The patch titled Subject: ocfs2: do not set OCFS2_LOCK_UPCONVERT_FINISHING if nonblocking lock can not be granted at once has been removed from the -mm tree. Its filename was ocfs2-do-not-set-ocfs2_lock_upconvert_finishing-if-nonblocking-lock-can-not-be-granted-at-once.patch This patch was dropped because it was merged into mainline or a subsystem tree ------------------------------------------------------ From: Xue jiufei <xuejiufei@xxxxxxxxxx> Subject: ocfs2: do not set OCFS2_LOCK_UPCONVERT_FINISHING if nonblocking lock can not be granted at once ocfs2_readpages() use nonblocking flag to avoid page lock inversion. It will trigger cluster hang because that flag OCFS2_LOCK_UPCONVERT_FINISHING is not cleared if nonblocking lock cannot be granted at once. The flag would prevent dc thread from downconverting. So other nodes cannot acheive this lockres for ever. So we should not set OCFS2_LOCK_UPCONVERT_FINISHING when receiving ast if nonblocking lock had already returned. Signed-off-by: joyce.xue <xuejiufei@xxxxxxxxxx> Reviewed-by: Junxiao Bi <junxiao.bi@xxxxxxxxxx> Cc: Mark Fasheh <mfasheh@xxxxxxxx> Cc: Joel Becker <jlbec@xxxxxxxxxxxx> Signed-off-by: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx> --- fs/ocfs2/dlmglue.c | 37 +++++++++++++++++++++++++++++++------ fs/ocfs2/ocfs2.h | 6 ++++++ 2 files changed, 37 insertions(+), 6 deletions(-) diff -puN fs/ocfs2/dlmglue.c~ocfs2-do-not-set-ocfs2_lock_upconvert_finishing-if-nonblocking-lock-can-not-be-granted-at-once fs/ocfs2/dlmglue.c --- a/fs/ocfs2/dlmglue.c~ocfs2-do-not-set-ocfs2_lock_upconvert_finishing-if-nonblocking-lock-can-not-be-granted-at-once +++ a/fs/ocfs2/dlmglue.c @@ -861,8 +861,13 @@ static inline void ocfs2_generic_handle_ * We set the OCFS2_LOCK_UPCONVERT_FINISHING flag before clearing * the OCFS2_LOCK_BUSY flag to prevent the dc thread from * downconverting the lock before the upconvert has fully completed. + * Do not prevent the dc thread from downconverting if NONBLOCK lock + * had already returned. */ - lockres_or_flags(lockres, OCFS2_LOCK_UPCONVERT_FINISHING); + if (!(lockres->l_flags & OCFS2_LOCK_NONBLOCK_FINISHED)) + lockres_or_flags(lockres, OCFS2_LOCK_UPCONVERT_FINISHING); + else + lockres_clear_flags(lockres, OCFS2_LOCK_NONBLOCK_FINISHED); lockres_clear_flags(lockres, OCFS2_LOCK_BUSY); } @@ -1324,13 +1329,12 @@ static void lockres_add_mask_waiter(stru /* returns 0 if the mw that was removed was already satisfied, -EBUSY * if the mask still hadn't reached its goal */ -static int lockres_remove_mask_waiter(struct ocfs2_lock_res *lockres, +static int __lockres_remove_mask_waiter(struct ocfs2_lock_res *lockres, struct ocfs2_mask_waiter *mw) { - unsigned long flags; int ret = 0; - spin_lock_irqsave(&lockres->l_lock, flags); + assert_spin_locked(&lockres->l_lock); if (!list_empty(&mw->mw_item)) { if ((lockres->l_flags & mw->mw_mask) != mw->mw_goal) ret = -EBUSY; @@ -1338,6 +1342,18 @@ static int lockres_remove_mask_waiter(st list_del_init(&mw->mw_item); init_completion(&mw->mw_complete); } + + return ret; +} + +static int lockres_remove_mask_waiter(struct ocfs2_lock_res *lockres, + struct ocfs2_mask_waiter *mw) +{ + unsigned long flags; + int ret = 0; + + spin_lock_irqsave(&lockres->l_lock, flags); + ret = __lockres_remove_mask_waiter(lockres, mw); spin_unlock_irqrestore(&lockres->l_lock, flags); return ret; @@ -1373,6 +1389,7 @@ static int __ocfs2_cluster_lock(struct o unsigned long flags; unsigned int gen; int noqueue_attempted = 0; + int dlm_locked = 0; ocfs2_init_mask_waiter(&mw); @@ -1481,6 +1498,7 @@ again: ocfs2_recover_from_dlm_error(lockres, 1); goto out; } + dlm_locked = 1; mlog(0, "lock %s, successful return from ocfs2_dlm_lock\n", lockres->l_name); @@ -1514,10 +1532,17 @@ out: if (wait && arg_flags & OCFS2_LOCK_NONBLOCK && mw.mw_mask & (OCFS2_LOCK_BUSY|OCFS2_LOCK_BLOCKED)) { wait = 0; - if (lockres_remove_mask_waiter(lockres, &mw)) + spin_lock_irqsave(&lockres->l_lock, flags); + if (__lockres_remove_mask_waiter(lockres, &mw)) { + if (dlm_locked) + lockres_or_flags(lockres, + OCFS2_LOCK_NONBLOCK_FINISHED); + spin_unlock_irqrestore(&lockres->l_lock, flags); ret = -EAGAIN; - else + } else { + spin_unlock_irqrestore(&lockres->l_lock, flags); goto again; + } } if (wait) { ret = ocfs2_wait_for_mask(&mw); diff -puN fs/ocfs2/ocfs2.h~ocfs2-do-not-set-ocfs2_lock_upconvert_finishing-if-nonblocking-lock-can-not-be-granted-at-once fs/ocfs2/ocfs2.h --- a/fs/ocfs2/ocfs2.h~ocfs2-do-not-set-ocfs2_lock_upconvert_finishing-if-nonblocking-lock-can-not-be-granted-at-once +++ a/fs/ocfs2/ocfs2.h @@ -144,6 +144,12 @@ enum ocfs2_unlock_action { * before the upconvert * has completed */ +#define OCFS2_LOCK_NONBLOCK_FINISHED (0x00001000) /* NONBLOCK cluster + * lock has already + * returned, do not block + * dc thread from + * downconverting */ + struct ocfs2_lock_res_ops; typedef void (*ocfs2_lock_callback)(int status, unsigned long data); _ Patches currently in -mm which might be from xuejiufei@xxxxxxxxxx are origin.patch ocfs2-free-inode-when-i_count-becomes-zero.patch ocfs2-dlm-fix-race-between-dispatched_work-and-dlm_lockres_grab_inflight_worker.patch ocfs2-implement-ocfs2_direct_io_write-fix.patch -- To unsubscribe from this list: send the line "unsubscribe mm-commits" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html