Revert "LU-14541 llite: Check vmpage in releasepage"

This reverts commit c524079f4f59a39b99467d9868ee4aafdcf033e9,
because it breaks releasepage for Lustre and does not
completely fix the data consistency issue in LU-14541.

Breaking releasepage matters because it prevents direct I/O
from working if there is page cache data present, and
because it causes similar issues with GDS, which must be
able to flush page cache pages before doing I/O.

With patches:
"LU-16160 llite: SIGBUS is possible on a race with page reclaim"/
d9c23a7934747eb19e23470b30806482a1aa60f8
and
"LU-14541 llite: Check for page deletion after fault"/
19678e30147f50f813e72e8216cfb0453fe0ca6e
LU-14541 is fully resolved, so we can revert this patch.

Lustre-change: https://review.whamcloud.com/49654
Lustre-commit: e3cfb688ed7116a57b2c7f89a3e4f28291a0b69f

Signed-off-by: Patrick Farrell <pfarrell@whamcloud.com>
Change-Id: I613bdb4f27161ffc3638d1d8ea38827af5a7bd47
Reviewed-on: https://review.whamcloud.com/c/fs/lustre-release/+/50599
Tested-by: jenkins <devops@whamcloud.com>
Tested-by: Maloo <maloo@whamcloud.com>
Reviewed-by: Qian Yingjin <qian@ddn.com>
Reviewed-by: Oleg Drokin <green@whamcloud.com>

diff --git a/lustre/include/cl_object.h b/lustre/include/cl_object.h
index 7d6d509..e4fb55a 100644 (file)
--- a/lustre/include/cl_object.h
+++ b/lustre/include/cl_object.h
@@ -1070,14 +1070,6 @@ static inline bool __page_in_use(const struct cl_page *page, int refc)
  */
 #define cl_page_in_use_noref(pg) __page_in_use(pg, 0)
 
-/* references: cl_page, page cache, optional + refcount for caller reference
- * (always 0 or 1 currently)
- */
-static inline int vmpage_in_use(struct page *vmpage, int refcount)
-{
-       return (page_count(vmpage) - page_mapcount(vmpage) > 2 + refcount);
-}
-
 /** @} cl_page */
 
 /** \addtogroup cl_lock cl_lock

diff --git a/lustre/llite/rw26.c b/lustre/llite/rw26.c
index 6bb71a5..5e0a900 100644 (file)
--- a/lustre/llite/rw26.c
+++ b/lustre/llite/rw26.c
@@ -168,7 +168,7 @@ static int ll_releasepage(struct page *vmpage, RELEASEPAGE_ARG_TYPE gfp_mask)
 {
        struct lu_env           *env;
        struct cl_object        *obj;
-       struct cl_page          *clpage;
+       struct cl_page          *page;
        struct address_space    *mapping;
        int result = 0;
 
@@ -184,23 +184,16 @@ static int ll_releasepage(struct page *vmpage, RELEASEPAGE_ARG_TYPE gfp_mask)
        if (obj == NULL)
                return 1;
 
-       clpage = cl_vmpage_page(vmpage, obj);
-       if (clpage == NULL)
+       page = cl_vmpage_page(vmpage, obj);
+       if (page == NULL)
                return 1;
 
        env = cl_env_percpu_get();
        LASSERT(!IS_ERR(env));
 
-       /* we must not delete the cl_page if the vmpage is in use, otherwise we
-        * disconnect the vmpage from Lustre while it's still alive(!), which
-        * means we won't find it to discard on lock cancellation.
-        *
-        * References here are: caller + cl_page + page cache.
-        * Any other references are potentially transient and must be ignored.
-        */
-       if (!cl_page_in_use(clpage) && !vmpage_in_use(vmpage, 1)) {
+       if (!cl_page_in_use(page)) {
                result = 1;
-               cl_page_delete(env, clpage);
+               cl_page_delete(env, page);
        }
 
        /* To use percpu env array, the call path can not be rescheduled;
@@ -217,7 +210,7 @@ static int ll_releasepage(struct page *vmpage, RELEASEPAGE_ARG_TYPE gfp_mask)
         * that we won't get into object delete path.
         */
        LASSERT(cl_object_refc(obj) > 1);
-       cl_page_put(env, clpage);
+       cl_page_put(env, page);
 
        cl_env_percpu_put(env);
        return result;

diff --git a/lustre/osc/osc_page.c b/lustre/osc/osc_page.c
index fd4c1f3..fa5d86e 100644 (file)
--- a/lustre/osc/osc_page.c
+++ b/lustre/osc/osc_page.c
@@ -539,13 +539,8 @@ static inline bool lru_page_busy(struct client_obd *cli, struct cl_page *page)
        if (cli->cl_cache->ccc_unstable_check) {
                struct page *vmpage = cl_page_vmpage(page);
 
-               /* this check is racy because the vmpage is not locked, but
-                * that's OK - the code which does the actual page release
-                * checks this again before releasing
-                *
-                * vmpage have two known users: cl_page and VM page cache
-                */
-               if (vmpage_in_use(vmpage, 0))
+               /* vmpage have two known users: cl_page and VM page cache */
+               if (page_count(vmpage) - page_mapcount(vmpage) > 2)
                        return true;
        }
        return false;

diff --git a/lustre/tests/sanity.sh b/lustre/tests/sanity.sh
index 934a333..fec778e 100755 (executable)
--- a/lustre/tests/sanity.sh
+++ b/lustre/tests/sanity.sh
@@ -41,8 +41,8 @@ init_logging
 ALWAYS_EXCEPT="$SANITY_EXCEPT "
 # bug number for skipped test: LU-9693 LU-6493 LU-9693
 ALWAYS_EXCEPT+="               42a     42b     42c "
-# bug number:    LU-8411 LU-14541
-ALWAYS_EXCEPT+=" 407     277 "
+# bug number:    LU-8411
+ALWAYS_EXCEPT+=" 407 "
 
 if $SHARED_KEY; then
        # bug number:    LU-14181 LU-14181
@@ -24371,21 +24371,18 @@ test_398a() { # LU-4198
        $LFS setstripe -c 1 -i 0 $DIR/$tfile
        $LCTL set_param ldlm.namespaces.*.lru_size=clear
 
-       # Disabled: DIO does not push out buffered I/O pages, see LU-12587
        # request a new lock on client
-       #dd if=/dev/zero of=$DIR/$tfile bs=1M count=1
+       dd if=/dev/zero of=$DIR/$tfile bs=1M count=1
 
-       #dd if=/dev/zero of=$DIR/$tfile bs=1M count=1 oflag=direct conv=notrunc
-       #local lock_count=$($LCTL get_param -n \
-       #                  ldlm.namespaces.$imp_name.lru_size)
-       #[[ $lock_count -eq 0 ]] || error "lock should be cancelled by direct IO"
+       dd if=/dev/zero of=$DIR/$tfile bs=1M count=1 oflag=direct conv=notrunc
+       local lock_count=$($LCTL get_param -n \
+                          ldlm.namespaces.$imp_name.lru_size)
+       [[ $lock_count -eq 0 ]] || error "lock should be cancelled by direct IO"
 
        $LCTL set_param ldlm.namespaces.*-OST0000-osc-ffff*.lru_size=clear
 
        # no lock cached, should use lockless DIO and not enqueue new lock
-       dd if=/dev/zero of=$DIR/$tfile bs=1M count=1 oflag=direct \
-               conv=notrunc ||
-               error "dio write failed"
+       dd if=/dev/zero of=$DIR/$tfile bs=1M count=1 oflag=direct conv=notrunc
        lock_count=$($LCTL get_param -n \
                     ldlm.namespaces.$imp_name.lru_size)
        [[ $lock_count -eq 0 ]] || error "no lock should be held by direct IO"