On Fri, Mar 13, 2009 at 12:55 AM, Greg Freemyer <greg.freemyer@xxxxxxxxx> wrote: > On Thu, Mar 12, 2009 at 2:37 PM, Greg Freemyer <greg.freemyer@xxxxxxxxx> wrote: >> On Thu, Mar 12, 2009 at 1:59 PM, Sandeep K Sinha >> <sandeepksinha@xxxxxxxxx> wrote: >>> On Thu, Mar 12, 2009 at 9:43 PM, Greg Freemyer <greg.freemyer@xxxxxxxxx> wrote: >>>> On Thu, Mar 12, 2009 at 11:16 AM, Sandeep K Sinha >>>> <sandeepksinha@xxxxxxxxx> wrote: >>>>> Hi all, >>>>> >>>>> I am listing above two codes for re-allocating new blocks for a given inode. >>>>> >>>>> These code snippets are specific to ext2. >>>>> The problem here is that it is performing slower than cp and dd. >>>>> >>>>> Can anyone help in determining what can be the possible reason(s) ? >>>> >>>> Excessive use of sync_dirty_buffer(); >>>> >>>> You're calling it for every single block you copy. Pretty sure it is a >>>> very aggressive call that forces the data out to the disk immediately, >>>> thus the benefits of caching and elevators are lost. And those are >>>> big benefits. >>>> >>> >>> probably the reason being that the ext2 has some issues with sync. >>> http://kerneltrap.org/index.php?q=mailarchive/linux-fsdevel/2007/2/14/316836/thread >>> >>> Hence, if we don't do this we see a difference in the original and >>> reallocated file. >> >> Sandeep, >> > Sandeep, > > I dropped ext4. > > Seperately, if you want to try to get your flip page approach to work > without sync_dirty_buffer(), I think you need to re-code it as: > > /* Swap data pointers for the data buffers */ > > lock_buffer(src_bhptr); > lock_buffer(dst_bhptr); > > oldpage = dst_bhptr->b_page; > olddata = dst_bhptr->b_data; > dst_bhptr->b_data = src_bhptr->b_data; > dst_bhptr->b_page = src_bhptr->b_page; > src_bhptr->b_data = oldpage; > src_bhptr->b_page = olddata; > flush_dcache_page(src_bhptr->b_page); > flush_dcache_page(dst_bhptr->b_page); > > unlock_buffer(src_bhptr); > unlock_buffer(dst_bhptr); > > /* buffer dirty so it will automatically get written to > disk by io_scheduler */ > mark_buffer_dirty(dst_bhptr); > > brelse(src_bhptr); /* src is not dirty, will be > released immediately */ > brelse(dst_bhptr); /* dst is dirty, will be released > as soon as the data is written to disk and the buffer becomes clean */ > > } /* End of block copy loop */ > > ext2_sync_inode(dest_ind); > > I threw in some comments just for my own sake. > > The above is totally untested, but it looks much better "to me" than > the logic you currently have. Personally I would get your memcpy > approach perfected first. It should be able to run at the same speed > as cp from userspace. > > Then come back and experiment with the above small speed optimization > which might run faster, but more importantly should show a smaller CPU > load since you aren't having to move the data around one extra time. > Will see this once we are clean with memcopy algorithm. Thanks. > Greg > -- > Greg Freemyer > Head of EDD Tape Extraction and Processing team > Litigation Triage Solutions Specialist > http://www.linkedin.com/in/gregfreemyer > First 99 Days Litigation White Paper - > http://www.norcrossgroup.com/forms/whitepapers/99%20Days%20whitepaper.pdf > > The Norcross Group > The Intersection of Evidence & Technology > http://www.norcrossgroup.com > -- Regards, Sandeep. “To learn is to change. Education is a process that changes the learner.” -- To unsubscribe from this list: send an email with "unsubscribe kernelnewbies" to ecartis@xxxxxxxxxxxx Please read the FAQ at http://kernelnewbies.org/FAQ
