How to capture L1 cache stores using perf

How to capture L1 cache stores using perf

[Joy Arulraj]

Hi,

 I am using Perf to capture L1 cache STORE and LOAD events on an Intel Nehalem
machine. In the example below from Apache, I am trying to capture the store to
the shared variable buf.  An incorrect interleaving should leave the buf->outcnt
vairable in invalid state before it is written to, leading to corrupt Apache
logs. For reference, the bug report is available here :
https://issues.apache.org/bugzilla/show_bug.cgi?id=25520

 I am able to capture a LOAD  event of shared variable buf->outcnt  quite
frequently [buf->outcnt = buf->outcnt + len ] , but I am never able to see the
following STORE event performed after adding len to it.
I am sampling every 10th occurrence of the event ( -c 10 ) and I run perf
separately for LOAD and STORE events.

static apr_status_t ap_buffered_log_writer(request_rec *r, void *handle,  const
char **strs, int *strl, int nelts, apr_size_t len){
    char *str;
    char *s;
    int i;
    apr_status_t rv;
    buffered_log *buf = (buffered_log*)handle;

    if (len + buf->outcnt > LOG_BUFSIZE) {
            flush_log(buf);
    }
    if (len >= LOG_BUFSIZE) {
            apr_size_t w;

            str = apr_palloc(r->pool, len + 1);
            for (i = 0, s = str; i < nelts; ++i) {
                    memcpy(s, strs[i], strl[i]);
                    s += strl[i];
            }

   }
    else {
            for (i = 0, s = &buf->outbuf[buf->outcnt]; i < nelts; ++i) {
                    memcpy(s, strs[i], strl[i]);
                    s += strl[i];
            }

            buf->outcnt += tmp;  // ------  buf is a shared variable
            //for(j=0 ; j<1000 ; j++);           

            rv = APR_SUCCESS;
    }

    return rv;
}

 So, my question is :

 (a) Is there any way to increase the possibility of capturing the STORE event
of buf->outcnt  [ by adding some code ] ?

 (b) I am not able to see the ap_buffered_log_write function in the perf log
when I increase sampling rate to -c 5 or more. I guess this because of frequent
page faults or due to highly frequent interrupts. I observed that -c 10 is good
for this example. 

Currently I use perf to sample all events in a script that sets up apache, wget
files and then stops apache. This involves sampling across the entire script.

once.sh :
killall -9 httpd  &> /dev/null
./install/bin/apachectl -k start
./get_files 8
./install/bin/apachectl -k stop
file -b ./install/logs/access_log

Then,
 perf record -R -e r841:u,r140:u -c 10 -a bash once.sh 

I only need to monitor httpd logging process. How can I get more fine grained
control over the specific logging process rather than running perf on the entire
Apache script ?


Thanks.
Joy