Documentation/locking/lockstat.rst

   1 ===============
   2 Lock Statistics
   3 ===============
   4
   5 What
   6 ====
   7
   8 As the name suggests, it provides statistics on locks.
   9
  10
  11 Why
  12 ===
  13
  14 Because things like lock contention can severely impact performance.
  15
  16 How
  17 ===
  18
  19 Lockdep already has hooks in the lock functions and maps lock instances to
  20 lock classes. We build on that (see Documentation/locking/lockdep-design.rst).
  21 The graph below shows the relation between the lock functions and the various
  22 hooks therein::
  23
  24         __acquire
  25             |
  26            lock _____
  27             |        \
  28             |    __contended
  29             |         |
  30             |       <wait>
  31             | _______/
  32             |/
  33             |
  34        __acquired
  35             |
  36             .
  37           <hold>
  38             .
  39             |
  40        __release
  41             |
  42          unlock
  43
  44   lock, unlock  - the regular lock functions
  45   __*           - the hooks
  46   <>            - states
  47
  48 With these hooks we provide the following statistics:
  49
  50  con-bounces
  51         - number of lock contention that involved x-cpu data
  52  contentions
  53         - number of lock acquisitions that had to wait
  54  wait time
  55      min
  56         - shortest (non-0) time we ever had to wait for a lock
  57      max
  58         - longest time we ever had to wait for a lock
  59      total
  60         - total time we spend waiting on this lock
  61      avg
  62         - average time spent waiting on this lock
  63  acq-bounces
  64         - number of lock acquisitions that involved x-cpu data
  65  acquisitions
  66         - number of times we took the lock
  67  hold time
  68      min
  69         - shortest (non-0) time we ever held the lock
  70      max
  71         - longest time we ever held the lock
  72      total
  73         - total time this lock was held
  74      avg
  75         - average time this lock was held
  76
  77 These numbers are gathered per lock class, per read/write state (when
  78 applicable).
  79
  80 It also tracks 4 contention points per class. A contention point is a call site
  81 that had to wait on lock acquisition.
  82
  83 Configuration
  84 -------------
  85
  86 Lock statistics are enabled via CONFIG_LOCK_STAT.
  87
  88 Usage
  89 -----
  90
  91 Enable collection of statistics::
  92
  93         # echo 1 >/proc/sys/kernel/lock_stat
  94
  95 Disable collection of statistics::
  96
  97         # echo 0 >/proc/sys/kernel/lock_stat
  98
  99 Look at the current lock statistics::
 100
 101   ( line numbers not part of actual output, done for clarity in the explanation
 102     below )
 103
 104   # less /proc/lock_stat
 105
 106   01 lock_stat version 0.4
 107   02-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
 108   03                              class name    con-bounces    contentions   waittime-min   waittime-max waittime-total   waittime-avg    acq-bounces   acquisitions   holdtime-min   holdtime-max holdtime-total   holdtime-avg
 109   04-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
 110   05
 111   06                         &mm->mmap_sem-W:            46             84           0.26         939.10       16371.53         194.90          47291        2922365           0.16     2220301.69 17464026916.32        5975.99
 112   07                         &mm->mmap_sem-R:            37            100           1.31      299502.61      325629.52        3256.30         212344       34316685           0.10        7744.91    95016910.20           2.77
 113   08                         ---------------
 114   09                           &mm->mmap_sem              1          [<ffffffff811502a7>] khugepaged_scan_mm_slot+0x57/0x280
 115   10                           &mm->mmap_sem             96          [<ffffffff815351c4>] __do_page_fault+0x1d4/0x510
 116   11                           &mm->mmap_sem             34          [<ffffffff81113d77>] vm_mmap_pgoff+0x87/0xd0
 117   12                           &mm->mmap_sem             17          [<ffffffff81127e71>] vm_munmap+0x41/0x80
 118   13                         ---------------
 119   14                           &mm->mmap_sem              1          [<ffffffff81046fda>] dup_mmap+0x2a/0x3f0
 120   15                           &mm->mmap_sem             60          [<ffffffff81129e29>] SyS_mprotect+0xe9/0x250
 121   16                           &mm->mmap_sem             41          [<ffffffff815351c4>] __do_page_fault+0x1d4/0x510
 122   17                           &mm->mmap_sem             68          [<ffffffff81113d77>] vm_mmap_pgoff+0x87/0xd0
 123   18
 124   19.............................................................................................................................................................................................................................
 125   20
 126   21                         unix_table_lock:           110            112           0.21          49.24         163.91           1.46          21094          66312           0.12         624.42       31589.81           0.48
 127   22                         ---------------
 128   23                         unix_table_lock             45          [<ffffffff8150ad8e>] unix_create1+0x16e/0x1b0
 129   24                         unix_table_lock             47          [<ffffffff8150b111>] unix_release_sock+0x31/0x250
 130   25                         unix_table_lock             15          [<ffffffff8150ca37>] unix_find_other+0x117/0x230
 131   26                         unix_table_lock              5          [<ffffffff8150a09f>] unix_autobind+0x11f/0x1b0
 132   27                         ---------------
 133   28                         unix_table_lock             39          [<ffffffff8150b111>] unix_release_sock+0x31/0x250
 134   29                         unix_table_lock             49          [<ffffffff8150ad8e>] unix_create1+0x16e/0x1b0
 135   30                         unix_table_lock             20          [<ffffffff8150ca37>] unix_find_other+0x117/0x230
 136   31                         unix_table_lock              4          [<ffffffff8150a09f>] unix_autobind+0x11f/0x1b0
 137
 138
 139 This excerpt shows the first two lock class statistics. Line 01 shows the
 140 output version - each time the format changes this will be updated. Line 02-04
 141 show the header with column descriptions. Lines 05-18 and 20-31 show the actual
 142 statistics. These statistics come in two parts; the actual stats separated by a
 143 short separator (line 08, 13) from the contention points.
 144
 145 Lines 09-12 show the first 4 recorded contention points (the code
 146 which tries to get the lock) and lines 14-17 show the first 4 recorded
 147 contended points (the lock holder). It is possible that the max
 148 con-bounces point is missing in the statistics.
 149
 150 The first lock (05-18) is a read/write lock, and shows two lines above the
 151 short separator. The contention points don't match the column descriptors,
 152 they have two: contentions and [<IP>] symbol. The second set of contention
 153 points are the points we're contending with.
 154
 155 The integer part of the time values is in us.
 156
 157 Dealing with nested locks, subclasses may appear::
 158
 159   32...........................................................................................................................................................................................................................
 160   33
 161   34                               &rq->lock:       13128          13128           0.43         190.53      103881.26           7.91          97454        3453404           0.00         401.11    13224683.11           3.82
 162   35                               ---------
 163   36                               &rq->lock          645          [<ffffffff8103bfc4>] task_rq_lock+0x43/0x75
 164   37                               &rq->lock          297          [<ffffffff8104ba65>] try_to_wake_up+0x127/0x25a
 165   38                               &rq->lock          360          [<ffffffff8103c4c5>] select_task_rq_fair+0x1f0/0x74a
 166   39                               &rq->lock          428          [<ffffffff81045f98>] scheduler_tick+0x46/0x1fb
 167   40                               ---------
 168   41                               &rq->lock           77          [<ffffffff8103bfc4>] task_rq_lock+0x43/0x75
 169   42                               &rq->lock          174          [<ffffffff8104ba65>] try_to_wake_up+0x127/0x25a
 170   43                               &rq->lock         4715          [<ffffffff8103ed4b>] double_rq_lock+0x42/0x54
 171   44                               &rq->lock          893          [<ffffffff81340524>] schedule+0x157/0x7b8
 172   45
 173   46...........................................................................................................................................................................................................................
 174   47
 175   48                             &rq->lock/1:        1526          11488           0.33         388.73      136294.31          11.86          21461          38404           0.00          37.93      109388.53           2.84
 176   49                             -----------
 177   50                             &rq->lock/1        11526          [<ffffffff8103ed58>] double_rq_lock+0x4f/0x54
 178   51                             -----------
 179   52                             &rq->lock/1         5645          [<ffffffff8103ed4b>] double_rq_lock+0x42/0x54
 180   53                             &rq->lock/1         1224          [<ffffffff81340524>] schedule+0x157/0x7b8
 181   54                             &rq->lock/1         4336          [<ffffffff8103ed58>] double_rq_lock+0x4f/0x54
 182   55                             &rq->lock/1          181          [<ffffffff8104ba65>] try_to_wake_up+0x127/0x25a
 183
 184 Line 48 shows statistics for the second subclass (/1) of &rq->lock class
 185 (subclass starts from 0), since in this case, as line 50 suggests,
 186 double_rq_lock actually acquires a nested lock of two spinlocks.
 187
 188 View the top contending locks::
 189
 190   # grep : /proc/lock_stat | head
 191                         clockevents_lock:       2926159        2947636           0.15       46882.81  1784540466.34         605.41        3381345        3879161           0.00        2260.97    53178395.68          13.71
 192                      tick_broadcast_lock:        346460         346717           0.18        2257.43    39364622.71         113.54        3642919        4242696           0.00        2263.79    49173646.60          11.59
 193                   &mapping->i_mmap_mutex:        203896         203899           3.36      645530.05 31767507988.39      155800.21        3361776        8893984           0.17        2254.15    14110121.02           1.59
 194                                &rq->lock:        135014         136909           0.18         606.09      842160.68           6.15        1540728       10436146           0.00         728.72    17606683.41           1.69
 195                &(&zone->lru_lock)->rlock:         93000          94934           0.16          59.18      188253.78           1.98        1199912        3809894           0.15         391.40     3559518.81           0.93
 196                          tasklist_lock-W:         40667          41130           0.23        1189.42      428980.51          10.43         270278         510106           0.16         653.51     3939674.91           7.72
 197                          tasklist_lock-R:         21298          21305           0.20        1310.05      215511.12          10.12         186204         241258           0.14        1162.33     1179779.23           4.89
 198                               rcu_node_1:         47656          49022           0.16         635.41      193616.41           3.95         844888        1865423           0.00         764.26     1656226.96           0.89
 199        &(&dentry->d_lockref.lock)->rlock:         39791          40179           0.15        1302.08       88851.96           2.21        2790851       12527025           0.10        1910.75     3379714.27           0.27
 200                               rcu_node_0:         29203          30064           0.16         786.55     1555573.00          51.74          88963         244254           0.00         398.87      428872.51           1.76
 201
 202 Clear the statistics::
 203
 204   # echo 0 > /proc/lock_stat