#include "Python.h" #include "pycore_ceval.h" // _PyPerf_Callbacks #include "pycore_frame.h" #include "pycore_interp.h" #ifdef PY_HAVE_PERF_TRAMPOLINE #include <fcntl.h> #include <stdio.h> #include <stdlib.h> #include <sys/mman.h> // mmap() #include <sys/types.h> #include <unistd.h> // sysconf() #include <sys/time.h> // gettimeofday() #include <sys/syscall.h> // ---------------------------------- // Perf jitdump API // ---------------------------------- PerfMapJitState; static PerfMapJitState perf_jit_map_state; /* Usually the binary and libraries are mapped in separate region like below: address -> --+---------------------+--//--+---------------------+-- | .text | .data | ... | | .text | .data | ... | --+---------------------+--//--+---------------------+-- myprog libc.so So it'd be easy and straight-forward to find a mapped binary or library from an address. But for JIT code, the code arena only cares about the code section. But the resulting DSOs (which is generated by perf inject -j) contain ELF headers and unwind info too. Then it'd generate following address space with synthesized MMAP events. Let's say it has a sample between address B and C. sample | address -> A B v C --------------------------------------------------------------------------------------------------- /tmp/jitted-PID-0.so | (headers) | .text | unwind info | /tmp/jitted-PID-1.so | (headers) | .text | unwind info | /tmp/jitted-PID-2.so | (headers) | .text | unwind info | ... --------------------------------------------------------------------------------------------------- If it only maps the .text section, it'd find the jitted-PID-1.so but cannot see the unwind info. If it maps both .text section and unwind sections, the sample could be mapped to either jitted-PID-0.so or jitted-PID-1.so and it's confusing which one is right. So to make perf happy we have non-overlapping ranges for each DSO: address -> ------------------------------------------------------------------------------------------------------- /tmp/jitted-PID-0.so | (headers) | .text | unwind info | /tmp/jitted-PID-1.so | (headers) | .text | unwind info | /tmp/jitted-PID-2.so | (headers) | .text | unwind info | ... ------------------------------------------------------------------------------------------------------- As the trampolines are constant, we add a constant padding but in general the padding needs to have the size of the unwind info rounded to 16 bytes. In general, for our trampolines this is 0x50 */ #define PERF_JIT_CODE_PADDING … #define trampoline_api … uword; CodeComments; #define Pd … #define MB … #define EM_386 … #define EM_X86_64 … #define EM_ARM … #define EM_AARCH64 … #define EM_RISCV … #define TARGET_ARCH_IA32 … #define TARGET_ARCH_X64 … #define TARGET_ARCH_ARM … #define TARGET_ARCH_ARM64 … #define TARGET_ARCH_RISCV32 … #define TARGET_ARCH_RISCV64 … #define FLAG_generate_perf_jitdump … #define FLAG_write_protect_code … #define FLAG_write_protect_vm_isolate … #define FLAG_code_comments … #define UNREACHABLE() … static uword GetElfMachineArchitecture(void) { … } Header; enum PerfEvent { … }; struct BaseEvent { … }; CodeLoadEvent; CodeUnwindingInfoEvent; static const intptr_t nanoseconds_per_second = …; // Dwarf encoding constants static const uint8_t DwarfUData4 = …; static const uint8_t DwarfSData4 = …; static const uint8_t DwarfPcRel = …; static const uint8_t DwarfDataRel = …; // static uint8_t DwarfOmit = 0xff; EhFrameHeader; static int64_t get_current_monotonic_ticks(void) { … } static int64_t get_current_time_microseconds(void) { … } static size_t round_up(int64_t value, int64_t multiple) { … } static void perf_map_jit_write_fully(const void* buffer, size_t size) { … } static void perf_map_jit_write_header(int pid, FILE* out_file) { … } static void* perf_map_jit_init(void) { … } /* DWARF definitions. */ #define DWRF_CIE_VERSION … enum { … }; enum { … }; enum { … }; enum { … }; enum { … }; enum { … }; enum { … }; enum { … }; enum { … }; ELFObjectContext; /* Append a null-terminated string. */ static uint32_t elfctx_append_string(ELFObjectContext* ctx, const char* str) { … } /* Append a SLEB128 value. */ static void elfctx_append_sleb128(ELFObjectContext* ctx, int32_t v) { … } /* Append a ULEB128 to buffer. */ static void elfctx_append_uleb128(ELFObjectContext* ctx, uint32_t v) { … } /* Shortcuts to generate DWARF structures. */ #define DWRF_U8(x) … #define DWRF_I8(x) … #define DWRF_U16(x) … #define DWRF_U32(x) … #define DWRF_ADDR(x) … #define DWRF_UV(x) … #define DWRF_SV(x) … #define DWRF_STR(str) … #define DWRF_ALIGNNOP(s) … #define DWRF_SECTION(name, stmt) … /* Initialize .eh_frame section. */ static void elf_init_ehframe(ELFObjectContext* ctx) { … } static void perf_map_jit_write_entry(void *state, const void *code_addr, unsigned int code_size, PyCodeObject *co) { … } static int perf_map_jit_fini(void* state) { … } _PyPerf_Callbacks _Py_perfmap_jit_callbacks = …; #endif
Codebase Browser
cpython