On 6/29/22 13:45, Eelco Chaudron wrote:
> 
> 
> On 29 Jun 2022, at 12:22, Ilya Maximets wrote:
> 
>> dp_netdev_input_outer_avx512 allocates a 16KB scratch pad per PMD
>> thread, but it's never freed.  This may cause significant memory
>> drain in dynamic environments.
>>
>>   ==4068109==ERROR: LeakSanitizer: detected memory leaks
>>
>>   Direct leak of 38656 byte(s) in 2 object(s) allocated from:
>>    0 0xf069fd in posix_memalign (vswitchd/ovs-vswitchd+0xf069fd)
>>    1 0x1d7ed14 in xmalloc_size_align lib/util.c:254:13
>>    2 0x1d7ed14 in xmalloc_pagealign lib/util.c:352:12
>>    3 0x2098254 in dp_netdev_input_outer_avx512 lib/dpif-netdev-avx512.c:69:17
>>    4 0x191591a in dp_netdev_process_rxq_port lib/dpif-netdev.c:5332:19
>>    5 0x190a961 in pmd_thread_main lib/dpif-netdev.c:6963:17
>>    6 0x1c4b69a in ovsthread_wrapper lib/ovs-thread.c:422:12
>>    7 0x7fd5ea6f1179 in start_thread pthread_create.c
>>
>>  SUMMARY: AddressSanitizer: 38656 byte(s) leaked in 2 allocation(s).
>>
>> Fixes: 9ac84a1a3698 ("dpif-avx512: Add ISA implementation of dpif.")
>> Signed-off-by: Ilya Maximets <[email protected]>
>> ---
> 
> Good catch... Do you have some script to run all kinds of sanitizers?

No.  For this one I just built with
  CFLAGS="-msse4.2 -O1 -fno-omit-frame-pointer -fno-common -g 
-fsanitize=address -fsanitize=undefined" CC=clang

And then tried to run with asan environment variables:

  ASAN_OPTIONS='detect_leaks=1:abort_on_error=true:log_path=asan'
  UBSAN_OPTIONS='print_stacktrace=1:halt_on_error=true:log_path=ubsan'

LeakSanitizer is part of the AddressSanitizer, you just need to ask to
detect leaks (we're disabling that by default in tests/atlocal.in).

I didn't check, but I guess, you should be able to reproduce the
issue by just building with Asan and running:

  ASAN_OPTIONS='detect_leaks=1:abort_on_error=true:log_path=asan' make 
check-dpdk

> 
> 
> Acked-by: Eelco Chaudron <[email protected]>
> 

_______________________________________________
dev mailing list
[email protected]
https://mail.openvswitch.org/mailman/listinfo/ovs-dev

Reply via email to