Actually,
this was first question I've asked myself ;) So I've made test app:
#include <stdio.h>
#include <sys/types.h>
#include <unistd.h>
int main(void)
{
int i;
int c = 0;
for (i = 0; i < 100000000;i++) {
// c = c + getpid();
// c = c + rand();
}
printf("%u\n", c);
return 0;
}
And if I uncomment c = c + getpid line, runtime is (VM on ~3 years old
quite low-end Xeon, compiled with -O2) ->
real 0m0.515s
user 0m0.511s
sys 0m0.004s
For c = c + rand(), it's:
real 0m1.436s
user 0m1.432s
sys 0m0.004s
In other words, it's for sure not 100% free, but 100000000 in 0.5s (->
one getpid takes ~ .00000000515000000746 sec) is probably ok.
Honza
Fabio M. Di Nitto napsal(a):
> The patch looks good, but I have one simple question... isn't expensive
> to call getpid() on each log_rec?
>
> Cheers
> Fabio
>
> On 03/25/2013 04:13 PM, Jan Friesse wrote:
>> When serivce plugin calls fork and child uses logsys, it may lead to
>> corrupted fdata (idx is updated but it's not shared between child and
>> parent, but fdata is mmaped file and this is shared).
>>
>> Solutions is to:
>> - Store corosync pid
>> - On log, check if current pid == corosync pid
>> - If so -> store in fdata, if not -> don't store
>>
>> Signed-off-by: Jan Friesse <[email protected]>
>> ---
>> exec/logsys.c | 28 ++++++++++++++++++----------
>> 1 files changed, 18 insertions(+), 10 deletions(-)
>>
>> diff --git a/exec/logsys.c b/exec/logsys.c
>> index c4611bc..cd6a311 100644
>> --- a/exec/logsys.c
>> +++ b/exec/logsys.c
>> @@ -205,6 +205,8 @@ static int logsys_dropped_messages = 0;
>>
>> void *logsys_rec_end;
>>
>> +static pid_t startup_pid = 0;
>> +
>> static DECLARE_LIST_INIT(logsys_print_finished_records);
>>
>> #define FDMAX_ARGS 64
>> @@ -1242,16 +1244,20 @@ void _logsys_log_vprintf (
>> short_file_name++; /* move past the "/" */
>> #endif /* BUILDING_IN_PLACE */
>>
>> - /*
>> - * Create a log record
>> - */
>> - _logsys_log_rec (
>> - rec_ident,
>> - function_name,
>> - short_file_name,
>> - file_line,
>> - logsys_print_buffer, len + 1,
>> - LOGSYS_REC_END);
>> + if (startup_pid == 0 || startup_pid == getpid()) {
>> + /*
>> + * Create a log record if we are really true corosync
>> + * process (not fork of some service) or if we didn't finished
>> + * initialization yet.
>> + */
>> + _logsys_log_rec (
>> + rec_ident,
>> + function_name,
>> + short_file_name,
>> + file_line,
>> + logsys_print_buffer, len + 1,
>> + LOGSYS_REC_END);
>> + }
>>
>> /*
>> * If logsys is not going to print a message to a log target don't
>> @@ -1324,6 +1330,8 @@ int _logsys_config_subsys_get (const char *subsys)
>> void logsys_fork_completed (void)
>> {
>> logsys_loggers[LOGSYS_MAX_SUBSYS_COUNT].mode &= ~LOGSYS_MODE_FORK;
>> + startup_pid = getpid();
>> +
>> (void)_logsys_wthread_create ();
>> }
>>
>>
>
> _______________________________________________
> discuss mailing list
> [email protected]
> http://lists.corosync.org/mailman/listinfo/discuss
_______________________________________________
discuss mailing list
[email protected]
http://lists.corosync.org/mailman/listinfo/discuss