lib/profiler/Profiler.cc

   1 /*
   2  * DEBUG: section 81    CPU Profiling Routines
   3  * AUTHOR: Andres Kroonmaa, Sep.2000
   4  *
   5  * SQUID Internet Object Cache  http://squid.nlanr.net/Squid/
   6  * ----------------------------------------------------------
   7  *
   8  *  Squid is the result of efforts by numerous individuals from the
   9  *  Internet community.  Development is led by Duane Wessels of the
  10  *  National Laboratory for Applied Network Research and funded by the
  11  *  National Science Foundation.  Squid is Copyrighted (C) 1998 by
  12  *  the Regents of the University of California.  Please see the
  13  *  COPYRIGHT file for full details.  Squid incorporates software
  14  *  developed and/or copyrighted by other sources.  Please see the
  15  *  CREDITS file for full details.
  16  *
  17  *  This program is free software; you can redistribute it and/or modify
  18  *  it under the terms of the GNU General Public License as published by
  19  *  the Free Software Foundation; either version 2 of the License, or
  20  *  (at your option) any later version.
  21  *
  22  *  This program is distributed in the hope that it will be useful,
  23  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
  24  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  25  *  GNU General Public License for more details.
  26  *
  27  *  You should have received a copy of the GNU General Public License
  28  *  along with this program; if not, write to the Free Software
  29  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111, USA.
  30  *
  31  */
  32
  33 /**
  34  * CPU Profiling implementation.
  35  *
  36  * \par
  37  *  This library implements the Probes needed to gather stats.
  38  *  See src/ProfStats.c which implements historical recording and
  39  *  presentation in CacheMgr.cgi.
  40  *
  41  * \par
  42  *  For timing we prefer on-CPU ops that retrieve cpu ticks counter.
  43  *  For Intel, this is "rdtsc", which is 64-bit counter that virtually
  44  *  never wraps. For alpha, this is "rpcc" which is 32-bit counter and
  45  *  wraps every few seconds. Currently, no handling of wrapping counters
  46  *  is implemented. Other CPU's are also not covered. Potentially all
  47  *  modern CPU's has similar counters.
  48  *
  49  * Usage.
  50  *  Insert macro PROF_state(probename) in strategic places in code.
  51  *    PROF_start(probename);
  52  *     ...  section of code measured ...
  53  *    PROF_stop(probename);
  54  *
  55  *   probename must be added to the xprof_type.h enum list
  56  *   with prepended "XPROF_" string.
  57  *
  58  * \section description Description.
  59  * \par PROF
  60  *  gathers stats per probename into structures. It indexes these
  61  *  structures by enum type index in an array.
  62  *
  63  * \par PROF
  64  *  records best, best, average and worst values for delta time,
  65  *  also, if UNACCED is defined, it measures "empty" time during which
  66  *  no probes are in measuring state. This allows to see time "unaccounted"
  67  *  for. If OVERHEAD is defined, additional calculations are made at every
  68  *  probe to measure approximate overhead of the probe code itself.
  69  *
  70  * \par
  71  *  Probe data is stored in linked-list, so the more probes you define,
  72  *  the more overhead is added to find the deepest nested probe. To reduce
  73  *  average overhead, linked list is manipulated each time PR_start is
  74  *  called, so that probe just started is moved 1 position up in linkedlist.
  75  *  This way frequently used probes are moved closer to the head of list,
  76  *  reducing average overhead.
  77  *  Note that all overhead is on the scale of one hundred of CPU clock
  78  *  ticks, which on the scale of submicroseconds. Yet, to optimise really
  79  *  fast and frequent sections of code, we want to reduce this overhead
  80  *  to absolute minimum possible.
  81  *
  82  * \par
  83  *  For actual measurements, probe overhead cancels out mostly. Still,
  84  *  do not take the measured times as facts, they should be viewed in
  85  *  relative comparison to overall CPU time and on the same platform.
  86  *
  87  * \par
  88  *  Every 1 second, Event within squid is called that parses gathered
  89  *  statistics of every probe, and accumulates that into historical
  90  *  structures for last 1,5,30 secs, 1,5,30 mins, and 1,5 and 24 hours.
  91  *  Each second active probe stats are reset, and only historical data
  92  *  is presented in cachemgr output.
  93  *
  94  * \section reading Reading stats.
  95  * \par
  96  *  "Worst case" may be misleading. Anything can happen at any section
  97  *  of code that could delay reaching to probe stop. For eg. system may
  98  *  need to service interrupt routine, task switch could occur, or page
  99  *  fault needs to be handled. In this sense, this is quite meaningless
 100  *  metric. "Best case" shows fastest completion of probe section, and
 101  *  is also somewhat useless, unless you know that amount of work is
 102  *  constant. Best metric to watch is "average time" and total cumulated
 103  *  time in given timeframe, which really show percentage of time spent
 104  *  in given section of code, and its average completion time. This data
 105  *  could be used to detect bottlenecks withing squid and optimise them.
 106  *
 107  * \par
 108  *  TOTALS are quite off reality. Its there just to summarise cumulative
 109  *  times and percent column. Percent values over 100% shows that there
 110  *  have been some probes nested into each other.
 111  *
 112  */
 113
 114 #include "squid.h"
 115 #include "profiler/Profiler.h"
 116
 117 #if USE_XPROF_STATS
 118
 119 #if HAVE_ASSERT_H
 120 #include <assert.h>
 121 #endif
 122 #if HAVE_GNUMALLLOC_H
 123 #include <gnumalloc.h>
 124 #elif HAVE_MALLOC_H
 125 #include <malloc.h>
 126 #endif
 127 #if HAVE_UNISTD_H
 128 #include <unistd.h>
 129 #endif
 130
 131 /* Exported Data */
 132 TimersArray *xprof_Timers = NULL;
 133
 134 /* Private stuff */
 135
 136 /* new stuff */
 137 #define MAXSTACKDEPTH   512
 138
 139 struct _callstack_entry {
 140     int timer;          /* index into timers array */
 141     const char *name;
 142     hrtime_t start, stop, accum;
 143 };
 144
 145 struct _callstack_entry cstack[MAXSTACKDEPTH];
 146 int cstack_head = 0;
 147
 148 #if defined(_MSC_VER) /* Microsoft C Compiler ONLY */
 149 static __inline void
 150 #else
 151 static inline void
 152 #endif
 153 xprof_update(xprof_stats_data * head)
 154 {
 155     if (head->delta < head->best)
 156         head->best = head->delta;
 157     if (head->worst < head->delta)
 158         head->worst = head->delta;
 159     head->summ += head->delta;
 160     ++head->count;
 161 }
 162
 163 static xprof_stats_data *xp_UNACCOUNTED;
 164 static int xprof_inited = 0;
 165
 166 static void
 167 xprof_InitLib(void)
 168 {
 169     if (xprof_inited)
 170         return;
 171
 172     xprof_Timers = static_cast<TimersArray *>(calloc(XPROF_LAST + 2, sizeof(xprof_stats_node)));
 173
 174     xprof_Timers[XPROF_PROF_UNACCOUNTED]->name = "PROF_UNACCOUNTED";
 175     xprof_Timers[XPROF_PROF_UNACCOUNTED]->accu.start = get_tick();
 176     xp_UNACCOUNTED = &xprof_Timers[XPROF_PROF_UNACCOUNTED]->accu;
 177     cstack_head = 0;
 178     xprof_inited = 1;
 179 }
 180
 181 void
 182 xprof_start(xprof_type type, const char *timer)
 183 {
 184     hrtime_t tt = get_tick();
 185     if (!xprof_inited)
 186         xprof_InitLib();
 187
 188     /* If nested, stop current stack frame */
 189     if (cstack_head > 0) {
 190         cstack[cstack_head - 1].accum += get_tick() - cstack[cstack_head - 1].start;
 191         cstack[cstack_head - 1].start = -1;
 192     }
 193
 194     /* Are we at the first level? If so; stop the unaccounted timer */
 195     if (cstack_head == 0) {
 196         assert(xp_UNACCOUNTED->start != -1);
 197         xp_UNACCOUNTED->delta = tt - xp_UNACCOUNTED->start;
 198         xp_UNACCOUNTED->start = -1;
 199         xprof_update(xp_UNACCOUNTED);
 200     }
 201
 202     /* Allocate new stack frame */
 203     cstack[cstack_head].start = tt;
 204     cstack[cstack_head].stop = -1;
 205     cstack[cstack_head].accum = 0;
 206     cstack[cstack_head].timer = type;
 207     cstack[cstack_head].name = timer;
 208     ++cstack_head;
 209     assert(cstack_head < MAXSTACKDEPTH);
 210
 211 }
 212
 213 void
 214 xprof_stop(xprof_type type, const char *timer)
 215 {
 216     hrtime_t tt = get_tick();
 217     assert(cstack_head > 0);
 218     --cstack_head;
 219     assert(cstack[cstack_head].timer == type);
 220
 221     /* Record timer details */
 222     cstack[cstack_head].accum += tt - cstack[cstack_head].start;
 223     xprof_Timers[type]->accu.delta = cstack[cstack_head].accum;
 224     xprof_Timers[type]->name = timer;
 225
 226     /* Update */
 227     xprof_update(&xprof_Timers[type]->accu);
 228
 229     /* Restart previous timer if we're not at the top level */
 230     if (cstack_head > 0) {
 231         cstack[cstack_head - 1].start = tt;
 232         cstack[cstack_head - 1].stop = 0;
 233         return;
 234     }
 235     /* Get here? We're at the top level; unaccounted */
 236     xp_UNACCOUNTED->start = tt;
 237 }
 238
 239 #endif /* USE_XPROF_STATS */