LU-13799 llite: Do not get/put DIO pages

We've already told the kernel we're working with these pages
using the get/put_user_pages functions, and userspace must
hold references on them throughout the i/o anyway.

So getting/putting these vmpages is unnecessary.  This
saves around 7% of the time in DIO page submission, netting
about that much of a performance improvement.

This patch reduces i/o time in ms/GiB by:
Write: 22 ms/GiB
Read: 19 ms/GiB

Totals:
Write: 135 ms/GiB
Read: 143 ms/GiB

mpirun -np 1  $IOR -w -r -t 64M -b 64G -o ./iorfile --posix.odirect

With previous patches in series:
write     6470 MiB/s
read      6354 MiB/s

Plus this patch:
write     7531 MiB/s
read      7179 MiB/s

Signed-off-by: Patrick Farrel <pfarrell@whamcloud.com>
Change-Id: Ic457c21ebca9624da2422463da453b535dcfd10e
Reviewed-on: https://review.whamcloud.com/39438
Reviewed-by: Shaun Tancheff <shaun.tancheff@hpe.com>
Reviewed-by: Andreas Dilger <adilger@whamcloud.com>
Reviewed-by: Yingjin Qian <qian@ddn.com>
Tested-by: jenkins <devops@whamcloud.com>
Tested-by: Maloo <maloo@whamcloud.com>
Reviewed-by: Oleg Drokin <green@whamcloud.com>

diff --git a/lustre/llite/vvp_page.c b/lustre/llite/vvp_page.c
index 3a3bc9e..c841fac 100644 (file)
--- a/lustre/llite/vvp_page.c
+++ b/lustre/llite/vvp_page.c
@@ -52,20 +52,6 @@
  * Page operations.
  *
  */
-
-static void vvp_page_fini_common(struct vvp_page *vpg, struct pagevec *pvec)
-{
-       struct page *vmpage = vpg->vpg_page;
-
-       LASSERT(vmpage != NULL);
-       if (pvec) {
-               if (!pagevec_add(pvec, vmpage))
-                       pagevec_release(pvec);
-       } else {
-               put_page(vmpage);
-       }
-}
-
 static void vvp_page_fini(const struct lu_env *env,
                          struct cl_page_slice *slice,
                          struct pagevec *pvec)
@@ -78,7 +64,13 @@ static void vvp_page_fini(const struct lu_env *env,
         * VPG_FREEING state.
         */
        LASSERT((struct cl_page *)vmpage->private != slice->cpl_page);
-       vvp_page_fini_common(vpg, pvec);
+       LASSERT(vmpage != NULL);
+       if (pvec) {
+               if (!pagevec_add(pvec, vmpage))
+                       pagevec_release(pvec);
+       } else {
+               put_page(vmpage);
+       }
 }
 
 static int vvp_page_own(const struct lu_env *env,
@@ -451,18 +443,8 @@ static int vvp_transient_page_is_vmlocked(const struct lu_env *env,
        return -EBUSY;
 }
 
-static void vvp_transient_page_fini(const struct lu_env *env,
-                                   struct cl_page_slice *slice,
-                                   struct pagevec *pvec)
-{
-       struct vvp_page *vpg = cl2vvp_page(slice);
-
-       vvp_page_fini_common(vpg, pvec);
-}
-
 static const struct cl_page_operations vvp_transient_page_ops = {
        .cpo_discard            = vvp_transient_page_discard,
-       .cpo_fini               = vvp_transient_page_fini,
        .cpo_is_vmlocked        = vvp_transient_page_is_vmlocked,
        .cpo_print              = vvp_page_print,
 };
@@ -476,7 +458,6 @@ int vvp_page_init(const struct lu_env *env, struct cl_object *obj,
        CLOBINVRNT(env, obj, vvp_object_invariant(obj));
 
        vpg->vpg_page = vmpage;
-       get_page(vmpage);
 
        if (page->cp_type == CPT_TRANSIENT) {
                /* DIO pages are referenced by userspace, we don't need to take
@@ -485,6 +466,7 @@ int vvp_page_init(const struct lu_env *env, struct cl_object *obj,
                cl_page_slice_add(page, &vpg->vpg_cl, obj,
                                  &vvp_transient_page_ops);
        } else {
+               get_page(vmpage);
                /* in cache, decref in vvp_page_delete */
                atomic_inc(&page->cp_ref);
                SetPagePrivate(vmpage);