From: Ram Pai <linuxram@us.ibm.com>

I have enclosed a patch that fixes a bunch of performance bugs in the
readahead code.

Below is a brief summary of the problems noticed and the proposed fixes
with some results:
   
 Problem 1:  Readahead code closes the readahead window and goes into
 slowread path, if a file is accessed the first time at an offset
 notequal to zero.  In the case of databases(especially in db2), a file
 may not be accessed at offset 0 the first time though the i/o's are
 sequential.

 Fix to Problem 1:
          min = get_min_readahead(ra);
         orig_next_size = ra-next_size;

 -       if (ra-next_size == 0 && offset == 0) {
 +       if (ra-next_size == 0) {

------------------------------------------------------------------------

 ------------------------------------------------------------------------
 Problem 2: After fixing Problem, the readahead window still does not
 open up the first time, if all the pages requested are already in the
 page cache. This time the window closes because of pagecache hits
 instead of misses. To fix this we put in these changes.

 -               check_ra_success(ra, ra-size, actual, orig_next_size);
 +               if(!first_access) {
 +                       check_ra_success(ra, ra-size, actual, orig_next_size);
 +               }

 ------------------------------------------------------------------------

 ------------------------------------------------------------------------
 Problem 3: In the case of large random reads, the readahead window is
 read in, the moment there is a hit in the active window. And it turns
 out that in most of the cases the readahead window gets scrapped,
 because the next large random read does not even touch any of the pages
 in that readahead window. We fixed this by introducing lazy readahead.
 Basically we wait till the last page in the active window gets a hit.
 And once the last page is hit, the readahead window is then read in.
 This fix gave a tremendous boost in the performance.
 To fix this the changes we put in were:

                 /*
                  * This read request is within the current window.  It is time
                  * to submit I/O for the ahead window while the application is
                  * crunching through the current window.
                  */
 -               if (ra-ahead_start == 0) {
 +               if (ra-ahead_start == 0 && offset == (ra-start + ra-size -1)) {

 ------------------------------------------------------------------------

 ------------------------------------------------------------------------
 Problem 4:
       If the request page does not fall in the active window and is not
the  first page of the read ahead window, we scrap both the active
window and the readahaed window and read in the active window. But it
turns out that we read in a lot of pages in the active window based on
the size of the 'projected readahead window size' (the next_size
variable). And we end up using part of the active window and waste the
remaining. We put  in a fix where we read in just as many pages in the
active window based on the number of pages used in the recent past.
 Again this gave us another big boost in performance and ended up
beating the performance of aio patch on a DSS workload.

 The fix to this is:
                  * ahead window and get some I/O underway for the new
                  * current window.
                  */
 +               if (!first_access && preoffset = ra-start &&
 +                               preoffset < (ra-start + ra-size)) {
 +                       ra-size = preoffset - ra-start + 2;
 +               } else {
 +                       ra-size = ra-next_size;

 ------------------------------------------------------------------------

 ------------------------------------------------------------------------
 Problem 5:
 With all the above fixes there is very low chance that the readahead
 window shall close.  But however if it does, we found that the slow
 read  path is really slow.  Any loss of sequentiality in the slow read
 path is penalized heavily by closing the window back to zero. So we
 fixed this  by decreasing the window size by one anytime we loose
 sequentiality and  increasing in by 1 if we didn't.

                 if (offset != ra-prev_page + 1) {
 -                       ra-size = 0;            /* Not sequential */
 +                       ra-size = ra-size?ra-size-1:0; /*Notsequential */

 ------------------------------------------------------------------------

With the above set of fixes we got about 28% improvement in DSS workload
which is about 5% more than what we got with the suparna's aio patch.

This patch compared equivalent to suparna's aio patch with aio-stress
run.

It fared better than aio patch for large random io.

We are yet to run a bunch of other benchmarks to evaluate this patch.
We would like to get your inputs on this patch and any suggestions you
may have to improve it. I have enclosed a patch with all these changes
along with some changes to the comments that reflect the new behaviour.
NOTE: the above patch reverts suparna's aio patch.