Home | History | Annotate | Download | only in simpleperf
      1 /*
      2  * Copyright (C) 2015 The Android Open Source Project
      3  *
      4  * Licensed under the Apache License, Version 2.0 (the "License");
      5  * you may not use this file except in compliance with the License.
      6  * You may obtain a copy of the License at
      7  *
      8  *      http://www.apache.org/licenses/LICENSE-2.0
      9  *
     10  * Unless required by applicable law or agreed to in writing, software
     11  * distributed under the License is distributed on an "AS IS" BASIS,
     12  * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
     13  * See the License for the specific language governing permissions and
     14  * limitations under the License.
     15  */
     16 #define ATRACE_TAG ATRACE_TAG_ALWAYS
     17 #include "event_fd.h"
     18 
     19 #include <fcntl.h>
     20 #include <stdio.h>
     21 #include <string.h>
     22 #include <sys/ioctl.h>
     23 #include <sys/mman.h>
     24 #include <sys/syscall.h>
     25 #include <sys/types.h>
     26 #include <atomic>
     27 #include <memory>
     28 #include <cutils/trace.h>
     29 #include <utils/Trace.h>
     30 
     31 #include <android-base/file.h>
     32 #include <android-base/logging.h>
     33 #include <android-base/stringprintf.h>
     34 
     35 #include "environment.h"
     36 #include "event_attr.h"
     37 #include "event_type.h"
     38 #include "perf_event.h"
     39 #include "utils.h"
     40 
     41 static int perf_event_open(const perf_event_attr& attr, pid_t pid, int cpu,
     42                            int group_fd, unsigned long flags) {  // NOLINT
     43   return syscall(__NR_perf_event_open, &attr, pid, cpu, group_fd, flags);
     44 }
     45 
     46 std::unique_ptr<EventFd> EventFd::OpenEventFile(const perf_event_attr& attr,
     47                                                 pid_t tid, int cpu,
     48                                                 EventFd* group_event_fd,
     49                                                 bool report_error) {
     50   std::string event_name = GetEventNameByAttr(attr);
     51   int group_fd = -1;
     52   if (group_event_fd != nullptr) {
     53     group_fd = group_event_fd->perf_event_fd_;
     54   }
     55   perf_event_attr real_attr = attr;
     56   if (attr.freq) {
     57     uint64_t max_sample_freq;
     58     if (GetMaxSampleFrequency(&max_sample_freq) && max_sample_freq < attr.sample_freq) {
     59       static bool warned = false;
     60       if (!warned) {
     61         warned = true;
     62         LOG(INFO) << "Adjust sample freq to max allowed sample freq " << max_sample_freq;
     63       }
     64       real_attr.sample_freq = max_sample_freq;
     65     }
     66   }
     67   int perf_event_fd = perf_event_open(real_attr, tid, cpu, group_fd, 0);
     68   if (perf_event_fd == -1) {
     69     if (report_error) {
     70       PLOG(ERROR) << "open perf_event_file (event " << event_name << ", tid "
     71                   << tid << ", cpu " << cpu << ", group_fd " << group_fd
     72                   << ") failed";
     73     } else {
     74       PLOG(DEBUG) << "open perf_event_file (event " << event_name << ", tid "
     75                   << tid << ", cpu " << cpu << ", group_fd " << group_fd
     76                   << ") failed";
     77     }
     78     return nullptr;
     79   }
     80   if (fcntl(perf_event_fd, F_SETFD, FD_CLOEXEC) == -1) {
     81     if (report_error) {
     82       PLOG(ERROR) << "fcntl(FD_CLOEXEC) for perf_event_file (event "
     83                   << event_name << ", tid " << tid << ", cpu " << cpu
     84                   << ", group_fd " << group_fd << ") failed";
     85     } else {
     86       PLOG(DEBUG) << "fcntl(FD_CLOEXEC) for perf_event_file (event "
     87                   << event_name << ", tid " << tid << ", cpu " << cpu
     88                   << ", group_fd " << group_fd << ") failed";
     89     }
     90     return nullptr;
     91   }
     92   return std::unique_ptr<EventFd>(
     93       new EventFd(real_attr, perf_event_fd, event_name, tid, cpu));
     94 }
     95 
     96 EventFd::~EventFd() {
     97   DestroyMappedBuffer();
     98   close(perf_event_fd_);
     99 }
    100 
    101 std::string EventFd::Name() const {
    102   return android::base::StringPrintf(
    103       "perf_event_file(event %s, tid %d, cpu %d)", event_name_.c_str(), tid_,
    104       cpu_);
    105 }
    106 
    107 uint64_t EventFd::Id() const {
    108   if (id_ == 0) {
    109     PerfCounter counter;
    110     if (InnerReadCounter(&counter)) {
    111       id_ = counter.id;
    112     }
    113   }
    114   return id_;
    115 }
    116 
    117 bool EventFd::EnableEvent() {
    118   int result = ioctl(perf_event_fd_, PERF_EVENT_IOC_ENABLE, 0);
    119   if (result < 0) {
    120     PLOG(ERROR) << "ioctl(enable) " << Name() << " failed";
    121     return false;
    122   }
    123   return true;
    124 }
    125 
    126 bool EventFd::InnerReadCounter(PerfCounter* counter) const {
    127   CHECK(counter != nullptr);
    128   if (!android::base::ReadFully(perf_event_fd_, counter, sizeof(*counter))) {
    129     PLOG(ERROR) << "ReadCounter from " << Name() << " failed";
    130     return false;
    131   }
    132   return true;
    133 }
    134 
    135 bool EventFd::ReadCounter(PerfCounter* counter) {
    136   if (!InnerReadCounter(counter)) {
    137     return false;
    138   }
    139   // Trace is always available to systrace if enabled
    140   if (tid_ > 0) {
    141     ATRACE_INT64(android::base::StringPrintf(
    142                    "%s_tid%d_cpu%d", event_name_.c_str(), tid_,
    143                    cpu_).c_str(), counter->value - last_counter_value_);
    144   } else {
    145     ATRACE_INT64(android::base::StringPrintf(
    146                    "%s_cpu%d", event_name_.c_str(),
    147                    cpu_).c_str(), counter->value - last_counter_value_);
    148   }
    149   last_counter_value_ = counter->value;
    150   return true;
    151 }
    152 
    153 bool EventFd::CreateMappedBuffer(size_t mmap_pages, bool report_error) {
    154   CHECK(IsPowerOfTwo(mmap_pages));
    155   size_t page_size = sysconf(_SC_PAGE_SIZE);
    156   size_t mmap_len = (mmap_pages + 1) * page_size;
    157   void* mmap_addr = mmap(nullptr, mmap_len, PROT_READ | PROT_WRITE, MAP_SHARED,
    158                          perf_event_fd_, 0);
    159   if (mmap_addr == MAP_FAILED) {
    160     bool is_perm_error = (errno == EPERM);
    161     if (report_error) {
    162       PLOG(ERROR) << "mmap(" << mmap_pages << ") failed for " << Name();
    163     } else {
    164       PLOG(DEBUG) << "mmap(" << mmap_pages << ") failed for " << Name();
    165     }
    166     if (report_error && is_perm_error) {
    167       LOG(ERROR)
    168           << "It seems the kernel doesn't allow allocating enough "
    169           << "buffer for dumping samples, consider decreasing mmap pages(-m).";
    170     }
    171     return false;
    172   }
    173   mmap_addr_ = mmap_addr;
    174   mmap_len_ = mmap_len;
    175   mmap_metadata_page_ = reinterpret_cast<perf_event_mmap_page*>(mmap_addr_);
    176   mmap_data_buffer_ = reinterpret_cast<char*>(mmap_addr_) + page_size;
    177   mmap_data_buffer_size_ = mmap_len_ - page_size;
    178   return true;
    179 }
    180 
    181 bool EventFd::ShareMappedBuffer(const EventFd& event_fd, bool report_error) {
    182   CHECK(!HasMappedBuffer());
    183   CHECK(event_fd.HasMappedBuffer());
    184   int result =
    185       ioctl(perf_event_fd_, PERF_EVENT_IOC_SET_OUTPUT, event_fd.perf_event_fd_);
    186   if (result != 0) {
    187     if (report_error) {
    188       PLOG(ERROR) << "failed to share mapped buffer of "
    189                   << event_fd.perf_event_fd_ << " with " << perf_event_fd_;
    190     }
    191     return false;
    192   }
    193   return true;
    194 }
    195 
    196 void EventFd::DestroyMappedBuffer() {
    197   if (HasMappedBuffer()) {
    198     munmap(mmap_addr_, mmap_len_);
    199     mmap_addr_ = nullptr;
    200     mmap_len_ = 0;
    201     mmap_metadata_page_ = nullptr;
    202     mmap_data_buffer_ = nullptr;
    203     mmap_data_buffer_size_ = 0;
    204   }
    205 }
    206 
    207 size_t EventFd::GetAvailableMmapData(std::vector<char>& buffer, size_t& buffer_pos) {
    208   if (!HasMappedBuffer()) {
    209     return 0;
    210   }
    211   // The mmap_data_buffer is used as a ring buffer between the kernel and
    212   // simpleperf. The kernel continuously writes records to the buffer, and
    213   // simpleperf continuously read records out.
    214   //         _________________________________________
    215   // buffer | can write   |   can read   |  can write |
    216   //                      ^              ^
    217   //                    read_head       write_head
    218   //
    219   // So simpleperf can read records in [read_head, write_head), and the kernel
    220   // can write records in [write_head, read_head). The kernel is responsible
    221   // for updating write_head, and simpleperf is responsible for updating
    222   // read_head.
    223 
    224   size_t buf_mask = mmap_data_buffer_size_ - 1;
    225   size_t write_head =
    226       static_cast<size_t>(mmap_metadata_page_->data_head & buf_mask);
    227   size_t read_head =
    228       static_cast<size_t>(mmap_metadata_page_->data_tail & buf_mask);
    229 
    230   if (read_head == write_head) {
    231     // No available data.
    232     return 0;
    233   }
    234   size_t read_bytes;
    235   if (read_head < write_head) {
    236     read_bytes = write_head - read_head;
    237   } else {
    238     read_bytes = mmap_data_buffer_size_ - read_head + write_head;
    239   }
    240   // Extend the buffer if it is not big enough.
    241   if (buffer.size() < buffer_pos + read_bytes) {
    242     buffer.resize(buffer_pos + read_bytes);
    243   }
    244 
    245   // rmb() used to ensure reading data after reading data_head.
    246   __sync_synchronize();
    247 
    248   // Copy records from mapped buffer. Note that records can be wrapped at the
    249   // end of the mapped buffer.
    250   char* to = &buffer[buffer_pos];
    251   if (read_head < write_head) {
    252     char* from = mmap_data_buffer_ + read_head;
    253     size_t n = write_head - read_head;
    254     memcpy(to, from, n);
    255   } else {
    256     char* from = mmap_data_buffer_ + read_head;
    257     size_t n = mmap_data_buffer_size_ - read_head;
    258     memcpy(to, from, n);
    259     to += n;
    260     from = mmap_data_buffer_;
    261     n = write_head;
    262     memcpy(to, from, n);
    263   }
    264   buffer_pos += read_bytes;
    265   DiscardMmapData(read_bytes);
    266   return read_bytes;
    267 }
    268 
    269 void EventFd::DiscardMmapData(size_t discard_size) {
    270   // mb() used to ensure finish reading data before writing data_tail.
    271   __sync_synchronize();
    272   mmap_metadata_page_->data_tail += discard_size;
    273 }
    274 
    275 bool EventFd::StartPolling(IOEventLoop& loop,
    276                            const std::function<bool()>& callback) {
    277   ioevent_ref_ = loop.AddReadEvent(perf_event_fd_, callback);
    278   return ioevent_ref_ != nullptr;
    279 }
    280 
    281 bool EventFd::StopPolling() { return IOEventLoop::DelEvent(ioevent_ref_); }
    282 
    283 bool IsEventAttrSupported(const perf_event_attr& attr) {
    284   if (attr.type == SIMPLEPERF_TYPE_USER_SPACE_SAMPLERS &&
    285       attr.config == SIMPLEPERF_CONFIG_INPLACE_SAMPLER) {
    286     // User space samplers don't need kernel support.
    287     return true;
    288   }
    289   std::unique_ptr<EventFd> event_fd = EventFd::OpenEventFile(attr, getpid(), -1, nullptr, false);
    290   return event_fd != nullptr;
    291 }
    292