[LU-16272] High CPU usage in libcfs_hash_for_each_empty()

Related — Fri, 9 Feb 2024 16:56:28 +0000

Here are ftrace logs for cancelling of 10k ldlm locks on client disconnect
ldlm_cancel_locks_for_export() / cfs_hash_for_each_relax():

First function calls are fast:

# tracer: function_graph
#
# CPU  DURATION                  FUNCTION CALLS
# |     |   |                     |   |   |   |
 0) + 30.305 us   |  cfs_hash_for_each_relax [libcfs]();
 0) + 13.570 us   |  cfs_hash_for_each_relax [libcfs]();
 0) + 16.458 us   |  cfs_hash_for_each_relax [libcfs]();
 1) + 22.303 us   |  cfs_hash_for_each_relax [libcfs]();
 0) + 11.710 us   |  cfs_hash_for_each_relax [libcfs]();
 0) + 10.227 us   |  cfs_hash_for_each_relax [libcfs]();
 1) + 12.137 us   |  cfs_hash_for_each_relax [libcfs]();
 0) + 10.632 us   |  cfs_hash_for_each_relax [libcfs]();
 1) + 11.427 us   |  cfs_hash_for_each_relax [libcfs]();

the last ones are critically slow:

3) ! 408.016 us  |  cfs_hash_for_each_relax [libcfs]();
 3) ! 425.611 us  |  cfs_hash_for_each_relax [libcfs]();
 3) ! 411.456 us  |  cfs_hash_for_each_relax [libcfs]();
 3) ! 408.476 us  |  cfs_hash_for_each_relax [libcfs]();
 3) ! 410.266 us  |  cfs_hash_for_each_relax [libcfs]();
 3) ! 409.032 us  |  cfs_hash_for_each_relax [libcfs]();
 3) ! 408.790 us  |  cfs_hash_for_each_relax [libcfs]();
 3) ! 409.062 us  |  cfs_hash_for_each_relax [libcfs]();
 3) ! 409.492 us  |  cfs_hash_for_each_relax [libcfs]();
 3) ! 839.014 us  |  cfs_hash_for_each_relax [libcfs]();
 3) ! 409.712 us  |  cfs_hash_for_each_relax [libcfs]();

i.e. cfs_hash_for_each_relax becoming slow when first hash buckets are emptied.
it matches the perf (cpu profiling) results in one of the above comments, that CPU consumed mostly by
cfs_hash_hh_hhead() – the function is used to check whether hlists in a bucket are empty.

     3.83%     0.00%  [k] ret_from_fork                                         -      -            
            |          
             --3.81%--kthread
                       |          
                       |--1.85%--ldlm_bl_thread_main
                       |          ldlm_export_cancel_locks
                       |          cfs_hash_for_each_empty
                       |          cfs_hash_for_each_relax
                       |          |          
                       |           --1.09%--cfs_hash_hh_hhead
                       |

ldlm_export_cancel_locks() call consumes 2.3 seconds:

[root@devvm3 tests]# cat ~/trace.txt
# tracer: function_graph
#
# CPU  DURATION                  FUNCTION CALLS
# |     |   |                     |   |   |   |
 2) $ 2397334 us  |  ldlm_export_cancel_locks [ptlrpc]();
 1)               |  ldlm_export_cancel_locks [ptlrpc]() {
 0) $ 2384598 us  |  } /* ldlm_export_cancel_locks [ptlrpc] */
[root@devvm3 tests]#

with a small libcfs_hash improvement (will be submitted soon), the function call takes only 0.1 sec:

# tracer: function_graph
#
# CPU  DURATION                  FUNCTION CALLS
# |     |   |                     |   |   |   |
 0)               |  ldlm_export_cancel_locks [ptlrpc]() {
 0)   8.154 us    |    lu_env_init [obdclass]();
 0) * 91503.19 us |    cfs_hash_for_each_empty [libcfs]();
 3)               |  ldlm_export_cancel_locks [ptlrpc]() {
 3)   7.818 us    |    lu_env_init [obdclass]();
 3) * 92138.99 us |    cfs_hash_for_each_empty [libcfs]();
 0) + 34.008 us   |    ldlm_reprocess_list [ptlrpc]();
 0)   1.183 us    |    libcfs_debug_msg [libcfs]();
 0) + 13.812 us   |    lu_env_fini [obdclass]();
 0) * 91564.21 us |  }
 3) + 36.190 us   |    ldlm_reprocess_list [ptlrpc]();
 3)   1.249 us    |    libcfs_debug_msg [libcfs]();
 3)   9.312 us    |    lu_env_fini [obdclass]();
 3) * 92197.71 us |  }

my test script, for a reference:

cd /sys/kernel/debug/tracing/
echo ldlm_export_cancel_locks > set_graph_function 
# echo cfs_hash_for_each_relax > set_graph_function
# echo ldlm_cancel_locks_for_export_cb > set_graph_function 
# echo ldlm_export_cancel_locks > /sys/kernel/debug/tracing/set_ftrace_filter 
echo > /sys/kernel/debug/tracing/set_ftrace_filter 
echo 1 > max_graph_depth
echo function_graph > current_tracer 
echo > trace
cd -

for n in {1..20}; do touch /mnt/lustre/foo-$n; for x in {1..1000}; do echo ladvise -a lockahead -s $(( 2 * x ))M -l 1M -m WRITE /mnt/lustre/foo-$n; done; done | ../utils/lfs 
../utils/lctl get_param ldlm.namespaces.filter-*.lock_count
LUUID=$(../utils/lctl get_param -n llite.*.uuid)
../utils/lctl set_param obdfilter.lustre-OST*.evict_client=$LUUID
sleep 3
../utils/lctl get_param ldlm.namespaces.filter-*.lock_count
cat /sys/kernel/debug/tracing/trace > ~/trace.txt
cat ~/trace.txt

Whamcloud Community JIRA

[LU-16272] High CPU usage in libcfs_hash_for_each_empty()