t pm_pmcid; /* allocated pmc id */
560 #include <sys/malloc.h>
561 #include <sys/sysctl.h>
563 #include <machine/frame.h>
565 #define PMC_HASH_SIZE 16
566 #define PMC_MTXPOOL_SIZE 32
567 #define PMC_LOG_BUFFER_SIZE 4
568 #define PMC_NLOGBUFFERS 16
569 #define PMC_NSAMPLES 32
570 #define PMC_CALLCHAIN_DEPTH 8
572 #define PMC_SYSCTL_NAME_PREFIX "kern." PMC_MODULE_NAME "."
577 * (b) - pmc_bufferlist_mtx (spin lock)
578 * (k) - pmc_kthread_mtx (sleep lock)
579 * (o) - po->po_mtx (spin lock)
586 struct pmc_syscall_args {
587 uint32_t pmop_code; /* one of PMC_OP_* */
588 void *pmop_data; /* syscall parameter */
592 * Interface to processor specific s1tuff
598 * Machine independent (i.e., the common parts) of a human readable
603 char pd_name[PMC_NAME_MAX]; /* name */
604 uint32_t pd_caps; /* capabilities */
605 enum pmc_class pd_class; /* class of the PMC */
606 uint32_t pd_width; /* width in bits */
612 * This structure records all the target processes associated with a
617 LIST_ENTRY(pmc_target) pt_next;
618 struct pmc_process *pt_process; /* target descriptor */
624 * Describes each allocated PMC.
626 * Each PMC has precisely one owner, namely the process that allocated
629 * A PMC may be attached to multiple target processes. The
630 * 'pm_targets' field links all the target processes being monitored
633 * The 'pm_savedvalue' field is protected by a mutex.
635 * On a multi-cpu machine, multiple target threads associated with a
636 * process-virtual PMC could be concurrently executing on different
637 * CPUs. The 'pm_runcount' field is atomically incremented every time
638 * the PMC gets scheduled on a CPU and atomically decremented when it
639 * get descheduled. Deletion of a PMC is only permitted when this
645 LIST_HEAD(,pmc_target) pm_targets; /* list of target processes */
646 LIST_ENTRY(pmc) pm_next; /* owner's list */
649 * System-wide PMCs are allocated on a CPU and are not moved
650 * around. For system-wide PMCs we record the CPU the PMC was
651 * allocated on in the 'CPU' field of the pmc ID.
653 * Virtual PMCs run on whichever CPU is currently executing
654 * their targets' threads. For these PMCs we need to save
655 * their current PMC counter values when they are taken off
660 pmc_value_t pm_savedvalue; /* Virtual PMCS */
664 * For sampling mode PMCs, we keep track of the PMC's "reload
665 * count", which is the counter value to be loaded in when
666 * arming the PMC for the next counting session. For counting
667 * modes on PMCs that are read-only (e.g., the x86 TSC), we
668 * keep track of the initial value at the start of
669 * counting-mode operation.
673 pmc_value_t pm_reloadcount; /* sampling PMC modes */
674 pmc_value_t pm_initial; /* counting PMC modes */
677 uint32_t pm_stalled; /* marks stalled sampling PMCs */
678 uint32_t pm_caps; /* PMC capabilities */
679 enum pmc_event pm_event; /* event being measured */
680 uint32_t pm_flags; /* additional flags PMC_F_... */
681 struct pmc_owner *pm_owner; /* owner thread state */
682 int pm_runcount; /* #cpus currently on */
683 enum pmc_state pm_state; /* current PMC state */
686 * The PMC ID field encodes the row-index for the PMC, its
687 * mode, class and the CPU# associated with the PMC.
690 pmc_id_t pm_id; /* allocated PMC id */
693 union pmc_md_pmc pm_md;
697 * Accessor macros for 'struct pmc'
700 #define PMC_TO_MODE(P) PMC_ID_TO_MODE((P)->pm_id)
701 #define PMC_TO_CLASS(P) PMC_ID_TO_CLASS((P)->pm_id)
702 #define PMC_TO_ROWINDEX(P) PMC_ID_TO_ROWINDEX((P)->pm_id)
703 #define PMC_TO_CPU(P) PMC_ID_TO_CPU((P)->pm_id)
709 * Record a 'target' process being profiled.
711 * The target process being profiled could be different from the owner
712 * process which allocated the PMCs. Each target process descriptor
713 * is associated with NHWPMC 'struct pmc *' pointers. Each PMC at a
714 * given hardware row-index 'n' will use slot 'n' of the 'pp_pmcs[]'
715 * array. The size of this structure is thus PMC architecture
720 struct pmc_targetstate {
721 struct pmc *pp_pmc; /* target PMC */
722 pmc_value_t pp_pmcval; /* per-process value */
726 LIST_ENTRY(pmc_process) pp_next; /* hash chain */
727 int pp_refcnt; /* reference count */
728 uint32_t pp_flags; /* flags PMC_PP_* */
729 struct proc *pp_proc; /* target thread */
730 struct pmc_targetstate pp_pmcs[]; /* NHWPMCs */
733 #define PMC_PP_ENABLE_MSR_ACCESS 0x00000001
738 * We associate a PMC with an 'owner' process.
740 * A process can be associated with 0..NCPUS*NHWPMC PMCs during its
741 * lifetime, where NCPUS is the numbers of CPUS in the system and
742 * NHWPMC is the number of hardware PMCs per CPU. These are
743 * maintained in the list headed by the 'po_pmcs' to save on space.
748 LIST_ENTRY(pmc_owner) po_next; /* hash chain */
749 LIST_ENTRY(pmc_owner) po_ssnext; /* list of SS PMC owners */
750 LIST_HEAD(, pmc) po_pmcs; /* owned PMC list */
751 TAILQ_HEAD(, pmclog_buffer) po_logbuffers; /* (o) logbuffer list */
752 struct mtx po_mtx; /* spin lock for (o) */
753 struct proc *po_owner; /* owner proc */
754 uint32_t po_flags; /* (k) flags PMC_PO_* */
755 struct proc *po_kthread; /* (k) helper kthread */
756 struct pmclog_buffer *po_curbuf; /* current log buffer */
757 struct file *po_file; /* file reference */
758 int po_error; /* recorded error */
759 short po_sscount; /* # SS PMCs owned */
760 short po_logprocmaps; /* global mappings done */
763 #define PMC_PO_OWNS_LOGFILE 0x00000001 /* has a log file */
764 #define PMC_PO_SHUTDOWN 0x00000010 /* in the process of shutdown */
765 #define PMC_PO_INITIAL_MAPPINGS_DONE 0x00000020
768 * struct pmc_hw -- describe the state of the PMC hardware
770 * When in use, a HW PMC is associated with one allocated 'struct pmc'
771 * pointed to by field 'phw_pmc'. When inactive, this field is NULL.
773 * On an SMP box, one or more HW PMC's in process virtual mode with
774 * the same 'phw_pmc' could be executing on different CPUs. In order
775 * to handle this case correctly, we need to ensure that only
776 * incremental counts get added to the saved value in the associated
777 * 'struct pmc'. The 'phw_save' field is used to keep the saved PMC
778 * value at the time the hardware is started during this context
779 * switch (i.e., the difference between the new (hardware) count and
780 * the saved count is atomically added to the count field in 'struct
781 * pmc' at context switch time).
786 uint32_t phw_state; /* see PHW_* macros below */
787 struct pmc *phw_pmc; /* current thread PMC */
790 #define PMC_PHW_RI_MASK 0x000000FF
791 #define PMC_PHW_CPU_SHIFT 8
792 #define PMC_PHW_CPU_MASK 0x0000FF00
793 #define PMC_PHW_FLAGS_SHIFT 16
794 #define PMC_PHW_FLAGS_MASK 0xFFFF0000
796 #define PMC_PHW_INDEX_TO_STATE(ri) ((ri) & PMC_PHW_RI_MASK)
797 #define PMC_PHW_STATE_TO_INDEX(state) ((state) & PMC_PHW_RI_MASK)
798 #define PMC_PHW_CPU_TO_STATE(cpu) (((cpu) << PMC_PHW_CPU_SHIFT) & \
800 #define PMC_PHW_STATE_TO_CPU(state) (((state) & PMC_PHW_CPU_MASK) >> \
802 #define PMC_PHW_FLAGS_TO_STATE(flags) (((flags) << PMC_PHW_FLAGS_SHIFT) & \
804 #define PMC_PHW_STATE_TO_FLAGS(state) (((state) & PMC_PHW_FLAGS_MASK) >> \
806 #define PMC_PHW_FLAG_IS_ENABLED (PMC_PHW_FLAGS_TO_STATE(0x01))
807 #define PMC_PHW_FLAG_IS_SHAREABLE (PMC_PHW_FLAGS_TO_STATE(0x02))
812 * Space for N (tunable) PC samples and associated control data.
816 uint16_t ps_nsamples; /* callchain depth */
817 uint8_t ps_cpu; /* cpu number */
818 uint8_t ps_flags; /* other flags */
819 pid_t ps_pid; /* process PID or -1 */
820 struct thread *ps_td; /* which thread */
821 struct pmc *ps_pmc; /* interrupting PMC */
822 uintptr_t *ps_pc; /* (const) callchain start */
825 #define PMC_SAMPLE_FREE ((uint16_t) 0)
826 #define PMC_SAMPLE_INUSE ((uint16_t) 0xFFFF)
828 struct pmc_samplebuffer {
829 struct pmc_sample * volatile ps_read; /* read pointer */
830 struct pmc_sample * volatile ps_write; /* write pointer */
831 uintptr_t *ps_callchains; /* all saved call chains */
832 struct pmc_sample *ps_fence; /* one beyond ps_samples[] */
833 struct pmc_sample ps_samples[]; /* array of sample entries */
838 * struct pmc_cpustate
840 * A CPU is modelled as a collection of HW PMCs with space for additional
845 uint32_t pc_state; /* physical cpu number + flags */
846 struct pmc_samplebuffer *pc_sb; /* space for samples */
847 struct pmc_hw *pc_hwpmcs[]; /* 'npmc' pointers */
850 #define PMC_PCPU_CPU_MASK 0x000000FF
851 #define PMC_PCPU_FLAGS_MASK 0xFFFFFF00
852 #define PMC_PCPU_FLAGS_SHIFT 8
853 #define PMC_PCPU_STATE_TO_CPU(S) ((S) & PMC_PCPU_CPU_MASK)
854 #define PMC_PCPU_STATE_TO_FLAGS(S) (((S) & PMC_PCPU_FLAGS_MASK) >> PMC_PCPU_FLAGS_SHIFT)
855 #define PMC_PCPU_FLAGS_TO_STATE(F) (((F) << PMC_PCPU_FLAGS_SHIFT) & PMC_PCPU_FLAGS_MASK)
856 #define PMC_PCPU_CPU_TO_STATE(C) ((C) & PMC_PCPU_CPU_MASK)
857 #define PMC_PCPU_FLAG_HTT (PMC_PCPU_FLAGS_TO_STATE(0x1))
862 * CPU binding information.
866 int pb_bound; /* is bound? */
867 int pb_cpu; /* if so, to which CPU */
874 * struct pmc_classdep
876 * PMC class-dependent operations.
878 struct pmc_classdep {
879 uint32_t pcd_caps; /* class capabilities */
880 enum pmc_class pcd_class; /* class id */
881 int pcd_num; /* number of PMCs */
882 int pcd_ri; /* row index of the first PMC in class */
883 int pcd_width; /* width of the PMC */
885 /* configuring/reading/writing the hardware PMCs */
886 int (*pcd_config_pmc)(int _cpu, int _ri, struct pmc *_pm);
887 int (*pcd_get_config)(int _cpu, int _ri, struct pmc **_ppm);
888 int (*pcd_read_pmc)(int _cpu, int _ri, pmc_value_t *_value);
889 int (*pcd_write_pmc)(int _cpu, int _ri, pmc_value_t _value);
891 /* pmc allocation/release */
892 int (*pcd_allocate_pmc)(int _cpu, int _ri, struct pmc *_t,
893 const struct pmc_op_pmcallocate *_a);
894 int (*pcd_release_pmc)(int _cpu, int _ri, struct pmc *_pm);
896 /* starting and stopping PMCs */
897 int (*pcd_start_pmc)(int _cpu, int _ri);
898 int (*pcd_stop_pmc)(int _cpu, int _ri);
901 int (*pcd_describe)(int _cpu, int _ri, struct pmc_info *_pi,
904 /* class-dependent initialization & finalization */
905 int (*pcd_pcpu_init)(struct pmc_mdep *_md, int _cpu);
906 int (*pcd_pcpu_fini)(struct pmc_mdep *_md, int _cpu);
908 /* machine-specific interface */
909 int (*pcd_get_msr)(int _ri, uint32_t *_msr);
915 * Machine dependent bits needed per CPU type.
919 uint32_t pmd_cputype; /* from enum pmc_cputype */
920 uint32_t pmd_npmc; /* number of PMCs per CPU */
921 uint32_t pmd_nclass; /* number of PMC classes present */
924 * Machine dependent methods.
927 /* per-cpu initialization and finalization */
928 int (*pmd_pcpu_init)(struct pmc_mdep *_md, int _cpu);
929 int (*pmd_pcpu_fini)(struct pmc_mdep *_md, int _cpu);
931 /* thread context switch in/out */
932 int (*pmd_switch_in)(struct pmc_cpu *_p, struct pmc_process *_pp);
933 int (*pmd_switch_out)(struct pmc_cpu *_p, struct pmc_process *_pp);
935 /* handle a PMC interrupt */
936 int (*pmd_intr)(int _cpu, struct trapframe *_tf);
939 * PMC class dependent information.
941 struct pmc_classdep pmd_classdep[];
945 * Per-CPU state. This is an array of 'mp_ncpu' pointers
946 * to struct pmc_cpu descriptors.
949 extern struct pmc_cpu **pmc_pcpu;
951 /* driver statistics */
952 extern struct pmc_op_getdriverstats pmc_stats;
954 #if defined(DEBUG) && DEBUG
956 /* debug flags, major flag groups */
957 struct pmc_debugflags {
969 extern struct pmc_debugflags pmc_debugflags;
971 #define PMC_DEBUG_STRSIZE 128
972 #define PMC_DEBUG_DEFAULT_FLAGS { 0, 0, 0, 0, 0, 0, 0, 0 }
974 #define PMCDBG(M,N,L,F,...) do { \
975 if (pmc_debugflags.pdb_ ## M & (1 << PMC_DEBUG_MIN_ ## N)) \
976 printf(#M ":" #N ":" #L ": " F "\n", __VA_ARGS__); \
980 #define PMC_DEBUG_MAJ_CPU 0 /* cpu switches */
981 #define PMC_DEBUG_MAJ_CSW 1 /* context switches */
982 #define PMC_DEBUG_MAJ_LOG 2 /* logging */
983 #define PMC_DEBUG_MAJ_MDP 3 /* machine dependent */
984 #define PMC_DEBUG_MAJ_MOD 4 /* misc module infrastructure */
985 #define PMC_DEBUG_MAJ_OWN 5 /* owner */
986 #define PMC_DEBUG_MAJ_PMC 6 /* pmc management */
987 #define PMC_DEBUG_MAJ_PRC 7 /* processes */
988 #define PMC_DEBUG_MAJ_SAM 8 /* sampling */
992 /* Common (8 bits) */
993 #define PMC_DEBUG_MIN_ALL 0 /* allocation */
994 #define PMC_DEBUG_MIN_REL 1 /* release */
995 #define PMC_DEBUG_MIN_OPS 2 /* ops: start, stop, ... */
996 #define PMC_DEBUG_MIN_INI 3 /* init */
997 #define PMC_DEBUG_MIN_FND 4 /* find */
1000 #define PMC_DEBUG_MIN_PMH 14 /* pmc_hook */
1001 #define PMC_DEBUG_MIN_PMS 15 /* pmc_syscall */
1004 #define PMC_DEBUG_MIN_ORM 8 /* owner remove */
1005 #define PMC_DEBUG_MIN_OMR 9 /* owner maybe remove */
1008 #define PMC_DEBUG_MIN_TLK 8 /* link target */
1009 #define PMC_DEBUG_MIN_TUL 9 /* unlink target */
1010 #define PMC_DEBUG_MIN_EXT 10 /* process exit */
1011 #define PMC_DEBUG_MIN_EXC 11 /* process exec */
1012 #define PMC_DEBUG_MIN_FRK 12 /* process fork */
1013 #define PMC_DEBUG_MIN_ATT 13 /* attach/detach */
1014 #define PMC_DEBUG_MIN_SIG 14 /* signalling */
1016 /* CONTEXT SWITCHES */
1017 #define PMC_DEBUG_MIN_SWI 8 /* switch in */
1018 #define PMC_DEBUG_MIN_SWO 9 /* switch out */
1021 #define PMC_DEBUG_MIN_REG 8 /* pmc register */
1022 #define PMC_DEBUG_MIN_ALR 9 /* allocate row */
1024 /* MACHINE DEPENDENT LAYER */
1025 #define PMC_DEBUG_MIN_REA 8 /* read */
1026 #define PMC_DEBUG_MIN_WRI 9 /* write */
1027 #define PMC_DEBUG_MIN_CFG 10 /* config */
1028 #define PMC_DEBUG_MIN_STA 11 /* start */
1029 #define PMC_DEBUG_MIN_STO 12 /* stop */
1030 #define PMC_DEBUG_MIN_INT 13 /* interrupts */
1033 #define PMC_DEBUG_MIN_BND 8 /* bind */
1034 #define PMC_DEBUG_MIN_SEL 9 /* select */
1037 #define PMC_DEBUG_MIN_GTB 8 /* get buf */
1038 #define PMC_DEBUG_MIN_SIO 9 /* schedule i/o */
1039 #define PMC_DEBUG_MIN_FLS 10 /* flush */
1040 #define PMC_DEBUG_MIN_SAM 11 /* sample */
1041 #define PMC_DEBUG_MIN_CLO 12 /* close */
1044 #define PMCDBG(M,N,L,F,...) /* nothing */
1047 /* declare a dedicated memory pool */
1048 MALLOC_DECLARE(M_PMC);
1054 struct pmc_mdep *pmc_md_initialize(void); /* MD init function */
1055 void pmc_md_finalize(struct pmc_mdep *_md); /* MD fini function */
1056 int pmc_getrowdisp(int _ri);
1057 int pmc_process_interrupt(int _cpu, struct pmc *_pm,
1058 struct trapframe *_tf, int _inuserspace);
1059 int pmc_save_kernel_callchain(uintptr_t *_cc, int _maxsamples,
1060 struct trapframe *_tf);
1061 int pmc_save_user_callchain(uintptr_t *_cc, int _maxsamples,
1062 struct trapframe *_tf);
1063 #endif /* _KERNEL */
1064 #endif /* _SYS_PMC_H_ */