From c53309b912fdd652eb3c2de44626aac4b0fc9bf9 Mon Sep 17 00:00:00 2001 From: David Teigland Date: Fri, 15 Mar 2024 17:11:39 -0500 Subject: dlm: add comments about forced waiters reset When a lock is waiting for a reply for a remote operation, and recovery interrupts this "waiters" state, the remote operation is voided by the recovery, and no reply will be processed. The lkb waiters state for the remote operation is forcibly reset/cleared, so that the lock operation can be restarted after recovery. Improve the comments describing this. Signed-off-by: David Teigland --- fs/dlm/lock.c | 78 ++++++++++++++++++++++++++++++++++++++++++++--------------- 1 file changed, 58 insertions(+), 20 deletions(-) (limited to 'fs/dlm') diff --git a/fs/dlm/lock.c b/fs/dlm/lock.c index 6dca5cdbbc5d..fd752dd03896 100644 --- a/fs/dlm/lock.c +++ b/fs/dlm/lock.c @@ -5014,21 +5014,32 @@ static struct dlm_lkb *find_resend_waiter(struct dlm_ls *ls) return lkb; } -/* Deal with lookups and lkb's marked RESEND from _pre. We may now be the - master or dir-node for r. Processing the lkb may result in it being placed - back on waiters. */ - -/* We do this after normal locking has been enabled and any saved messages - (in requestqueue) have been processed. We should be confident that at - this point we won't get or process a reply to any of these waiting - operations. But, new ops may be coming in on the rsbs/locks here from - userspace or remotely. */ - -/* there may have been an overlap unlock/cancel prior to recovery or after - recovery. if before, the lkb may still have a pos wait_count; if after, the - overlap flag would just have been set and nothing new sent. we can be - confident here than any replies to either the initial op or overlap ops - prior to recovery have been received. */ +/* + * Forced state reset for locks that were in the middle of remote operations + * when recovery happened (i.e. lkbs that were on the waiters list, waiting + * for a reply from a remote operation.) The lkbs remaining on the waiters + * list need to be reevaluated; some may need resending to a different node + * than previously, and some may now need local handling rather than remote. + * + * First, the lkb state for the voided remote operation is forcibly reset, + * equivalent to what remove_from_waiters() would normally do: + * . lkb removed from ls_waiters list + * . lkb wait_type cleared + * . lkb waiters_count cleared + * . lkb ref count decremented for each waiters_count (almost always 1, + * but possibly 2 in case of cancel/unlock overlapping, which means + * two remote replies were being expected for the lkb.) + * + * Second, the lkb is reprocessed like an original operation would be, + * by passing it to _request_lock or _convert_lock, which will either + * process the lkb operation locally, or send it to a remote node again + * and put the lkb back onto the waiters list. + * + * When reprocessing the lkb, we may find that it's flagged for an overlapping + * force-unlock or cancel, either from before recovery began, or after recovery + * finished. If this is the case, the unlock/cancel is done directly, and the + * original operation is not initiated again (no _request_lock/_convert_lock.) + */ int dlm_recover_waiters_post(struct dlm_ls *ls) { @@ -5043,6 +5054,11 @@ int dlm_recover_waiters_post(struct dlm_ls *ls) break; } + /* + * Find an lkb from the waiters list that's been affected by + * recovery node changes, and needs to be reprocessed. Does + * hold_lkb(), adding a refcount. + */ lkb = find_resend_waiter(ls); if (!lkb) break; @@ -5051,6 +5067,11 @@ int dlm_recover_waiters_post(struct dlm_ls *ls) hold_rsb(r); lock_rsb(r); + /* + * If the lkb has been flagged for a force unlock or cancel, + * then the reprocessing below will be replaced by just doing + * the unlock/cancel directly. + */ mstype = lkb->lkb_wait_type; oc = test_and_clear_bit(DLM_IFL_OVERLAP_CANCEL_BIT, &lkb->lkb_iflags); @@ -5064,23 +5085,40 @@ int dlm_recover_waiters_post(struct dlm_ls *ls) r->res_nodeid, lkb->lkb_nodeid, lkb->lkb_wait_nodeid, dlm_dir_nodeid(r), oc, ou); - /* At this point we assume that we won't get a reply to any - previous op or overlap op on this lock. First, do a big - remove_from_waiters() for all previous ops. */ + /* + * No reply to the pre-recovery operation will now be received, + * so a forced equivalent of remove_from_waiters() is needed to + * reset the waiters state that was in place before recovery. + */ clear_bit(DLM_IFL_RESEND_BIT, &lkb->lkb_iflags); + + /* Forcibly clear wait_type */ lkb->lkb_wait_type = 0; - /* drop all wait_count references we still - * hold a reference for this iteration. + + /* + * Forcibly reset wait_count and associated refcount. The + * wait_count will almost always be 1, but in case of an + * overlapping unlock/cancel it could be 2: see where + * add_to_waiters() finds the lkb is already on the waiters + * list and does lkb_wait_count++; hold_lkb(). */ while (lkb->lkb_wait_count) { lkb->lkb_wait_count--; unhold_lkb(lkb); } + + /* Forcibly remove from waiters list */ mutex_lock(&ls->ls_waiters_mutex); list_del_init(&lkb->lkb_wait_reply); mutex_unlock(&ls->ls_waiters_mutex); + /* + * The lkb is now clear of all prior waiters state and can be + * processed locally, or sent to remote node again, or directly + * cancelled/unlocked. + */ + if (oc || ou) { /* do an unlock or cancel instead of resending */ switch (mstype) { -- cgit v1.2.3-58-ga151