Retrieving filename / inode from page fault probe

Retrieving filename / inode from page fault probe

[Yun-Chih Chen]

Hi, friends:

I am benchmarking a NFS server, on which I want to answer this question:

When there is a NFS read request coming in, are the pages of the
requested file in page cache?

My plan is to record timestamp of arriving read request with
"network/nfsd-trace.stp" and record the file that corresponds to every
page-cache-add and page-cache-remove.  Once the data is collected, I
can match read request timeline with that of page cache to see if the
request is a hit or miss.

The following script aims to record the file that an added page belongs to:

probe kernel.trace("mm_filemap_add_to_page_cache") {
 f = inode_path($page->mapping->host)
 printf("%s %s \n", tz_ctime(gettimeofday_s()), f);
}

This will result in a read fault runtime error:
ERROR: read fault [man error::fault] at 0xffffffffffffff78 (addr) near
operator '->' at /usr/share/systemtap/tapset/linux/dentry.stp:65:33

Well, I'm not sure if this is the correct way of achieving the goal
...  It is also not clear to me how the probe differs from Systemtap's
"vm.pagefault" probe.

Anyone has suggest on this?

Thanks,

Yunchih

Re: Retrieving filename / inode from page fault probe

[Yunchih Chen]

Well, I have managed to make it work with the following script:

probe kernel.trace("mm_filemap_add_to_page_cache") {
  if($page->mapping->host->i_ino){
        f = inode_path($page->mapping->host)
        printf("%s Added %s \n", tz_ctime(gettimeofday_s()), f);
  }
}

Some inode turn out to be 0 and causes the read fault!