Re: [PATCH 3/10] tmpfs: optimize clearing when writing

[Nick Piggin]

On 12 May 2012 22:04, Hugh Dickins <hughd@xxxxxxxxxx> wrote:
> Nick proposed years ago that tmpfs should avoid clearing its pages where
> write will overwrite them with new data, as ramfs has long done. ÂBut I
> messed it up and just got bad data. ÂTried again recently, it works fine.
>
> Here's time output for writing 4GiB 16 times on this Core i5 laptop:
>
> before: real  Â0m21.169s user Â0m0.028s sys  Â0m21.057s
>    Âreal  Â0m21.382s user Â0m0.016s sys  Â0m21.289s
>    Âreal  Â0m21.311s user Â0m0.020s sys  Â0m21.217s
>
> after: Âreal  Â0m18.273s user Â0m0.032s sys  Â0m18.165s
>    Âreal  Â0m18.354s user Â0m0.020s sys  Â0m18.265s
>    Âreal  Â0m18.440s user Â0m0.032s sys  Â0m18.337s
>
> ramfs: Âreal  Â0m16.860s user Â0m0.028s sys  Â0m16.765s
>    Âreal  Â0m17.382s user Â0m0.040s sys  Â0m17.273s
>    Âreal  Â0m17.133s user Â0m0.044s sys  Â0m17.021s

Cool, thanks Hugh! Very big speedup.


>
> Yes, I have done perf reports, but they need more explanation than they
> deserve: in summary, clear_page vanishes, its cache loading shifts into
> copy_user_generic_unrolled; shmem_getpage_gfp goes down, and surprisingly
> mark_page_accessed goes way up - I think because they are respectively
> where the cache gets to be reloaded after being purged by clear or copy.
>
> Suggested-by: Nick Piggin <npiggin@xxxxxxxxx>
> Signed-off-by: Hugh Dickins <hughd@xxxxxxxxxx>
> ---
> Âmm/shmem.c | Â 20 +++++++++++++++++---
> Â1 file changed, 17 insertions(+), 3 deletions(-)
>
> --- 3045N.orig/mm/shmem.c    2012-05-05 10:46:05.732062006 -0700
> +++ 3045N/mm/shmem.c  Â2012-05-05 10:46:12.316062172 -0700
> @@ -1095,9 +1095,14 @@ repeat:
> Â Â Â Â Â Â Â Âshmem_recalc_inode(inode);
> Â Â Â Â Â Â Â Âspin_unlock(&info->lock);
>
> - Â Â Â Â Â Â Â clear_highpage(page);
> - Â Â Â Â Â Â Â flush_dcache_page(page);
> - Â Â Â Â Â Â Â SetPageUptodate(page);
> + Â Â Â Â Â Â Â /*
> + Â Â Â Â Â Â Â Â* Let SGP_WRITE caller clear ends if write does not fill page
> + Â Â Â Â Â Â Â Â*/
> + Â Â Â Â Â Â Â if (sgp != SGP_WRITE) {
> + Â Â Â Â Â Â Â Â Â Â Â clear_highpage(page);
> + Â Â Â Â Â Â Â Â Â Â Â flush_dcache_page(page);
> + Â Â Â Â Â Â Â Â Â Â Â SetPageUptodate(page);
> + Â Â Â Â Â Â Â }
> Â Â Â Â Â Â Â Âif (sgp == SGP_DIRTY)
> Â Â Â Â Â Â Â Â Â Â Â Âset_page_dirty(page);
> Â Â Â Â}
> @@ -1307,6 +1312,14 @@ shmem_write_end(struct file *file, struc
> Â Â Â Âif (pos + copied > inode->i_size)
> Â Â Â Â Â Â Â Âi_size_write(inode, pos + copied);
>
> + Â Â Â if (!PageUptodate(page)) {
> + Â Â Â Â Â Â Â if (copied < PAGE_CACHE_SIZE) {
> + Â Â Â Â Â Â Â Â Â Â Â unsigned from = pos & (PAGE_CACHE_SIZE - 1);
> + Â Â Â Â Â Â Â Â Â Â Â zero_user_segments(page, 0, from,
> + Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â from + copied, PAGE_CACHE_SIZE);
> + Â Â Â Â Â Â Â }
> + Â Â Â Â Â Â Â SetPageUptodate(page);
> + Â Â Â }
> Â Â Â Âset_page_dirty(page);
> Â Â Â Âunlock_page(page);
> Â Â Â Âpage_cache_release(page);
> @@ -1768,6 +1781,7 @@ static int shmem_symlink(struct inode *d
> Â Â Â Â Â Â Â Âkaddr = kmap_atomic(page);
> Â Â Â Â Â Â Â Âmemcpy(kaddr, symname, len);
> Â Â Â Â Â Â Â Âkunmap_atomic(kaddr);
> + Â Â Â Â Â Â Â SetPageUptodate(page);
> Â Â Â Â Â Â Â Âset_page_dirty(page);
> Â Â Â Â Â Â Â Âunlock_page(page);
> Â Â Â Â Â Â Â Âpage_cache_release(page);
èº{.nÇ+‰·Ÿ®‰­†+%ŠËlzwm…ébëæìr¸›zX§»®w¥Š{ayºÊÚë,j­¢f£¢·hš‹àz¹®w¥¢¸¢·¦j:+v‰¨ŠwèjØm¶Ÿÿ¾«‘êçzZ+ƒùšŽŠÝj"ú!¶iO•æ¬z·švØ^¶m§ÿðÃnÆàþY&—