Fix: perf counters build against kernel headers < 3.12
authorMathieu Desnoyers <mathieu.desnoyers@efficios.com>
Thu, 18 Aug 2016 18:21:46 +0000 (14:21 -0400)
committerMathieu Desnoyers <mathieu.desnoyers@efficios.com>
Thu, 18 Aug 2016 23:30:51 +0000 (19:30 -0400)
Copy Linux kernel perf_event.h installed headers into lttng-ust to know
the recent ABI layout, and use the bit description detailed in the
following Linux kernel commit:

https://github.com/torvalds/linux/commit/fa7315871046b9a4c48627905691dbde57e51033

to check whether the kernel supports rdpmc.

Fall-back on the perf read system call for kernels prior to 3.12,
because older kernels have an ABI bug where a union was used for both
cap_usr_time and cap_usr_rdpmc.

Ensure setup_perf set the pc pointer value before checking whether we
need to the file descriptor open or not.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
LICENSE
liblttng-ust/Makefile.am
liblttng-ust/lttng-context-perf-counters.c
liblttng-ust/perf_event.h [new file with mode: 0644]

diff --git a/LICENSE b/LICENSE
index 3c2d3d1a62b463ffb2db8089bdcf2205f25391fe..361eede1465601447347c5584d5de8f4c66fc538 100644 (file)
--- a/LICENSE
+++ b/LICENSE
@@ -16,4 +16,6 @@ mit-license.txt for details.
 * GPLv2
 
 Library test code is distributed under the GPLv2 license, as specified in the
-per-file license. See gpl-2.0.txt for details.
+per-file license. See gpl-2.0.txt for details. A copy of the
+perf_event.h Linux kernel installed header file is also licensed under
+GPLv2.
index a2a32b4efcedfcd68d45491e1c1c5af939c30f05..27acfec16471f4111e96f231ae78e66d1a4d511b 100644 (file)
@@ -60,7 +60,8 @@ liblttng_ust_runtime_la_SOURCES = \
 
 if HAVE_PERF_EVENT
 liblttng_ust_runtime_la_SOURCES += \
-       lttng-context-perf-counters.c
+       lttng-context-perf-counters.c \
+       perf_event.h
 endif
 
 liblttng_ust_support_la_SOURCES = \
index dbb1e6b1dcf368d339053ba1c199f90347007e79..7e87be94cfa5a5f7c19d63fe449b8c3ae70d5a76 100644 (file)
@@ -28,7 +28,6 @@
 #include <stdbool.h>
 #include <sys/mman.h>
 #include <sys/syscall.h>
-#include <linux/perf_event.h>
 #include <lttng/ust-events.h>
 #include <lttng/ust-tracer.h>
 #include <lttng/ringbuffer-config.h>
@@ -39,6 +38,7 @@
 #include <urcu/ref.h>
 #include <usterr-signal-safe.h>
 #include <signal.h>
+#include "perf_event.h"
 #include "lttng-tracer-core.h"
 
 /*
@@ -108,6 +108,15 @@ uint64_t rdpmc(unsigned int counter)
        return low | ((uint64_t) high) << 32;
 }
 
+static
+bool has_rdpmc(struct perf_event_mmap_page *pc)
+{
+       if (caa_unlikely(!pc->cap_bit0_is_deprecated))
+               return false;
+       /* Since Linux kernel 3.12. */
+       return pc->cap_user_rdpmc;
+}
+
 static
 uint64_t arch_read_perf_counter(
                struct lttng_perf_counter_thread_field *thread_field)
@@ -124,7 +133,7 @@ uint64_t arch_read_perf_counter(
                cmm_barrier();
 
                idx = pc->index;
-               if (caa_likely(pc->cap_user_rdpmc && idx)) {
+               if (caa_likely(has_rdpmc(pc) && idx)) {
                        int64_t pmcval;
 
                        pmcval = rdpmc(idx - 1);
@@ -149,7 +158,7 @@ int arch_perf_keep_fd(struct lttng_perf_counter_thread_field *thread_field)
 
        if (!pc)
                return 0;
-       return !pc->cap_user_rdpmc;
+       return !has_rdpmc(pc);
 }
 
 #else
@@ -205,9 +214,7 @@ void close_perf_fd(int fd)
        }
 }
 
-static
-struct perf_event_mmap_page *setup_perf(
-               struct lttng_perf_counter_thread_field *thread_field)
+static void setup_perf(struct lttng_perf_counter_thread_field *thread_field)
 {
        void *perf_addr;
 
@@ -215,13 +222,12 @@ struct perf_event_mmap_page *setup_perf(
                        PROT_READ, MAP_SHARED, thread_field->fd, 0);
        if (perf_addr == MAP_FAILED)
                perf_addr = NULL;
+       thread_field->pc = perf_addr;
 
        if (!arch_perf_keep_fd(thread_field)) {
                close_perf_fd(thread_field->fd);
                thread_field->fd = -1;
        }
-
-       return perf_addr;
 }
 
 static
@@ -296,7 +302,7 @@ struct lttng_perf_counter_thread_field *
        thread_field->field = perf_field;
        thread_field->fd = open_perf_fd(&perf_field->attr);
        if (thread_field->fd >= 0)
-               thread_field->pc = setup_perf(thread_field);
+               setup_perf(thread_field);
        /*
         * Note: thread_field->pc can be NULL if setup_perf() fails.
         * Also, thread_field->fd can be -1 if open_perf_fd() fails.
diff --git a/liblttng-ust/perf_event.h b/liblttng-ust/perf_event.h
new file mode 100644 (file)
index 0000000..356d74b
--- /dev/null
@@ -0,0 +1,992 @@
+/*
+ * Performance events:
+ *
+ *    Copyright (C) 2008-2009, Thomas Gleixner <tglx@linutronix.de>
+ *    Copyright (C) 2008-2011, Red Hat, Inc., Ingo Molnar
+ *    Copyright (C) 2008-2011, Red Hat, Inc., Peter Zijlstra
+ *
+ * Data type definitions, declarations, prototypes.
+ *
+ *    Started by: Thomas Gleixner and Ingo Molnar
+ *
+ * Header copied from Linux kernel v4.7 installed headers.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, you can access it online at
+ * http://www.gnu.org/licenses/gpl-2.0.html.
+ */
+#ifndef _UAPI_LINUX_PERF_EVENT_H
+#define _UAPI_LINUX_PERF_EVENT_H
+
+#include <linux/types.h>
+#include <linux/ioctl.h>
+#include <asm/byteorder.h>
+
+/*
+ * User-space ABI bits:
+ */
+
+/*
+ * attr.type
+ */
+enum perf_type_id {
+       PERF_TYPE_HARDWARE                      = 0,
+       PERF_TYPE_SOFTWARE                      = 1,
+       PERF_TYPE_TRACEPOINT                    = 2,
+       PERF_TYPE_HW_CACHE                      = 3,
+       PERF_TYPE_RAW                           = 4,
+       PERF_TYPE_BREAKPOINT                    = 5,
+
+       PERF_TYPE_MAX,                          /* non-ABI */
+};
+
+/*
+ * Generalized performance event event_id types, used by the
+ * attr.event_id parameter of the sys_perf_event_open()
+ * syscall:
+ */
+enum perf_hw_id {
+       /*
+        * Common hardware events, generalized by the kernel:
+        */
+       PERF_COUNT_HW_CPU_CYCLES                = 0,
+       PERF_COUNT_HW_INSTRUCTIONS              = 1,
+       PERF_COUNT_HW_CACHE_REFERENCES          = 2,
+       PERF_COUNT_HW_CACHE_MISSES              = 3,
+       PERF_COUNT_HW_BRANCH_INSTRUCTIONS       = 4,
+       PERF_COUNT_HW_BRANCH_MISSES             = 5,
+       PERF_COUNT_HW_BUS_CYCLES                = 6,
+       PERF_COUNT_HW_STALLED_CYCLES_FRONTEND   = 7,
+       PERF_COUNT_HW_STALLED_CYCLES_BACKEND    = 8,
+       PERF_COUNT_HW_REF_CPU_CYCLES            = 9,
+
+       PERF_COUNT_HW_MAX,                      /* non-ABI */
+};
+
+/*
+ * Generalized hardware cache events:
+ *
+ *       { L1-D, L1-I, LLC, ITLB, DTLB, BPU, NODE } x
+ *       { read, write, prefetch } x
+ *       { accesses, misses }
+ */
+enum perf_hw_cache_id {
+       PERF_COUNT_HW_CACHE_L1D                 = 0,
+       PERF_COUNT_HW_CACHE_L1I                 = 1,
+       PERF_COUNT_HW_CACHE_LL                  = 2,
+       PERF_COUNT_HW_CACHE_DTLB                = 3,
+       PERF_COUNT_HW_CACHE_ITLB                = 4,
+       PERF_COUNT_HW_CACHE_BPU                 = 5,
+       PERF_COUNT_HW_CACHE_NODE                = 6,
+
+       PERF_COUNT_HW_CACHE_MAX,                /* non-ABI */
+};
+
+enum perf_hw_cache_op_id {
+       PERF_COUNT_HW_CACHE_OP_READ             = 0,
+       PERF_COUNT_HW_CACHE_OP_WRITE            = 1,
+       PERF_COUNT_HW_CACHE_OP_PREFETCH         = 2,
+
+       PERF_COUNT_HW_CACHE_OP_MAX,             /* non-ABI */
+};
+
+enum perf_hw_cache_op_result_id {
+       PERF_COUNT_HW_CACHE_RESULT_ACCESS       = 0,
+       PERF_COUNT_HW_CACHE_RESULT_MISS         = 1,
+
+       PERF_COUNT_HW_CACHE_RESULT_MAX,         /* non-ABI */
+};
+
+/*
+ * Special "software" events provided by the kernel, even if the hardware
+ * does not support performance events. These events measure various
+ * physical and sw events of the kernel (and allow the profiling of them as
+ * well):
+ */
+enum perf_sw_ids {
+       PERF_COUNT_SW_CPU_CLOCK                 = 0,
+       PERF_COUNT_SW_TASK_CLOCK                = 1,
+       PERF_COUNT_SW_PAGE_FAULTS               = 2,
+       PERF_COUNT_SW_CONTEXT_SWITCHES          = 3,
+       PERF_COUNT_SW_CPU_MIGRATIONS            = 4,
+       PERF_COUNT_SW_PAGE_FAULTS_MIN           = 5,
+       PERF_COUNT_SW_PAGE_FAULTS_MAJ           = 6,
+       PERF_COUNT_SW_ALIGNMENT_FAULTS          = 7,
+       PERF_COUNT_SW_EMULATION_FAULTS          = 8,
+       PERF_COUNT_SW_DUMMY                     = 9,
+       PERF_COUNT_SW_BPF_OUTPUT                = 10,
+
+       PERF_COUNT_SW_MAX,                      /* non-ABI */
+};
+
+/*
+ * Bits that can be set in attr.sample_type to request information
+ * in the overflow packets.
+ */
+enum perf_event_sample_format {
+       PERF_SAMPLE_IP                          = 1U << 0,
+       PERF_SAMPLE_TID                         = 1U << 1,
+       PERF_SAMPLE_TIME                        = 1U << 2,
+       PERF_SAMPLE_ADDR                        = 1U << 3,
+       PERF_SAMPLE_READ                        = 1U << 4,
+       PERF_SAMPLE_CALLCHAIN                   = 1U << 5,
+       PERF_SAMPLE_ID                          = 1U << 6,
+       PERF_SAMPLE_CPU                         = 1U << 7,
+       PERF_SAMPLE_PERIOD                      = 1U << 8,
+       PERF_SAMPLE_STREAM_ID                   = 1U << 9,
+       PERF_SAMPLE_RAW                         = 1U << 10,
+       PERF_SAMPLE_BRANCH_STACK                = 1U << 11,
+       PERF_SAMPLE_REGS_USER                   = 1U << 12,
+       PERF_SAMPLE_STACK_USER                  = 1U << 13,
+       PERF_SAMPLE_WEIGHT                      = 1U << 14,
+       PERF_SAMPLE_DATA_SRC                    = 1U << 15,
+       PERF_SAMPLE_IDENTIFIER                  = 1U << 16,
+       PERF_SAMPLE_TRANSACTION                 = 1U << 17,
+       PERF_SAMPLE_REGS_INTR                   = 1U << 18,
+
+       PERF_SAMPLE_MAX = 1U << 19,             /* non-ABI */
+};
+
+/*
+ * values to program into branch_sample_type when PERF_SAMPLE_BRANCH is set
+ *
+ * If the user does not pass priv level information via branch_sample_type,
+ * the kernel uses the event's priv level. Branch and event priv levels do
+ * not have to match. Branch priv level is checked for permissions.
+ *
+ * The branch types can be combined, however BRANCH_ANY covers all types
+ * of branches and therefore it supersedes all the other types.
+ */
+enum perf_branch_sample_type_shift {
+       PERF_SAMPLE_BRANCH_USER_SHIFT           = 0, /* user branches */
+       PERF_SAMPLE_BRANCH_KERNEL_SHIFT         = 1, /* kernel branches */
+       PERF_SAMPLE_BRANCH_HV_SHIFT             = 2, /* hypervisor branches */
+
+       PERF_SAMPLE_BRANCH_ANY_SHIFT            = 3, /* any branch types */
+       PERF_SAMPLE_BRANCH_ANY_CALL_SHIFT       = 4, /* any call branch */
+       PERF_SAMPLE_BRANCH_ANY_RETURN_SHIFT     = 5, /* any return branch */
+       PERF_SAMPLE_BRANCH_IND_CALL_SHIFT       = 6, /* indirect calls */
+       PERF_SAMPLE_BRANCH_ABORT_TX_SHIFT       = 7, /* transaction aborts */
+       PERF_SAMPLE_BRANCH_IN_TX_SHIFT          = 8, /* in transaction */
+       PERF_SAMPLE_BRANCH_NO_TX_SHIFT          = 9, /* not in transaction */
+       PERF_SAMPLE_BRANCH_COND_SHIFT           = 10, /* conditional branches */
+
+       PERF_SAMPLE_BRANCH_CALL_STACK_SHIFT     = 11, /* call/ret stack */
+       PERF_SAMPLE_BRANCH_IND_JUMP_SHIFT       = 12, /* indirect jumps */
+       PERF_SAMPLE_BRANCH_CALL_SHIFT           = 13, /* direct call */
+
+       PERF_SAMPLE_BRANCH_NO_FLAGS_SHIFT       = 14, /* no flags */
+       PERF_SAMPLE_BRANCH_NO_CYCLES_SHIFT      = 15, /* no cycles */
+
+       PERF_SAMPLE_BRANCH_MAX_SHIFT            /* non-ABI */
+};
+
+enum perf_branch_sample_type {
+       PERF_SAMPLE_BRANCH_USER         = 1U << PERF_SAMPLE_BRANCH_USER_SHIFT,
+       PERF_SAMPLE_BRANCH_KERNEL       = 1U << PERF_SAMPLE_BRANCH_KERNEL_SHIFT,
+       PERF_SAMPLE_BRANCH_HV           = 1U << PERF_SAMPLE_BRANCH_HV_SHIFT,
+
+       PERF_SAMPLE_BRANCH_ANY          = 1U << PERF_SAMPLE_BRANCH_ANY_SHIFT,
+       PERF_SAMPLE_BRANCH_ANY_CALL     = 1U << PERF_SAMPLE_BRANCH_ANY_CALL_SHIFT,
+       PERF_SAMPLE_BRANCH_ANY_RETURN   = 1U << PERF_SAMPLE_BRANCH_ANY_RETURN_SHIFT,
+       PERF_SAMPLE_BRANCH_IND_CALL     = 1U << PERF_SAMPLE_BRANCH_IND_CALL_SHIFT,
+       PERF_SAMPLE_BRANCH_ABORT_TX     = 1U << PERF_SAMPLE_BRANCH_ABORT_TX_SHIFT,
+       PERF_SAMPLE_BRANCH_IN_TX        = 1U << PERF_SAMPLE_BRANCH_IN_TX_SHIFT,
+       PERF_SAMPLE_BRANCH_NO_TX        = 1U << PERF_SAMPLE_BRANCH_NO_TX_SHIFT,
+       PERF_SAMPLE_BRANCH_COND         = 1U << PERF_SAMPLE_BRANCH_COND_SHIFT,
+
+       PERF_SAMPLE_BRANCH_CALL_STACK   = 1U << PERF_SAMPLE_BRANCH_CALL_STACK_SHIFT,
+       PERF_SAMPLE_BRANCH_IND_JUMP     = 1U << PERF_SAMPLE_BRANCH_IND_JUMP_SHIFT,
+       PERF_SAMPLE_BRANCH_CALL         = 1U << PERF_SAMPLE_BRANCH_CALL_SHIFT,
+
+       PERF_SAMPLE_BRANCH_NO_FLAGS     = 1U << PERF_SAMPLE_BRANCH_NO_FLAGS_SHIFT,
+       PERF_SAMPLE_BRANCH_NO_CYCLES    = 1U << PERF_SAMPLE_BRANCH_NO_CYCLES_SHIFT,
+
+       PERF_SAMPLE_BRANCH_MAX          = 1U << PERF_SAMPLE_BRANCH_MAX_SHIFT,
+};
+
+#define PERF_SAMPLE_BRANCH_PLM_ALL \
+       (PERF_SAMPLE_BRANCH_USER|\
+        PERF_SAMPLE_BRANCH_KERNEL|\
+        PERF_SAMPLE_BRANCH_HV)
+
+/*
+ * Values to determine ABI of the registers dump.
+ */
+enum perf_sample_regs_abi {
+       PERF_SAMPLE_REGS_ABI_NONE       = 0,
+       PERF_SAMPLE_REGS_ABI_32         = 1,
+       PERF_SAMPLE_REGS_ABI_64         = 2,
+};
+
+/*
+ * Values for the memory transaction event qualifier, mostly for
+ * abort events. Multiple bits can be set.
+ */
+enum {
+       PERF_TXN_ELISION        = (1 << 0), /* From elision */
+       PERF_TXN_TRANSACTION    = (1 << 1), /* From transaction */
+       PERF_TXN_SYNC           = (1 << 2), /* Instruction is related */
+       PERF_TXN_ASYNC          = (1 << 3), /* Instruction not related */
+       PERF_TXN_RETRY          = (1 << 4), /* Retry possible */
+       PERF_TXN_CONFLICT       = (1 << 5), /* Conflict abort */
+       PERF_TXN_CAPACITY_WRITE = (1 << 6), /* Capacity write abort */
+       PERF_TXN_CAPACITY_READ  = (1 << 7), /* Capacity read abort */
+
+       PERF_TXN_MAX            = (1 << 8), /* non-ABI */
+
+       /* bits 32..63 are reserved for the abort code */
+
+       PERF_TXN_ABORT_MASK  = (0xffffffffULL << 32),
+       PERF_TXN_ABORT_SHIFT = 32,
+};
+
+/*
+ * The format of the data returned by read() on a perf event fd,
+ * as specified by attr.read_format:
+ *
+ * struct read_format {
+ *     { u64           value;
+ *       { u64         time_enabled; } && PERF_FORMAT_TOTAL_TIME_ENABLED
+ *       { u64         time_running; } && PERF_FORMAT_TOTAL_TIME_RUNNING
+ *       { u64         id;           } && PERF_FORMAT_ID
+ *     } && !PERF_FORMAT_GROUP
+ *
+ *     { u64           nr;
+ *       { u64         time_enabled; } && PERF_FORMAT_TOTAL_TIME_ENABLED
+ *       { u64         time_running; } && PERF_FORMAT_TOTAL_TIME_RUNNING
+ *       { u64         value;
+ *         { u64       id;           } && PERF_FORMAT_ID
+ *       }             cntr[nr];
+ *     } && PERF_FORMAT_GROUP
+ * };
+ */
+enum perf_event_read_format {
+       PERF_FORMAT_TOTAL_TIME_ENABLED          = 1U << 0,
+       PERF_FORMAT_TOTAL_TIME_RUNNING          = 1U << 1,
+       PERF_FORMAT_ID                          = 1U << 2,
+       PERF_FORMAT_GROUP                       = 1U << 3,
+
+       PERF_FORMAT_MAX = 1U << 4,              /* non-ABI */
+};
+
+#define PERF_ATTR_SIZE_VER0    64      /* sizeof first published struct */
+#define PERF_ATTR_SIZE_VER1    72      /* add: config2 */
+#define PERF_ATTR_SIZE_VER2    80      /* add: branch_sample_type */
+#define PERF_ATTR_SIZE_VER3    96      /* add: sample_regs_user */
+                                       /* add: sample_stack_user */
+#define PERF_ATTR_SIZE_VER4    104     /* add: sample_regs_intr */
+#define PERF_ATTR_SIZE_VER5    112     /* add: aux_watermark */
+
+/*
+ * Hardware event_id to monitor via a performance monitoring event:
+ */
+struct perf_event_attr {
+
+       /*
+        * Major type: hardware/software/tracepoint/etc.
+        */
+       __u32                   type;
+
+       /*
+        * Size of the attr structure, for fwd/bwd compat.
+        */
+       __u32                   size;
+
+       /*
+        * Type specific configuration information.
+        */
+       __u64                   config;
+
+       union {
+               __u64           sample_period;
+               __u64           sample_freq;
+       };
+
+       __u64                   sample_type;
+       __u64                   read_format;
+
+       __u64                   disabled       :  1, /* off by default        */
+                               inherit        :  1, /* children inherit it   */
+                               pinned         :  1, /* must always be on PMU */
+                               exclusive      :  1, /* only group on PMU     */
+                               exclude_user   :  1, /* don't count user      */
+                               exclude_kernel :  1, /* ditto kernel          */
+                               exclude_hv     :  1, /* ditto hypervisor      */
+                               exclude_idle   :  1, /* don't count when idle */
+                               mmap           :  1, /* include mmap data     */
+                               comm           :  1, /* include comm data     */
+                               freq           :  1, /* use freq, not period  */
+                               inherit_stat   :  1, /* per task counts       */
+                               enable_on_exec :  1, /* next exec enables     */
+                               task           :  1, /* trace fork/exit       */
+                               watermark      :  1, /* wakeup_watermark      */
+                               /*
+                                * precise_ip:
+                                *
+                                *  0 - SAMPLE_IP can have arbitrary skid
+                                *  1 - SAMPLE_IP must have constant skid
+                                *  2 - SAMPLE_IP requested to have 0 skid
+                                *  3 - SAMPLE_IP must have 0 skid
+                                *
+                                *  See also PERF_RECORD_MISC_EXACT_IP
+                                */
+                               precise_ip     :  2, /* skid constraint       */
+                               mmap_data      :  1, /* non-exec mmap data    */
+                               sample_id_all  :  1, /* sample_type all events */
+
+                               exclude_host   :  1, /* don't count in host   */
+                               exclude_guest  :  1, /* don't count in guest  */
+
+                               exclude_callchain_kernel : 1, /* exclude kernel callchains */
+                               exclude_callchain_user   : 1, /* exclude user callchains */
+                               mmap2          :  1, /* include mmap with inode data     */
+                               comm_exec      :  1, /* flag comm events that are due to an exec */
+                               use_clockid    :  1, /* use @clockid for time fields */
+                               context_switch :  1, /* context switch data */
+                               write_backward :  1, /* Write ring buffer from end to beginning */
+                               __reserved_1   : 36;
+
+       union {
+               __u32           wakeup_events;    /* wakeup every n events */
+               __u32           wakeup_watermark; /* bytes before wakeup   */
+       };
+
+       __u32                   bp_type;
+       union {
+               __u64           bp_addr;
+               __u64           config1; /* extension of config */
+       };
+       union {
+               __u64           bp_len;
+               __u64           config2; /* extension of config1 */
+       };
+       __u64   branch_sample_type; /* enum perf_branch_sample_type */
+
+       /*
+        * Defines set of user regs to dump on samples.
+        * See asm/perf_regs.h for details.
+        */
+       __u64   sample_regs_user;
+
+       /*
+        * Defines size of the user stack to dump on samples.
+        */
+       __u32   sample_stack_user;
+
+       __s32   clockid;
+       /*
+        * Defines set of regs to dump for each sample
+        * state captured on:
+        *  - precise = 0: PMU interrupt
+        *  - precise > 0: sampled instruction
+        *
+        * See asm/perf_regs.h for details.
+        */
+       __u64   sample_regs_intr;
+
+       /*
+        * Wakeup watermark for AUX area
+        */
+       __u32   aux_watermark;
+       __u32   __reserved_2;   /* align to __u64 */
+};
+
+#define perf_flags(attr)       (*(&(attr)->read_format + 1))
+
+/*
+ * Ioctls that can be done on a perf event fd:
+ */
+#define PERF_EVENT_IOC_ENABLE          _IO ('$', 0)
+#define PERF_EVENT_IOC_DISABLE         _IO ('$', 1)
+#define PERF_EVENT_IOC_REFRESH         _IO ('$', 2)
+#define PERF_EVENT_IOC_RESET           _IO ('$', 3)
+#define PERF_EVENT_IOC_PERIOD          _IOW('$', 4, __u64)
+#define PERF_EVENT_IOC_SET_OUTPUT      _IO ('$', 5)
+#define PERF_EVENT_IOC_SET_FILTER      _IOW('$', 6, char *)
+#define PERF_EVENT_IOC_ID              _IOR('$', 7, __u64 *)
+#define PERF_EVENT_IOC_SET_BPF         _IOW('$', 8, __u32)
+#define PERF_EVENT_IOC_PAUSE_OUTPUT    _IOW('$', 9, __u32)
+
+enum perf_event_ioc_flags {
+       PERF_IOC_FLAG_GROUP             = 1U << 0,
+};
+
+/*
+ * Structure of the page that can be mapped via mmap
+ */
+struct perf_event_mmap_page {
+       __u32   version;                /* version number of this structure */
+       __u32   compat_version;         /* lowest version this is compat with */
+
+       /*
+        * Bits needed to read the hw events in user-space.
+        *
+        *   u32 seq, time_mult, time_shift, index, width;
+        *   u64 count, enabled, running;
+        *   u64 cyc, time_offset;
+        *   s64 pmc = 0;
+        *
+        *   do {
+        *     seq = pc->lock;
+        *     barrier()
+        *
+        *     enabled = pc->time_enabled;
+        *     running = pc->time_running;
+        *
+        *     if (pc->cap_usr_time && enabled != running) {
+        *       cyc = rdtsc();
+        *       time_offset = pc->time_offset;
+        *       time_mult   = pc->time_mult;
+        *       time_shift  = pc->time_shift;
+        *     }
+        *
+        *     index = pc->index;
+        *     count = pc->offset;
+        *     if (pc->cap_user_rdpmc && index) {
+        *       width = pc->pmc_width;
+        *       pmc = rdpmc(index - 1);
+        *     }
+        *
+        *     barrier();
+        *   } while (pc->lock != seq);
+        *
+        * NOTE: for obvious reason this only works on self-monitoring
+        *       processes.
+        */
+       __u32   lock;                   /* seqlock for synchronization */
+       __u32   index;                  /* hardware event identifier */
+       __s64   offset;                 /* add to hardware event value */
+       __u64   time_enabled;           /* time event active */
+       __u64   time_running;           /* time event on cpu */
+       union {
+               __u64   capabilities;
+               struct {
+                       __u64   cap_bit0                : 1, /* Always 0, deprecated, see commit 860f085b74e9 */
+                               cap_bit0_is_deprecated  : 1, /* Always 1, signals that bit 0 is zero */
+
+                               cap_user_rdpmc          : 1, /* The RDPMC instruction can be used to read counts */
+                               cap_user_time           : 1, /* The time_* fields are used */
+                               cap_user_time_zero      : 1, /* The time_zero field is used */
+                               cap_____res             : 59;
+               };
+       };
+
+       /*
+        * If cap_user_rdpmc this field provides the bit-width of the value
+        * read using the rdpmc() or equivalent instruction. This can be used
+        * to sign extend the result like:
+        *
+        *   pmc <<= 64 - width;
+        *   pmc >>= 64 - width; // signed shift right
+        *   count += pmc;
+        */
+       __u16   pmc_width;
+
+       /*
+        * If cap_usr_time the below fields can be used to compute the time
+        * delta since time_enabled (in ns) using rdtsc or similar.
+        *
+        *   u64 quot, rem;
+        *   u64 delta;
+        *
+        *   quot = (cyc >> time_shift);
+        *   rem = cyc & (((u64)1 << time_shift) - 1);
+        *   delta = time_offset + quot * time_mult +
+        *              ((rem * time_mult) >> time_shift);
+        *
+        * Where time_offset,time_mult,time_shift and cyc are read in the
+        * seqcount loop described above. This delta can then be added to
+        * enabled and possible running (if index), improving the scaling:
+        *
+        *   enabled += delta;
+        *   if (index)
+        *     running += delta;
+        *
+        *   quot = count / running;
+        *   rem  = count % running;
+        *   count = quot * enabled + (rem * enabled) / running;
+        */
+       __u16   time_shift;
+       __u32   time_mult;
+       __u64   time_offset;
+       /*
+        * If cap_usr_time_zero, the hardware clock (e.g. TSC) can be calculated
+        * from sample timestamps.
+        *
+        *   time = timestamp - time_zero;
+        *   quot = time / time_mult;
+        *   rem  = time % time_mult;
+        *   cyc = (quot << time_shift) + (rem << time_shift) / time_mult;
+        *
+        * And vice versa:
+        *
+        *   quot = cyc >> time_shift;
+        *   rem  = cyc & (((u64)1 << time_shift) - 1);
+        *   timestamp = time_zero + quot * time_mult +
+        *               ((rem * time_mult) >> time_shift);
+        */
+       __u64   time_zero;
+       __u32   size;                   /* Header size up to __reserved[] fields. */
+
+               /*
+                * Hole for extension of the self monitor capabilities
+                */
+
+       __u8    __reserved[118*8+4];    /* align to 1k. */
+
+       /*
+        * Control data for the mmap() data buffer.
+        *
+        * User-space reading the @data_head value should issue an smp_rmb(),
+        * after reading this value.
+        *
+        * When the mapping is PROT_WRITE the @data_tail value should be
+        * written by userspace to reflect the last read data, after issueing
+        * an smp_mb() to separate the data read from the ->data_tail store.
+        * In this case the kernel will not over-write unread data.
+        *
+        * See perf_output_put_handle() for the data ordering.
+        *
+        * data_{offset,size} indicate the location and size of the perf record
+        * buffer within the mmapped area.
+        */
+       __u64   data_head;              /* head in the data section */
+       __u64   data_tail;              /* user-space written tail */
+       __u64   data_offset;            /* where the buffer starts */
+       __u64   data_size;              /* data buffer size */
+
+       /*
+        * AUX area is defined by aux_{offset,size} fields that should be set
+        * by the userspace, so that
+        *
+        *   aux_offset >= data_offset + data_size
+        *
+        * prior to mmap()ing it. Size of the mmap()ed area should be aux_size.
+        *
+        * Ring buffer pointers aux_{head,tail} have the same semantics as
+        * data_{head,tail} and same ordering rules apply.
+        */
+       __u64   aux_head;
+       __u64   aux_tail;
+       __u64   aux_offset;
+       __u64   aux_size;
+};
+
+#define PERF_RECORD_MISC_CPUMODE_MASK          (7 << 0)
+#define PERF_RECORD_MISC_CPUMODE_UNKNOWN       (0 << 0)
+#define PERF_RECORD_MISC_KERNEL                        (1 << 0)
+#define PERF_RECORD_MISC_USER                  (2 << 0)
+#define PERF_RECORD_MISC_HYPERVISOR            (3 << 0)
+#define PERF_RECORD_MISC_GUEST_KERNEL          (4 << 0)
+#define PERF_RECORD_MISC_GUEST_USER            (5 << 0)
+
+/*
+ * Indicates that /proc/PID/maps parsing are truncated by time out.
+ */
+#define PERF_RECORD_MISC_PROC_MAP_PARSE_TIMEOUT        (1 << 12)
+/*
+ * PERF_RECORD_MISC_MMAP_DATA and PERF_RECORD_MISC_COMM_EXEC are used on
+ * different events so can reuse the same bit position.
+ * Ditto PERF_RECORD_MISC_SWITCH_OUT.
+ */
+#define PERF_RECORD_MISC_MMAP_DATA             (1 << 13)
+#define PERF_RECORD_MISC_COMM_EXEC             (1 << 13)
+#define PERF_RECORD_MISC_SWITCH_OUT            (1 << 13)
+/*
+ * Indicates that the content of PERF_SAMPLE_IP points to
+ * the actual instruction that triggered the event. See also
+ * perf_event_attr::precise_ip.
+ */
+#define PERF_RECORD_MISC_EXACT_IP              (1 << 14)
+/*
+ * Reserve the last bit to indicate some extended misc field
+ */
+#define PERF_RECORD_MISC_EXT_RESERVED          (1 << 15)
+
+struct perf_event_header {
+       __u32   type;
+       __u16   misc;
+       __u16   size;
+};
+
+enum perf_event_type {
+
+       /*
+        * If perf_event_attr.sample_id_all is set then all event types will
+        * have the sample_type selected fields related to where/when
+        * (identity) an event took place (TID, TIME, ID, STREAM_ID, CPU,
+        * IDENTIFIER) described in PERF_RECORD_SAMPLE below, it will be stashed
+        * just after the perf_event_header and the fields already present for
+        * the existing fields, i.e. at the end of the payload. That way a newer
+        * perf.data file will be supported by older perf tools, with these new
+        * optional fields being ignored.
+        *
+        * struct sample_id {
+        *      { u32                   pid, tid; } && PERF_SAMPLE_TID
+        *      { u64                   time;     } && PERF_SAMPLE_TIME
+        *      { u64                   id;       } && PERF_SAMPLE_ID
+        *      { u64                   stream_id;} && PERF_SAMPLE_STREAM_ID
+        *      { u32                   cpu, res; } && PERF_SAMPLE_CPU
+        *      { u64                   id;       } && PERF_SAMPLE_IDENTIFIER
+        * } && perf_event_attr::sample_id_all
+        *
+        * Note that PERF_SAMPLE_IDENTIFIER duplicates PERF_SAMPLE_ID.  The
+        * advantage of PERF_SAMPLE_IDENTIFIER is that its position is fixed
+        * relative to header.size.
+        */
+
+       /*
+        * The MMAP events record the PROT_EXEC mappings so that we can
+        * correlate userspace IPs to code. They have the following structure:
+        *
+        * struct {
+        *      struct perf_event_header        header;
+        *
+        *      u32                             pid, tid;
+        *      u64                             addr;
+        *      u64                             len;
+        *      u64                             pgoff;
+        *      char                            filename[];
+        *      struct sample_id                sample_id;
+        * };
+        */
+       PERF_RECORD_MMAP                        = 1,
+
+       /*
+        * struct {
+        *      struct perf_event_header        header;
+        *      u64                             id;
+        *      u64                             lost;
+        *      struct sample_id                sample_id;
+        * };
+        */
+       PERF_RECORD_LOST                        = 2,
+
+       /*
+        * struct {
+        *      struct perf_event_header        header;
+        *
+        *      u32                             pid, tid;
+        *      char                            comm[];
+        *      struct sample_id                sample_id;
+        * };
+        */
+       PERF_RECORD_COMM                        = 3,
+
+       /*
+        * struct {
+        *      struct perf_event_header        header;
+        *      u32                             pid, ppid;
+        *      u32                             tid, ptid;
+        *      u64                             time;
+        *      struct sample_id                sample_id;
+        * };
+        */
+       PERF_RECORD_EXIT                        = 4,
+
+       /*
+        * struct {
+        *      struct perf_event_header        header;
+        *      u64                             time;
+        *      u64                             id;
+        *      u64                             stream_id;
+        *      struct sample_id                sample_id;
+        * };
+        */
+       PERF_RECORD_THROTTLE                    = 5,
+       PERF_RECORD_UNTHROTTLE                  = 6,
+
+       /*
+        * struct {
+        *      struct perf_event_header        header;
+        *      u32                             pid, ppid;
+        *      u32                             tid, ptid;
+        *      u64                             time;
+        *      struct sample_id                sample_id;
+        * };
+        */
+       PERF_RECORD_FORK                        = 7,
+
+       /*
+        * struct {
+        *      struct perf_event_header        header;
+        *      u32                             pid, tid;
+        *
+        *      struct read_format              values;
+        *      struct sample_id                sample_id;
+        * };
+        */
+       PERF_RECORD_READ                        = 8,
+
+       /*
+        * struct {
+        *      struct perf_event_header        header;
+        *
+        *      #
+        *      # Note that PERF_SAMPLE_IDENTIFIER duplicates PERF_SAMPLE_ID.
+        *      # The advantage of PERF_SAMPLE_IDENTIFIER is that its position
+        *      # is fixed relative to header.
+        *      #
+        *
+        *      { u64                   id;       } && PERF_SAMPLE_IDENTIFIER
+        *      { u64                   ip;       } && PERF_SAMPLE_IP
+        *      { u32                   pid, tid; } && PERF_SAMPLE_TID
+        *      { u64                   time;     } && PERF_SAMPLE_TIME
+        *      { u64                   addr;     } && PERF_SAMPLE_ADDR
+        *      { u64                   id;       } && PERF_SAMPLE_ID
+        *      { u64                   stream_id;} && PERF_SAMPLE_STREAM_ID
+        *      { u32                   cpu, res; } && PERF_SAMPLE_CPU
+        *      { u64                   period;   } && PERF_SAMPLE_PERIOD
+        *
+        *      { struct read_format    values;   } && PERF_SAMPLE_READ
+        *
+        *      { u64                   nr,
+        *        u64                   ips[nr];  } && PERF_SAMPLE_CALLCHAIN
+        *
+        *      #
+        *      # The RAW record below is opaque data wrt the ABI
+        *      #
+        *      # That is, the ABI doesn't make any promises wrt to
+        *      # the stability of its content, it may vary depending
+        *      # on event, hardware, kernel version and phase of
+        *      # the moon.
+        *      #
+        *      # In other words, PERF_SAMPLE_RAW contents are not an ABI.
+        *      #
+        *
+        *      { u32                   size;
+        *        char                  data[size];}&& PERF_SAMPLE_RAW
+        *
+        *      { u64                   nr;
+        *        { u64 from, to, flags } lbr[nr];} && PERF_SAMPLE_BRANCH_STACK
+        *
+        *      { u64                   abi; # enum perf_sample_regs_abi
+        *        u64                   regs[weight(mask)]; } && PERF_SAMPLE_REGS_USER
+        *
+        *      { u64                   size;
+        *        char                  data[size];
+        *        u64                   dyn_size; } && PERF_SAMPLE_STACK_USER
+        *
+        *      { u64                   weight;   } && PERF_SAMPLE_WEIGHT
+        *      { u64                   data_src; } && PERF_SAMPLE_DATA_SRC
+        *      { u64                   transaction; } && PERF_SAMPLE_TRANSACTION
+        *      { u64                   abi; # enum perf_sample_regs_abi
+        *        u64                   regs[weight(mask)]; } && PERF_SAMPLE_REGS_INTR
+        * };
+        */
+       PERF_RECORD_SAMPLE                      = 9,
+
+       /*
+        * The MMAP2 records are an augmented version of MMAP, they add
+        * maj, min, ino numbers to be used to uniquely identify each mapping
+        *
+        * struct {
+        *      struct perf_event_header        header;
+        *
+        *      u32                             pid, tid;
+        *      u64                             addr;
+        *      u64                             len;
+        *      u64                             pgoff;
+        *      u32                             maj;
+        *      u32                             min;
+        *      u64                             ino;
+        *      u64                             ino_generation;
+        *      u32                             prot, flags;
+        *      char                            filename[];
+        *      struct sample_id                sample_id;
+        * };
+        */
+       PERF_RECORD_MMAP2                       = 10,
+
+       /*
+        * Records that new data landed in the AUX buffer part.
+        *
+        * struct {
+        *      struct perf_event_header        header;
+        *
+        *      u64                             aux_offset;
+        *      u64                             aux_size;
+        *      u64                             flags;
+        *      struct sample_id                sample_id;
+        * };
+        */
+       PERF_RECORD_AUX                         = 11,
+
+       /*
+        * Indicates that instruction trace has started
+        *
+        * struct {
+        *      struct perf_event_header        header;
+        *      u32                             pid;
+        *      u32                             tid;
+        * };
+        */
+       PERF_RECORD_ITRACE_START                = 12,
+
+       /*
+        * Records the dropped/lost sample number.
+        *
+        * struct {
+        *      struct perf_event_header        header;
+        *
+        *      u64                             lost;
+        *      struct sample_id                sample_id;
+        * };
+        */
+       PERF_RECORD_LOST_SAMPLES                = 13,
+
+       /*
+        * Records a context switch in or out (flagged by
+        * PERF_RECORD_MISC_SWITCH_OUT). See also
+        * PERF_RECORD_SWITCH_CPU_WIDE.
+        *
+        * struct {
+        *      struct perf_event_header        header;
+        *      struct sample_id                sample_id;
+        * };
+        */
+       PERF_RECORD_SWITCH                      = 14,
+
+       /*
+        * CPU-wide version of PERF_RECORD_SWITCH with next_prev_pid and
+        * next_prev_tid that are the next (switching out) or previous
+        * (switching in) pid/tid.
+        *
+        * struct {
+        *      struct perf_event_header        header;
+        *      u32                             next_prev_pid;
+        *      u32                             next_prev_tid;
+        *      struct sample_id                sample_id;
+        * };
+        */
+       PERF_RECORD_SWITCH_CPU_WIDE             = 15,
+
+       PERF_RECORD_MAX,                        /* non-ABI */
+};
+
+#define PERF_MAX_STACK_DEPTH           127
+#define PERF_MAX_CONTEXTS_PER_STACK      8
+
+enum perf_callchain_context {
+       PERF_CONTEXT_HV                 = (__u64)-32,
+       PERF_CONTEXT_KERNEL             = (__u64)-128,
+       PERF_CONTEXT_USER               = (__u64)-512,
+
+       PERF_CONTEXT_GUEST              = (__u64)-2048,
+       PERF_CONTEXT_GUEST_KERNEL       = (__u64)-2176,
+       PERF_CONTEXT_GUEST_USER         = (__u64)-2560,
+
+       PERF_CONTEXT_MAX                = (__u64)-4095,
+};
+
+/**
+ * PERF_RECORD_AUX::flags bits
+ */
+#define PERF_AUX_FLAG_TRUNCATED                0x01    /* record was truncated to fit */
+#define PERF_AUX_FLAG_OVERWRITE                0x02    /* snapshot from overwrite mode */
+
+#define PERF_FLAG_FD_NO_GROUP          (1UL << 0)
+#define PERF_FLAG_FD_OUTPUT            (1UL << 1)
+#define PERF_FLAG_PID_CGROUP           (1UL << 2) /* pid=cgroup id, per-cpu mode only */
+#define PERF_FLAG_FD_CLOEXEC           (1UL << 3) /* O_CLOEXEC */
+
+union perf_mem_data_src {
+       __u64 val;
+       struct {
+               __u64   mem_op:5,       /* type of opcode */
+                       mem_lvl:14,     /* memory hierarchy level */
+                       mem_snoop:5,    /* snoop mode */
+                       mem_lock:2,     /* lock instr */
+                       mem_dtlb:7,     /* tlb access */
+                       mem_rsvd:31;
+       };
+};
+
+/* type of opcode (load/store/prefetch,code) */
+#define PERF_MEM_OP_NA         0x01 /* not available */
+#define PERF_MEM_OP_LOAD       0x02 /* load instruction */
+#define PERF_MEM_OP_STORE      0x04 /* store instruction */
+#define PERF_MEM_OP_PFETCH     0x08 /* prefetch */
+#define PERF_MEM_OP_EXEC       0x10 /* code (execution) */
+#define PERF_MEM_OP_SHIFT      0
+
+/* memory hierarchy (memory level, hit or miss) */
+#define PERF_MEM_LVL_NA                0x01  /* not available */
+#define PERF_MEM_LVL_HIT       0x02  /* hit level */
+#define PERF_MEM_LVL_MISS      0x04  /* miss level  */
+#define PERF_MEM_LVL_L1                0x08  /* L1 */
+#define PERF_MEM_LVL_LFB       0x10  /* Line Fill Buffer */
+#define PERF_MEM_LVL_L2                0x20  /* L2 */
+#define PERF_MEM_LVL_L3                0x40  /* L3 */
+#define PERF_MEM_LVL_LOC_RAM   0x80  /* Local DRAM */
+#define PERF_MEM_LVL_REM_RAM1  0x100 /* Remote DRAM (1 hop) */
+#define PERF_MEM_LVL_REM_RAM2  0x200 /* Remote DRAM (2 hops) */
+#define PERF_MEM_LVL_REM_CCE1  0x400 /* Remote Cache (1 hop) */
+#define PERF_MEM_LVL_REM_CCE2  0x800 /* Remote Cache (2 hops) */
+#define PERF_MEM_LVL_IO                0x1000 /* I/O memory */
+#define PERF_MEM_LVL_UNC       0x2000 /* Uncached memory */
+#define PERF_MEM_LVL_SHIFT     5
+
+/* snoop mode */
+#define PERF_MEM_SNOOP_NA      0x01 /* not available */
+#define PERF_MEM_SNOOP_NONE    0x02 /* no snoop */
+#define PERF_MEM_SNOOP_HIT     0x04 /* snoop hit */
+#define PERF_MEM_SNOOP_MISS    0x08 /* snoop miss */
+#define PERF_MEM_SNOOP_HITM    0x10 /* snoop hit modified */
+#define PERF_MEM_SNOOP_SHIFT   19
+
+/* locked instruction */
+#define PERF_MEM_LOCK_NA       0x01 /* not available */
+#define PERF_MEM_LOCK_LOCKED   0x02 /* locked transaction */
+#define PERF_MEM_LOCK_SHIFT    24
+
+/* TLB access */
+#define PERF_MEM_TLB_NA                0x01 /* not available */
+#define PERF_MEM_TLB_HIT       0x02 /* hit level */
+#define PERF_MEM_TLB_MISS      0x04 /* miss level */
+#define PERF_MEM_TLB_L1                0x08 /* L1 */
+#define PERF_MEM_TLB_L2                0x10 /* L2 */
+#define PERF_MEM_TLB_WK                0x20 /* Hardware Walker*/
+#define PERF_MEM_TLB_OS                0x40 /* OS fault handler */
+#define PERF_MEM_TLB_SHIFT     26
+
+#define PERF_MEM_S(a, s) \
+       (((__u64)PERF_MEM_##a##_##s) << PERF_MEM_##a##_SHIFT)
+
+/*
+ * single taken branch record layout:
+ *
+ *      from: source instruction (may not always be a branch insn)
+ *        to: branch target
+ *   mispred: branch target was mispredicted
+ * predicted: branch target was predicted
+ *
+ * support for mispred, predicted is optional. In case it
+ * is not supported mispred = predicted = 0.
+ *
+ *     in_tx: running in a hardware transaction
+ *     abort: aborting a hardware transaction
+ *    cycles: cycles from last branch (or 0 if not supported)
+ */
+struct perf_branch_entry {
+       __u64   from;
+       __u64   to;
+       __u64   mispred:1,  /* target mispredicted */
+               predicted:1,/* target predicted */
+               in_tx:1,    /* in transaction */
+               abort:1,    /* transaction abort */
+               cycles:16,  /* cycle count to last branch */
+               reserved:44;
+};
+
+#endif /* _UAPI_LINUX_PERF_EVENT_H */
This page took 0.036775 seconds and 4 git commands to generate.