1 /*
   2  * Copyright (c) 2018, Oracle and/or its affiliates. All rights reserved.
   3  * Copyright (c) 2018, Google and/or its affiliates. All rights reserved.
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This code is free software; you can redistribute it and/or modify it
   7  * under the terms of the GNU General Public License version 2 only, as
   8  * published by the Free Software Foundation.
   9  *
  10  * This code is distributed in the hope that it will be useful, but WITHOUT
  11  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  12  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  13  * version 2 for more details (a copy is included in the LICENSE file that
  14  * accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License version
  17  * 2 along with this work; if not, write to the Free Software Foundation,
  18  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  19  *
  20  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  21  * or visit www.oracle.com if you need additional information or have any
  22  * questions.
  23  *
  24  */
  25 
  26 #include "precompiled.hpp"
  27 #include "runtime/handles.inline.hpp"
  28 #include "runtime/orderAccess.hpp"
  29 #include "runtime/sharedRuntime.hpp"
  30 #include "runtime/threadHeapSampler.hpp"
  31 
  32 // Cheap random number generator
  33 uint64_t ThreadHeapSampler::_rnd;
  34 // Default is 512kb.
  35 int ThreadHeapSampler::_sampling_interval = 512 * 1024;
  36 int ThreadHeapSampler::_enabled;
  37 
  38 // Statics for the fast log
  39 static const int FastLogNumBits = 10;
  40 static const int FastLogMask = (1 << FastLogNumBits) - 1;
  41 static double log_table[1<<FastLogNumBits];  // Constant
  42 static bool log_table_initialized;
  43 
  44 // Returns the next prng value.
  45 // pRNG is: aX+b mod c with a = 0x5DEECE66D, b =  0xB, c = 1<<48
  46 // This is the lrand64 generator.
  47 static uint64_t next_random(uint64_t rnd) {
  48   const uint64_t PrngMult = 0x5DEECE66DLL;
  49   const uint64_t PrngAdd = 0xB;
  50   const uint64_t PrngModPower = 48;
  51   const uint64_t PrngModMask = ((uint64_t)1 << PrngModPower) - 1;
  52   //assert(IS_SAFE_SIZE_MUL(PrngMult, rnd), "Overflow on multiplication.");
  53   //assert(IS_SAFE_SIZE_ADD(PrngMult * rnd, PrngAdd), "Overflow on addition.");
  54   return (PrngMult * rnd + PrngAdd) & PrngModMask;
  55 }
  56 
  57 static double fast_log2(const double & d) {
  58   assert(d>0, "bad value passed to assert");
  59   uint64_t x = 0;
  60   assert(sizeof(d) == sizeof(x),
  61          "double and uint64_t do not have the same size");
  62   x = *reinterpret_cast<const uint64_t*>(&d);
  63   const uint32_t x_high = x >> 32;
  64   assert(FastLogNumBits <= 20, "FastLogNumBits should be less than 20.");
  65   const uint32_t y = x_high >> (20 - FastLogNumBits) & FastLogMask;
  66   const int32_t exponent = ((x_high >> 20) & 0x7FF) - 1023;
  67   return exponent + log_table[y];
  68 }
  69 
  70 // Generates a geometric variable with the specified mean (512K by default).
  71 // This is done by generating a random number between 0 and 1 and applying
  72 // the inverse cumulative distribution function for an exponential.
  73 // Specifically: Let m be the inverse of the sample interval, then
  74 // the probability distribution function is m*exp(-mx) so the CDF is
  75 // p = 1 - exp(-mx), so
  76 // q = 1 - p = exp(-mx)
  77 // log_e(q) = -mx
  78 // -log_e(q)/m = x
  79 // log_2(q) * (-log_e(2) * 1/m) = x
  80 // In the code, q is actually in the range 1 to 2**26, hence the -26 below
  81 void ThreadHeapSampler::pick_next_geometric_sample() {
  82   _rnd = next_random(_rnd);
  83   // Take the top 26 bits as the random number
  84   // (This plus a 1<<58 sampling bound gives a max possible step of
  85   // 5194297183973780480 bytes.  In this case,
  86   // for sample_parameter = 1<<19, max possible step is
  87   // 9448372 bytes (24 bits).
  88   const uint64_t PrngModPower = 48;  // Number of bits in prng
  89   // The uint32_t cast is to prevent a (hard-to-reproduce) NAN
  90   // under piii debug for some binaries.
  91   double q = static_cast<uint32_t>(_rnd >> (PrngModPower - 26)) + 1.0;
  92   // Put the computed p-value through the CDF of a geometric.
  93   // For faster performance (save ~1/20th exec time), replace
  94   // min(0.0, FastLog2(q) - 26)  by  (Fastlog2(q) - 26.000705)
  95   // The value 26.000705 is used rather than 26 to compensate
  96   // for inaccuracies in FastLog2 which otherwise result in a
  97   // negative answer.
  98   double log_val = (fast_log2(q) - 26);
  99   double result =
 100       (0.0 < log_val ? 0.0 : log_val) * (-log(2.0) * (get_sampling_interval())) + 1;
 101   assert(result > 0 && result < SIZE_MAX, "Result is not in an acceptable range.");
 102   size_t interval = static_cast<size_t>(result);
 103   _bytes_until_sample = interval;
 104 }
 105 
 106 void ThreadHeapSampler::pick_next_sample(size_t overflowed_bytes) {
 107   // Explicitly test if the sampling interval is 0, return 0 to sample every
 108   // allocation.
 109   if (get_sampling_interval() == 0) {
 110     _bytes_until_sample = 0;
 111     return;
 112   }
 113 
 114   pick_next_geometric_sample();
 115 
 116   // Try to correct sample size by removing extra space from last allocation.
 117   if (overflowed_bytes > 0 && _bytes_until_sample > overflowed_bytes) {
 118     _bytes_until_sample -= overflowed_bytes;
 119   }
 120 }
 121 
 122 void ThreadHeapSampler::check_for_sampling(oop obj, size_t allocation_size, size_t bytes_since_allocation) {
 123   size_t total_allocated_bytes = bytes_since_allocation + allocation_size;
 124 
 125   // If not yet time for a sample, skip it.
 126   if (total_allocated_bytes < _bytes_until_sample) {
 127     _bytes_until_sample -= total_allocated_bytes;
 128     return;
 129   }
 130 
 131   JvmtiExport::sampled_object_alloc_event_collector(obj);
 132 
 133   size_t overflow_bytes = total_allocated_bytes - _bytes_until_sample;
 134   pick_next_sample(overflow_bytes);
 135 }
 136 
 137 void ThreadHeapSampler::init_log_table() {
 138   MutexLockerEx mu(ThreadHeapSampler_lock, Mutex::_no_safepoint_check_flag);
 139 
 140   if (log_table_initialized) {
 141     return;
 142   }
 143 
 144   for (int i = 0; i < (1 << FastLogNumBits); i++) {
 145     log_table[i] = (log(1.0 + static_cast<double>(i+0.5) / (1 << FastLogNumBits))
 146                     / log(2.0));
 147   }
 148 
 149   log_table_initialized = true;
 150 }
 151 
 152 void ThreadHeapSampler::enable() {
 153   // Done here to be done when things have settled. This adds a mutex lock but
 154   // presumably, users won't be enabling and disabling all the time.
 155   init_log_table();
 156   OrderAccess::release_store(&_enabled, 1);
 157 }
 158 
 159 int ThreadHeapSampler::enabled() {
 160   return OrderAccess::load_acquire(&_enabled);
 161 }
 162 
 163 void ThreadHeapSampler::disable() {
 164   OrderAccess::release_store(&_enabled, 0);
 165 }
 166 
 167 int ThreadHeapSampler::get_sampling_interval() {
 168   return OrderAccess::load_acquire(&_sampling_interval);
 169 }
 170 
 171 void ThreadHeapSampler::set_sampling_interval(int sampling_interval) {
 172   OrderAccess::release_store(&_sampling_interval, sampling_interval);
 173 }
 174 
 175 // Methods used in assertion mode to check if a collector is present or not at
 176 // the moment of TLAB sampling, ie a slow allocation path.
 177 bool ThreadHeapSampler::sampling_collector_present() const {
 178   return _collectors_present > 0;
 179 }
 180 
 181 bool ThreadHeapSampler::remove_sampling_collector() {
 182   assert(_collectors_present > 0, "Problem with collector counter.");
 183   _collectors_present--;
 184   return true;
 185 }
 186 
 187 bool ThreadHeapSampler::add_sampling_collector() {
 188   _collectors_present++;
 189   return true;
 190 }