[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Message-ID: <CAK-6q+iedvC_b3_dO+7C6S15y2o9uuqwDaNJ=a0wZE3hW=+G1g@mail.gmail.com>
Date: Fri, 30 Aug 2024 08:55:31 -0400
From: Alexander Aring <aahringo@...hat.com>
To: teigland@...hat.com
Cc: gfs2@...ts.linux.dev, song@...nel.org, yukuai3@...wei.com,
agruenba@...hat.com, mark@...heh.com, jlbec@...lplan.org,
joseph.qi@...ux.alibaba.com, gregkh@...uxfoundation.org, rafael@...nel.org,
akpm@...ux-foundation.org, linux-kernel@...r.kernel.org,
linux-raid@...r.kernel.org, ocfs2-devel@...ts.linux.dev,
netdev@...r.kernel.org, vvidic@...entin-vidic.from.hr, heming.zhao@...e.com,
lucien.xin@...il.com, paulmck@...nel.org, rcu@...r.kernel.org,
juri.lelli@...hat.com, williams@...hat.com
Subject: Re: [RFC 7/7] rv: add dlm compatible lock state kernel verifier
Hi,
On Tue, Aug 27, 2024 at 2:03 PM Alexander Aring <aahringo@...hat.com> wrote:
...
> + set_holder_state(lk, our_nodeid, mode);
> + rv = check_valid_lock_holders(lk, mode, our_nodeid);
> + if (rv) {
> + /* the whole validation process, this event signals
> + * everything is fine and DLM works correctly there
> + * are no cluster-wide locks that violates DLM locking.
> + */
> + da_handle_event_dlm(lk, with_others_compatible_dlm);
> + } else {
> + /* print all holders of the lock when a invalid lock state is entered */
> + console_lock();
I can't hold this lock in some contexts the ast callback can be called from.
I will drop this lock as I don't care.
It would be nice to use this msg callback from the refactor but then I
somehow need to pass the lk pointer to it.
This however works for me that I know at least which nodes/modes are
incompatible if it hits.
> + pr_info("---\n");
> + pr_info("ls_id %u lkb_id: 0x%08x\n", ls_id, lkb_id);
> + pr_info("holders:\n");
> + list_for_each_entry(hl, &lk->holders, list) {
> + pr_info("\tnodeid: %u mode: %d\n", hl->nodeid,
> + hl->mode);
> + }
> + pr_info("---\n");
> + console_unlock();
> +
> + /* move into an invalid state change, we don't have a edge for that
> + * so we just use event_max_dlm.
> + */
> + da_handle_event_dlm(lk, event_max_dlm);
> + }
> + spin_unlock_bh(&dlm_rv_hash_lock);
> +}
> +
> +/* set the holder to transition state as lock downgrades can issue
> + * grant messages to other nodes we need to ignore if a lock on a
> + * specific node is in state transition. From point of DLM API
> + * the user cannot assume to still hold the lock at this point
> + * anyway.
> + */
> +static void set_holder_transition(uint32_t ls_id, const char *res_name,
> + size_t res_length, uint32_t our_nodeid)
> +{
> + struct dlm_rv_holder *hl;
> + struct dlm_rv_lock *lk;
> +
> + spin_lock_bh(&dlm_rv_hash_lock);
> + lk = lookup_lock(ls_id, res_name, res_length);
> + if (lk) {
> + hl = lookup_holder(lk, our_nodeid);
> + if (hl)
> + hl->mode = STATE_MODE_IN_TRANSITION;
> + }
> + spin_unlock_bh(&dlm_rv_hash_lock);
> +}
> +
> +/* after a lock request got validated it cannot fail */
> +static void handle_dlm_lock_validated(void *data, struct dlm_ls *ls,
> + struct dlm_lkb *lkb,
> + struct dlm_args *args,
> + const char *res_name, size_t res_length)
> +{
> + set_holder_transition(ls->ls_global_id, res_name,
> + res_length, ls->ls_dn->our_node->id);
> +}
> +
> +static void handle_dlm_unlock_validated(void *data, struct dlm_ls *ls,
> + struct dlm_lkb *lkb,
> + struct dlm_args *args)
> +{
we need to ignore unlock(CANCEL) requests.
> + set_holder_transition(ls->ls_global_id,
> + lkb->lkb_resource->res_name,
> + lkb->lkb_resource->res_length,
> + ls->ls_dn->our_node->id);
> +}
> +
- Alex
Powered by blists - more mailing lists