ClickHouse/libs/libcpuid/include/cpuid/libcpuid.h

886 lines
36 KiB
C
Raw Normal View History

/*
* Copyright 2008 Veselin Georgiev,
* anrieffNOSPAM @ mgail_DOT.com (convert to gmail)
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
*
* 1. Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* 2. Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
*
* THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
* IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
* OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
* IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
* INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
* NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
* DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
* THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
* THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*/
#ifndef __LIBCPUID_H__
#define __LIBCPUID_H__
/**
* @File libcpuid.h
* @Author Veselin Georgiev
* @Date Oct 2008
* @Version 0.2.1
*
* Version history:
*
* 0.1.0 (2008-10-15): initial adaptation from wxfractgui sources
* 0.1.1 (2009-07-06): Added intel_fn11 fields to cpu_raw_data_t to handle
* new processor topology enumeration required on Core i7
* 0.1.2 (2009-09-26): Added support for MSR reading through self-extracting
* kernel driver on Win32.
* 0.1.3 (2010-04-20): Added support for greater more accurate CPU clock
* measurements with cpu_clock_by_ic()
* 0.2.0 (2011-10-11): Support for AMD Bulldozer CPUs, 128-bit SSE unit size
* checking. A backwards-incompatible change, since the
* sizeof cpu_id_t is now different.
* 0.2.1 (2012-05-26): Support for Ivy Bridge, and detecting the presence of
* the RdRand instruction.
*/
/** @mainpage A simple libcpuid introduction
*
* LibCPUID provides CPU identification and access to the CPUID and RDTSC
* instructions on the x86.
* <p>
* To execute CPUID, use \ref cpu_exec_cpuid <br>
* To execute RDTSC, use \ref cpu_rdtsc <br>
* To fetch the CPUID info needed for CPU identification, use
* \ref cpuid_get_raw_data <br>
* To make sense of that data (decode, extract features), use \ref cpu_identify <br>
* </p>
*/
/** @defgroup libcpuid LibCPUID
@{ */
/* Include some integer type specifications: */
#include "libcpuid_types.h"
/* Some limits and other constants */
#include "libcpuid_constants.h"
#ifdef __cplusplus
extern "C" {
#endif
/**
* @brief CPU vendor, as guessed from the Vendor String.
*/
typedef enum {
VENDOR_INTEL = 0, /*!< Intel CPU */
VENDOR_AMD, /*!< AMD CPU */
VENDOR_CYRIX, /*!< Cyrix CPU */
VENDOR_NEXGEN, /*!< NexGen CPU */
VENDOR_TRANSMETA, /*!< Transmeta CPU */
VENDOR_UMC, /*!< x86 CPU by UMC */
VENDOR_CENTAUR, /*!< x86 CPU by IDT */
VENDOR_RISE, /*!< x86 CPU by Rise Technology */
VENDOR_SIS, /*!< x86 CPU by SiS */
VENDOR_NSC, /*!< x86 CPU by National Semiconductor */
NUM_CPU_VENDORS, /*!< Valid CPU vendor ids: 0..NUM_CPU_VENDORS - 1 */
VENDOR_UNKNOWN = -1,
} cpu_vendor_t;
#define NUM_CPU_VENDORS NUM_CPU_VENDORS
/**
* @brief Contains just the raw CPUID data.
*
* This contains only the most basic CPU data, required to do identification
* and feature recognition. Every processor should be identifiable using this
* data only.
*/
struct cpu_raw_data_t {
/** contains results of CPUID for eax = 0, 1, ...*/
uint32_t basic_cpuid[MAX_CPUID_LEVEL][4];
/** contains results of CPUID for eax = 0x80000000, 0x80000001, ...*/
uint32_t ext_cpuid[MAX_EXT_CPUID_LEVEL][4];
/** when the CPU is intel and it supports deterministic cache
information: this contains the results of CPUID for eax = 4
and ecx = 0, 1, ... */
uint32_t intel_fn4[MAX_INTELFN4_LEVEL][4];
/** when the CPU is intel and it supports leaf 0Bh (Extended Topology
enumeration leaf), this stores the result of CPUID with
eax = 11 and ecx = 0, 1, 2... */
uint32_t intel_fn11[MAX_INTELFN11_LEVEL][4];
};
/**
* @brief This contains the recognized CPU features/info
*/
struct cpu_id_t {
/** contains the CPU vendor string, e.g. "GenuineIntel" */
char vendor_str[VENDOR_STR_MAX];
/** contains the brand string, e.g. "Intel(R) Xeon(TM) CPU 2.40GHz" */
char brand_str[BRAND_STR_MAX];
/** contains the recognized CPU vendor */
cpu_vendor_t vendor;
/**
* contain CPU flags. Used to test for features. See
* the CPU_FEATURE_* macros below. @see Features
*/
uint8_t flags[CPU_FLAGS_MAX];
/** CPU family */
int32_t family;
/** CPU model */
int32_t model;
/** CPU stepping */
int32_t stepping;
/** CPU extended family */
int32_t ext_family;
/** CPU extended model */
int32_t ext_model;
/** Number of CPU cores on the current processor */
int32_t num_cores;
/**
* Number of logical processors on the current processor.
* Could be more than the number of physical cores,
* e.g. when the processor has HyperThreading.
*/
int32_t num_logical_cpus;
/**
* The total number of logical processors.
*
* This is num_logical_cpus * {total physical processors in the system}
*
* If you're writing a multithreaded program and you want to run it on
* all CPUs, this is the number of threads you need.
*/
int32_t total_logical_cpus;
/**
* L1 data cache size in KB. Could be zero, if the CPU lacks cache.
* If the size cannot be determined, it will be -1.
*/
int32_t l1_data_cache;
/**
* L1 instruction cache size in KB. Could be zero, if the CPU lacks
* cache. If the size cannot be determined, it will be -1.
* @note On some Intel CPUs, whose instruction cache is in fact
* a trace cache, the size will be expressed in K uOps.
*/
int32_t l1_instruction_cache;
/**
* L2 cache size in KB. Could be zero, if the CPU lacks L2 cache.
* If the size of the cache could not be determined, it will be -1
*/
int32_t l2_cache;
/** L3 cache size in KB. Zero on most systems */
int32_t l3_cache;
/** Cache associativity for the L1 data cache. -1 if undetermined */
int32_t l1_assoc;
/** Cache associativity for the L2 cache. -1 if undetermined */
int32_t l2_assoc;
/** Cache associativity for the L3 cache. -1 if undetermined */
int32_t l3_assoc;
/** Cache-line size for L1 data cache. -1 if undetermined */
int32_t l1_cacheline;
/** Cache-line size for L2 cache. -1 if undetermined */
int32_t l2_cacheline;
/** Cache-line size for L3 cache. -1 if undetermined */
int32_t l3_cacheline;
/**
* The brief and human-friendly CPU codename, which was recognized.<br>
* Examples:
* @code
* +--------+--------+-------+-------+-------+---------------------------------------+-----------------------+
* | Vendor | Family | Model | Step. | Cache | Brand String | cpu_id_t.cpu_codename |
* +--------+--------+-------+-------+-------+---------------------------------------+-----------------------+
* | AMD | 6 | 8 | 0 | 256 | (not available - will be ignored) | "K6-2" |
* | Intel | 15 | 2 | 5 | 512 | "Intel(R) Xeon(TM) CPU 2.40GHz" | "Xeon (Prestonia)" |
* | Intel | 6 | 15 | 11 | 4096 | "Intel(R) Core(TM)2 Duo CPU E6550..." | "Conroe (Core 2 Duo)" |
* | AMD | 15 | 35 | 2 | 1024 | "Dual Core AMD Opteron(tm) Proces..." | "Opteron (Dual Core)" |
* +--------+--------+-------+-------+-------+---------------------------------------+-----------------------+
* @endcode
*/
char cpu_codename[64];
/** SSE execution unit size (64 or 128; -1 if N/A) */
int32_t sse_size;
/**
* contain miscellaneous detection information. Used to test about specifics of
* certain detected features. See CPU_HINT_* macros below. @see Hints
*/
uint8_t detection_hints[CPU_HINTS_MAX];
};
/**
* @brief CPU feature identifiers
*
* Usage:
* @code
* ...
* struct cpu_raw_data_t raw;
* struct cpu_id_t id;
* if (cpuid_get_raw_data(&raw) == 0 && cpu_identify(&raw, &id) == 0) {
* if (id.flags[CPU_FEATURE_SSE2]) {
* // The CPU has SSE2...
* ...
* } else {
* // no SSE2
* }
* } else {
* // processor cannot be determined.
* }
* @endcode
*/
typedef enum {
CPU_FEATURE_FPU = 0, /*!< Floating point unit */
CPU_FEATURE_VME, /*!< Virtual mode extension */
CPU_FEATURE_DE, /*!< Debugging extension */
CPU_FEATURE_PSE, /*!< Page size extension */
CPU_FEATURE_TSC, /*!< Time-stamp counter */
CPU_FEATURE_MSR, /*!< Model-specific regsisters, RDMSR/WRMSR supported */
CPU_FEATURE_PAE, /*!< Physical address extension */
CPU_FEATURE_MCE, /*!< Machine check exception */
CPU_FEATURE_CX8, /*!< CMPXCHG8B instruction supported */
CPU_FEATURE_APIC, /*!< APIC support */
CPU_FEATURE_MTRR, /*!< Memory type range registers */
CPU_FEATURE_SEP, /*!< SYSENTER / SYSEXIT instructions supported */
CPU_FEATURE_PGE, /*!< Page global enable */
CPU_FEATURE_MCA, /*!< Machine check architecture */
CPU_FEATURE_CMOV, /*!< CMOVxx instructions supported */
CPU_FEATURE_PAT, /*!< Page attribute table */
CPU_FEATURE_PSE36, /*!< 36-bit page address extension */
CPU_FEATURE_PN, /*!< Processor serial # implemented (Intel P3 only) */
CPU_FEATURE_CLFLUSH, /*!< CLFLUSH instruction supported */
CPU_FEATURE_DTS, /*!< Debug store supported */
CPU_FEATURE_ACPI, /*!< ACPI support (power states) */
CPU_FEATURE_MMX, /*!< MMX instruction set supported */
CPU_FEATURE_FXSR, /*!< FXSAVE / FXRSTOR supported */
CPU_FEATURE_SSE, /*!< Streaming-SIMD Extensions (SSE) supported */
CPU_FEATURE_SSE2, /*!< SSE2 instructions supported */
CPU_FEATURE_SS, /*!< Self-snoop */
CPU_FEATURE_HT, /*!< Hyper-threading supported (but might be disabled) */
CPU_FEATURE_TM, /*!< Thermal monitor */
CPU_FEATURE_IA64, /*!< IA64 supported (Itanium only) */
CPU_FEATURE_PBE, /*!< Pending-break enable */
CPU_FEATURE_PNI, /*!< PNI (SSE3) instructions supported */
CPU_FEATURE_PCLMUL, /*!< PCLMULQDQ instruction supported */
CPU_FEATURE_DTS64, /*!< 64-bit Debug store supported */
CPU_FEATURE_MONITOR, /*!< MONITOR / MWAIT supported */
CPU_FEATURE_DS_CPL, /*!< CPL Qualified Debug Store */
CPU_FEATURE_VMX, /*!< Virtualization technology supported */
CPU_FEATURE_SMX, /*!< Safer mode exceptions */
CPU_FEATURE_EST, /*!< Enhanced SpeedStep */
CPU_FEATURE_TM2, /*!< Thermal monitor 2 */
CPU_FEATURE_SSSE3, /*!< SSSE3 instructionss supported (this is different from SSE3!) */
CPU_FEATURE_CID, /*!< Context ID supported */
CPU_FEATURE_CX16, /*!< CMPXCHG16B instruction supported */
CPU_FEATURE_XTPR, /*!< Send Task Priority Messages disable */
CPU_FEATURE_PDCM, /*!< Performance capabilities MSR supported */
CPU_FEATURE_DCA, /*!< Direct cache access supported */
CPU_FEATURE_SSE4_1, /*!< SSE 4.1 instructions supported */
CPU_FEATURE_SSE4_2, /*!< SSE 4.2 instructions supported */
CPU_FEATURE_SYSCALL, /*!< SYSCALL / SYSRET instructions supported */
CPU_FEATURE_XD, /*!< Execute disable bit supported */
CPU_FEATURE_MOVBE, /*!< MOVBE instruction supported */
CPU_FEATURE_POPCNT, /*!< POPCNT instruction supported */
CPU_FEATURE_AES, /*!< AES* instructions supported */
CPU_FEATURE_XSAVE, /*!< XSAVE/XRSTOR/etc instructions supported */
CPU_FEATURE_OSXSAVE, /*!< non-privileged copy of OSXSAVE supported */
CPU_FEATURE_AVX, /*!< Advanced vector extensions supported */
CPU_FEATURE_MMXEXT, /*!< AMD MMX-extended instructions supported */
CPU_FEATURE_3DNOW, /*!< AMD 3DNow! instructions supported */
CPU_FEATURE_3DNOWEXT, /*!< AMD 3DNow! extended instructions supported */
CPU_FEATURE_NX, /*!< No-execute bit supported */
CPU_FEATURE_FXSR_OPT, /*!< FFXSR: FXSAVE and FXRSTOR optimizations */
CPU_FEATURE_RDTSCP, /*!< RDTSCP instruction supported (AMD-only) */
CPU_FEATURE_LM, /*!< Long mode (x86_64/EM64T) supported */
CPU_FEATURE_LAHF_LM, /*!< LAHF/SAHF supported in 64-bit mode */
CPU_FEATURE_CMP_LEGACY, /*!< core multi-processing legacy mode */
CPU_FEATURE_SVM, /*!< AMD Secure virtual machine */
CPU_FEATURE_ABM, /*!< LZCNT instruction support */
CPU_FEATURE_MISALIGNSSE,/*!< Misaligned SSE supported */
CPU_FEATURE_SSE4A, /*!< SSE 4a from AMD */
CPU_FEATURE_3DNOWPREFETCH, /*!< PREFETCH/PREFETCHW support */
CPU_FEATURE_OSVW, /*!< OS Visible Workaround (AMD) */
CPU_FEATURE_IBS, /*!< Instruction-based sampling */
CPU_FEATURE_SSE5, /*!< SSE 5 instructions supported (deprecated, will never be 1) */
CPU_FEATURE_SKINIT, /*!< SKINIT / STGI supported */
CPU_FEATURE_WDT, /*!< Watchdog timer support */
CPU_FEATURE_TS, /*!< Temperature sensor */
CPU_FEATURE_FID, /*!< Frequency ID control */
CPU_FEATURE_VID, /*!< Voltage ID control */
CPU_FEATURE_TTP, /*!< THERMTRIP */
CPU_FEATURE_TM_AMD, /*!< AMD-specified hardware thermal control */
CPU_FEATURE_STC, /*!< Software thermal control */
CPU_FEATURE_100MHZSTEPS,/*!< 100 MHz multiplier control */
CPU_FEATURE_HWPSTATE, /*!< Hardware P-state control */
CPU_FEATURE_CONSTANT_TSC, /*!< TSC ticks at constant rate */
CPU_FEATURE_XOP, /*!< The XOP instruction set (same as the old CPU_FEATURE_SSE5) */
CPU_FEATURE_FMA3, /*!< The FMA3 instruction set */
CPU_FEATURE_FMA4, /*!< The FMA4 instruction set */
CPU_FEATURE_TBM, /*!< Trailing bit manipulation instruction support */
CPU_FEATURE_F16C, /*!< 16-bit FP convert instruction support */
CPU_FEATURE_RDRAND, /*!< RdRand instruction */
CPU_FEATURE_X2APIC, /*!< x2APIC, APIC_BASE.EXTD, MSRs 0000_0800h...0000_0BFFh 64-bit ICR (+030h but not +031h), no DFR (+00Eh), SELF_IPI (+040h) also see standard level 0000_000Bh */
CPU_FEATURE_CPB, /*!< Core performance boost */
CPU_FEATURE_APERFMPERF, /*!< MPERF/APERF MSRs support */
CPU_FEATURE_PFI, /*!< Processor Feedback Interface support */
CPU_FEATURE_PA, /*!< Processor accumulator */
CPU_FEATURE_AVX2, /*!< AVX2 instructions */
/* termination: */
NUM_CPU_FEATURES,
} cpu_feature_t;
/**
* @brief CPU detection hints identifiers
*
* Usage: similar to the flags usage
*/
typedef enum {
CPU_HINT_SSE_SIZE_AUTH = 0, /*!< SSE unit size is authoritative (not only a Family/Model guesswork, but based on an actual CPUID bit) */
/* termination */
NUM_CPU_HINTS,
} cpu_hint_t;
/**
* @brief Describes common library error codes
*/
typedef enum {
ERR_OK = 0, /*!< "No error" */
ERR_NO_CPUID = -1, /*!< "CPUID instruction is not supported" */
ERR_NO_RDTSC = -2, /*!< "RDTSC instruction is not supported" */
ERR_NO_MEM = -3, /*!< "Memory allocation failed" */
ERR_OPEN = -4, /*!< "File open operation failed" */
ERR_BADFMT = -5, /*!< "Bad file format" */
ERR_NOT_IMP = -6, /*!< "Not implemented" */
ERR_CPU_UNKN = -7, /*!< "Unsupported processor" */
ERR_NO_RDMSR = -8, /*!< "RDMSR instruction is not supported" */
ERR_NO_DRIVER= -9, /*!< "RDMSR driver error (generic)" */
ERR_NO_PERMS = -10, /*!< "No permissions to install RDMSR driver" */
ERR_EXTRACT = -11, /*!< "Cannot extract RDMSR driver (read only media?)" */
ERR_HANDLE = -12, /*!< "Bad handle" */
ERR_INVMSR = -13, /*!< "Invalid MSR" */
} cpu_error_t;
/**
* @brief Internal structure, used in cpu_tsc_mark, cpu_tsc_unmark and
* cpu_clock_by_mark
*/
struct cpu_mark_t {
uint64_t tsc; /*!< Time-stamp from RDTSC */
uint64_t sys_clock; /*!< In microsecond resolution */
};
/**
* @brief Returns the total number of CPUs even if CPUID is not present
* @retval Number of CPUs available
*/
int cpuid_get_total_cpus(void);
/**
* @brief Checks if the CPUID instruction is supported
* @retval 1 if CPUID is present
* @retval 0 the CPU doesn't have CPUID.
*/
int cpuid_present(void);
/**
* @brief Executes the CPUID instruction
* @param eax - the value of the EAX register when executing CPUID
* @param regs - the results will be stored here. regs[0] = EAX, regs[1] = EBX, ...
* @note CPUID will be executed with EAX set to the given value and EBX, ECX,
* EDX set to zero.
*/
void cpu_exec_cpuid(uint32_t eax, uint32_t* regs);
/**
* @brief Executes the CPUID instruction with the given input registers
* @note This is just a bit more generic version of cpu_exec_cpuid - it allows
* you to control all the registers.
* @param regs - Input/output. Prior to executing CPUID, EAX, EBX, ECX and
* EDX will be set to regs[0], regs[1], regs[2] and regs[3].
* After CPUID, this array will contain the results.
*/
void cpu_exec_cpuid_ext(uint32_t* regs);
/**
* @brief Obtains the raw CPUID data from the current CPU
* @param data - a pointer to cpu_raw_data_t structure
* @returns zero if successful, and some negative number on error.
* The error message can be obtained by calling \ref cpuid_error.
* @see cpu_error_t
*/
int cpuid_get_raw_data(struct cpu_raw_data_t* data);
/**
* @brief Writes the raw CPUID data to a text file
* @param data - a pointer to cpu_raw_data_t structure
* @param filename - the path of the file, where the serialized data should be
* written. If empty, stdout will be used.
* @note This is intended primarily for debugging. On some processor, which is
* not currently supported or not completely recognized by cpu_identify,
* one can still successfully get the raw data and write it to a file.
* libcpuid developers can later import this file and debug the detection
* code as if running on the actual hardware.
* The file is simple text format of "something=value" pairs. Version info
* is also written, but the format is not intended to be neither backward-
* nor forward compatible.
* @returns zero if successful, and some negative number on error.
* The error message can be obtained by calling \ref cpuid_error.
* @see cpu_error_t
*/
int cpuid_serialize_raw_data(struct cpu_raw_data_t* data, const char* filename);
/**
* @brief Reads raw CPUID data from file
* @param data - a pointer to cpu_raw_data_t structure. The deserialized data will
* be written here.
* @param filename - the path of the file, containing the serialized raw data.
* If empty, stdin will be used.
* @note This function may fail, if the file is created by different version of
* the library. Also, see the notes on cpuid_serialize_raw_data.
* @returns zero if successful, and some negative number on error.
* The error message can be obtained by calling \ref cpuid_error.
* @see cpu_error_t
*/
int cpuid_deserialize_raw_data(struct cpu_raw_data_t* data, const char* filename);
/**
* @brief Identifies the CPU
* @param raw - Input - a pointer to the raw CPUID data, which is obtained
* either by cpuid_get_raw_data or cpuid_deserialize_raw_data.
* Can also be NULL, in which case the functions calls
* cpuid_get_raw_data itself.
* @param data - Output - the decoded CPU features/info is written here.
* @note The function will not fail, even if some of the information
* cannot be obtained. Even when the CPU is new and thus unknown to
* libcpuid, some generic info, such as "AMD K9 family CPU" will be
* written to data.cpu_codename, and most other things, such as the
* CPU flags, cache sizes, etc. should be detected correctly anyway.
* However, the function CAN fail, if the CPU is completely alien to
* libcpuid.
* @note While cpu_identify() and cpuid_get_raw_data() are fast for most
* purposes, running them several thousand times per second can hamper
* performance significantly. Specifically, avoid writing "cpu feature
* checker" wrapping function, which calls cpu_identify and returns the
* value of some flag, if that function is going to be called frequently.
* @returns zero if successful, and some negative number on error.
* The error message can be obtained by calling \ref cpuid_error.
* @see cpu_error_t
*/
int cpu_identify(struct cpu_raw_data_t* raw, struct cpu_id_t* data);
/**
* @brief Returns the short textual representation of a CPU flag
* @param feature - the feature, whose textual representation is wanted.
* @returns a constant string like "fpu", "tsc", "sse2", etc.
* @note the names of the returned flags are compatible with those from
* /proc/cpuinfo in Linux, with the exception of `tm_amd'
*/
const char* cpu_feature_str(cpu_feature_t feature);
/**
* @brief Returns textual description of the last error
*
* libcpuid stores an `errno'-style error status, whose description
* can be obtained with this function.
* @note This function is not thread-safe
* @see cpu_error_t
*/
const char* cpuid_error(void);
/**
* @brief Executes RDTSC
*
* The RDTSC (ReaD Time Stamp Counter) instruction gives access to an
* internal 64-bit counter, which usually increments at each clock cycle.
* This can be used for various timing routines, and as a very precise
* clock source. It is set to zero on system startup. Beware that may not
* increment at the same frequency as the CPU. Consecutive calls of RDTSC
* are, however, guaranteed to return monotonically-increasing values.
*
* @param result - a pointer to a 64-bit unsigned integer, where the TSC value
* will be stored
*
* @note If 100% compatibility is a concern, you must first check if the
* RDTSC instruction is present (if it is not, your program will crash
* with "invalid opcode" exception). Only some very old processors (i486,
* early AMD K5 and some Cyrix CPUs) lack that instruction - they should
* have become exceedingly rare these days. To verify RDTSC presence,
* run cpu_identify() and check flags[CPU_FEATURE_TSC].
*
* @note The monotonically increasing nature of the TSC may be violated
* on SMP systems, if their TSC clocks run at different rate. If the OS
* doesn't account for that, the TSC drift may become arbitrary large.
*/
void cpu_rdtsc(uint64_t* result);
/**
* @brief Store TSC and timing info
*
* This function stores the current TSC value and current
* time info from a precise OS-specific clock source in the cpu_mark_t
* structure. The sys_clock field contains time with microsecond resolution.
* The values can later be used to measure time intervals, number of clocks,
* FPU frequency, etc.
* @see cpu_rdtsc
*
* @param mark [out] - a pointer to a cpu_mark_t structure
*/
void cpu_tsc_mark(struct cpu_mark_t* mark);
/**
* @brief Calculate TSC and timing difference
*
* @param mark - input/output: a pointer to a cpu_mark_t sturcture, which has
* already been initialized by cpu_tsc_mark. The difference in
* TSC and time will be written here.
*
* This function calculates the TSC and time difference, by obtaining the
* current TSC and timing values and subtracting the contents of the `mark'
* structure from them. Results are written in the same structure.
*
* Example:
* @code
* ...
* struct cpu_mark_t mark;
* cpu_tsc_mark(&mark);
* foo();
* cpu_tsc_unmark(&mark);
* printf("Foo finished. Executed in %llu cycles and %llu usecs\n",
* mark.tsc, mark.sys_clock);
* ...
* @endcode
*/
void cpu_tsc_unmark(struct cpu_mark_t* mark);
/**
* @brief Calculates the CPU clock
*
* @param mark - pointer to a cpu_mark_t structure, which has been initialized
* with cpu_tsc_mark and later `stopped' with cpu_tsc_unmark.
*
* @note For reliable results, the marked time interval should be at least about
* 10 ms.
*
* @returns the CPU clock frequency, in MHz. Due to measurement error, it will
* differ from the true value in a few least-significant bits. Accuracy depends
* on the timing interval - the more, the better. If the timing interval is
* insufficient, the result is -1. Also, see the comment on cpu_clock_measure
* for additional issues and pitfalls in using RDTSC for CPU frequency
* measurements.
*/
int cpu_clock_by_mark(struct cpu_mark_t* mark);
/**
* @brief Returns the CPU clock, as reported by the OS
*
* This function uses OS-specific functions to obtain the CPU clock. It may
* differ from the true clock for several reasons:<br><br>
*
* i) The CPU might be in some power saving state, while the OS reports its
* full-power frequency, or vice-versa.<br>
* ii) In some cases you can raise or lower the CPU frequency with overclocking
* utilities and the OS will not notice.
*
* @returns the CPU clock frequency in MHz. If the OS is not (yet) supported
* or lacks the necessary reporting machinery, the return value is -1
*/
int cpu_clock_by_os(void);
/**
* @brief Measure the CPU clock frequency
*
* @param millis - How much time to waste in the busy-wait cycle. In millisecs.
* Useful values 10 - 1000
* @param quad_check - Do a more thorough measurement if nonzero
* (see the explanation).
*
* The function performs a busy-wait cycle for the given time and calculates
* the CPU frequency by the difference of the TSC values. The accuracy of the
* calculation depends on the length of the busy-wait cycle: more is better,
* but 100ms should be enough for most purposes.
*
* While this will calculate the CPU frequency correctly in most cases, there are
* several reasons why it might be incorrect:<br>
*
* i) RDTSC doesn't guarantee it will run at the same clock as the CPU.
* Apparently there aren't CPUs at the moment, but still, there's no
* guarantee.<br>
* ii) The CPU might be in a low-frequency power saving mode, and the CPU
* might be switched to higher frequency at any time. If this happens
* during the measurement, the result can be anywhere between the
* low and high frequencies. Also, if you're interested in the
* high frequency value only, this function might return the low one
* instead.<br>
* iii) On SMP systems exhibiting TSC drift (see \ref cpu_rdtsc)
*
* the quad_check option will run four consecutive measurements and
* then return the average of the two most-consistent results. The total
* runtime of the function will still be `millis' - consider using
* a bit more time for the timing interval.
*
* Finally, for benchmarking / CPU intensive applications, the best strategy is
* to use the cpu_tsc_mark() / cpu_tsc_unmark() / cpu_clock_by_mark() method.
* Begin by mark()-ing about one second after application startup (allowing the
* power-saving manager to kick in and rise the frequency during that time),
* then unmark() just before application finishing. The result will most
* acurately represent at what frequency your app was running.
*
* @returns the CPU clock frequency in MHz (within some measurement error
* margin). If RDTSC is not supported, the result is -1.
*/
int cpu_clock_measure(int millis, int quad_check);
/**
* @brief Measure the CPU clock frequency using instruction-counting
*
* @param millis - how much time to allocate for each run, in milliseconds
* @param runs - how many runs to perform
*
* The function performs a busy-wait cycle using a known number of "heavy" (SSE)
* instructions. These instructions run at (more or less guaranteed) 1 IPC rate,
* so by running a busy loop for a fixed amount of time, and measuring the
* amount of instructions done, the CPU clock is accurately measured.
*
* Of course, this function is still affected by the power-saving schemes, so
* the warnings as of cpu_clock_measure() still apply. However, this function is
* immune to problems with detection, related to the Intel Nehalem's "Turbo"
* mode, where the internal clock is raised, but the RDTSC rate is unaffected.
*
* The function will run for about (millis * runs) milliseconds.
* You can make only a single busy-wait run (runs == 1); however, this can
* be affected by task scheduling (which will break the counting), so allowing
* more than one run is recommended. As run length is not imperative for
* accurate readings (e.g., 50ms is sufficient), you can afford a lot of short
* runs, e.g. 10 runs of 50ms or 20 runs of 25ms.
*
* Recommended values - millis = 50, runs = 4. For more robustness,
* increase the number of runs.
*
* NOTE: on Bulldozer and later CPUs, the busy-wait cycle runs at 1.4 IPC, thus
* the results are skewed. This is corrected internally by dividing the resulting
* value by 1.4.
* However, this only occurs if the thread is executed on a single CMT
* module - if there are other threads competing for resources, the results are
* unpredictable. Make sure you run cpu_clock_by_ic() on a CPU that is free from
* competing threads, or if there are such threads, they shouldn't exceed the
* number of modules. On a Bulldozer X8, that means 4 threads.
*
* @returns the CPU clock frequency in MHz (within some measurement error
* margin). If SSE is not supported, the result is -1. If the input parameters
* are incorrect, or some other internal fault is detected, the result is -2.
*/
int cpu_clock_by_ic(int millis, int runs);
/**
* @brief Get the CPU clock frequency (all-in-one method)
*
* This is an all-in-one method for getting the CPU clock frequency.
* It tries to use the OS for that. If the OS doesn't have this info, it
* uses cpu_clock_measure with 200ms time interval and quadruple checking.
*
* @returns the CPU clock frequency in MHz. If every possible method fails,
* the result is -1.
*/
int cpu_clock(void);
/**
* @brief Returns the libcpuid version
*
* @returns the string representation of the libcpuid version, like "0.1.1"
*/
const char* cpuid_lib_version(void);
typedef void (*libcpuid_warn_fn_t) (const char *msg);
/**
* @brief Sets the warning print function
*
* In some cases, the internal libcpuid machinery would like to emit useful
* debug warnings. By default, these warnings are written to stderr. However,
* you can set a custom function that will receive those warnings.
*
* @param warn_fun - the warning function you want to set. If NULL, warnings
* are disabled. The function takes const char* argument.
*
* @returns the current warning function. You can use the return value to
* keep the previous warning function and restore it at your discretion.
*/
libcpuid_warn_fn_t cpuid_set_warn_function(libcpuid_warn_fn_t warn_fun);
/**
* @brief Sets the verbosiness level
*
* When the verbosiness level is above zero, some functions might print
* diagnostic information about what are they doing. The higher the level is,
* the more detail is printed. Level zero is guaranteed to omit all such
* output. The output is written using the same machinery as the warnings,
* @see cpuid_set_warn_function()
*
* @param level the desired verbosiness level. Useful values 0..2 inclusive
*/
void cpuid_set_verbosiness_level(int level);
/**
* @brief a structure that holds a list of processor names
*/
struct cpu_list_t {
/** Number of entries in the list */
int num_entries;
/** Pointers to names. There will be num_entries of them */
char **names;
};
/**
* @brief Gets a list of all known CPU names from a specific vendor.
*
* This function compiles a list of all known CPU (code)names
* (i.e. the possible values of cpu_id_t::cpu_codename) for the given vendor.
*
* There are about 100 entries for Intel and AMD, and a few for the other
* vendors. The list is written out in approximate chronological introduction
* order of the parts.
*
* @param vendor the vendor to be queried
* @param list [out] the resulting list will be written here.
* NOTE: As the memory is dynamically allocated, be sure to call
* cpuid_free_cpu_list() after you're done with the data
* @see cpu_list_t
*/
void cpuid_get_cpu_list(cpu_vendor_t vendor, struct cpu_list_t* list);
/**
* @brief Frees a CPU list
*
* This function deletes all the memory associated with a CPU list, as obtained
* by cpuid_get_cpu_list()
*
* @param list - the list to be free()'d.
*/
void cpuid_free_cpu_list(struct cpu_list_t* list);
/**
* @brief Starts/opens a driver, needed to read MSRs (Model Specific Registers)
*
* On systems that support it, this function will create a temporary
* system driver, that has privileges to execute the RDMSR instruction.
* After the driver is created, you can read MSRs by calling \ref cpu_rdmsr
*
* @returns a handle to the driver on success, and NULL on error.
* The error message can be obtained by calling \ref cpuid_error.
* @see cpu_error_t
*/
struct msr_driver_t;
struct msr_driver_t* cpu_msr_driver_open(void);
/**
* @brief Reads a Model-Specific Register (MSR)
*
* If the CPU has MSRs (as indicated by the CPU_FEATURE_MSR flag), you can
* read a MSR with the given index by calling this function.
*
* There are several prerequisites you must do before reading MSRs:
* 1) You must ensure the CPU has RDMSR. Check the CPU_FEATURE_MSR flag
* in cpu_id_t::flags
* 2) You must ensure that the CPU implements the specific MSR you intend to
* read.
* 3) You must open a MSR-reader driver. RDMSR is a privileged instruction and
* needs ring-0 access in order to work. This temporary driver is created
* by calling \ref cpu_msr_driver_open
*
* @param handle - a handle to the MSR reader driver, as created by
* cpu_msr_driver_open
* @param msr_index - the numeric ID of the MSR you want to read
* @param result - a pointer to a 64-bit integer, where the MSR value is stored
*
* @returns zero if successful, and some negative number on error.
* The error message can be obtained by calling \ref cpuid_error.
* @see cpu_error_t
*/
int cpu_rdmsr(struct msr_driver_t* handle, int msr_index, uint64_t* result);
typedef enum {
INFO_MPERF, /*!< Maximum performance frequency clock. This
is a counter, which increments as a
proportion of the actual processor speed */
INFO_APERF, /*!< Actual performance frequency clock. This
accumulates the core clock counts when the
core is active. */
INFO_CUR_MULTIPLIER, /*!< Current CPU:FSB ratio, multiplied by 100.
e.g., a CPU:FSB value of 18.5 reads as
1850. */
INFO_MAX_MULTIPLIER, /*!< Maxumum CPU:FSB ratio for this CPU,
multiplied by 100 */
INFO_TEMPERATURE, /*!< The current core temperature in Celsius */
INFO_THROTTLING, /*!< 1 if the current logical processor is
throttling. 0 if it is running normally. */
} cpu_msrinfo_request_t;
/**
* @brief Reads extended CPU information from Model-Specific Registers.
* @param handle - a handle to an open MSR driver, @see cpu_msr_driver_open
* @param which - which info field should be returned. A list of
* available information entities is listed in the
* cpu_msrinfo_request_t enum.
* @retval - if the requested information is available for the current
* processor model, the respective value is returned.
* if no information is available, or the CPU doesn't support
* the query, the special value CPU_INVALID_VALUE is returned
*/
int cpu_msrinfo(struct msr_driver_t* handle, cpu_msrinfo_request_t which);
#define CPU_INVALID_VALUE 0x3fffffff
/**
* @brief Closes an open MSR driver
*
* This function unloads the MSR driver opened by cpu_msr_driver_open and
* frees any resources associated with it.
*
* @param handle - a handle to the MSR reader driver, as created by
* cpu_msr_driver_open
*
* @returns zero if successful, and some negative number on error.
* The error message can be obtained by calling \ref cpuid_error.
* @see cpu_error_t
*/
int cpu_msr_driver_close(struct msr_driver_t* handle);
#ifdef __cplusplus
}; /* extern "C" */
#endif
/** @} */
#endif /* __LIBCPUID_H__ */