Thank you Peng for doing this. On Tue, Jan 18, 2022, at 16:01, Peng He wrote: > From hepeng: > https://patchwork.ozlabs.org/project/openvswitch/patch/20200717015041.82746-1-hepeng.0...@bytedance.com/#2487473 > > also from guohongzhi <guohongz...@huawei.com>: > http://patchwork.ozlabs.org/project/openvswitch/patch/20200306130555.19884-1-guohongz...@huawei.com/ > > also from a discussion about the mixing use of RCU and refcount in the mail > list with Ilya Maximets, William Tu, Ben Pfaf, and Gaëtan Rivet. > > A summary, as quoted from Ilya: > > " > RCU for ofproto was introduced for one > and only one reason - to avoid freeing ofproto while rules are still > alive. This was done in commit f416c8d61601 ("ofproto: RCU postpone > rule destruction."). The goal was to allow using rules without > refcounting them within a single grace period. And that forced us > to postpone destruction of the ofproto for a single grace period. > Later commit 39c9459355b6 ("Use classifier versioning.") made it > possible for rules to be alive for more than one grace period, so > the commit made ofproto wait for 2 grace periods by double postponing. > As we can see now, that wasn't enough and we have to wait for more > than 2 grace periods in certain cases. > " > > In a short, the ofproto should have a longer life time than rule, if > the rule lasts for more than 2 grace periods, the ofproto should live > longer to ensure rule->ofproto is valid. It's hard to predict how long > a ofproto should live, thus we need to use refcount on ofproto to make > things easy. The controversial part is that we have already used RCU postpone > to delay ofproto destrution, if we have to add refcount, is it simpler to > use just refcount without RCU postpone? > > IMO, I think going back to the pure refcount solution is more > complicated than mixing using both. > > Gaëtan Rive asks some questions on guohongzhi's v2 patch: > > during ofproto_rule_create, should we use ofproto_ref > or ofproto_try_ref? how can we make sure the ofproto is alive? > > By using RCU, ofproto has three states: > > state 1: alive, with refcount >= 1 > state 2: dying, with refcount == 0, however pointer is valid > state 3: died, memory freed, pointer might be dangling. > > Without using RCU, there is no state 2, thus, we have to be very careful > every time we see a ofproto pointer. In contrast, with RCU, we can be sure > that it's alive at least in this grace peroid, so we can just check if > it is dying by ofproto_try_ref. > > This shows that by mixing use of RCU and refcount we can save a lot of work > worrying if ofproto is dangling. > > In short, the RCU part makes sure the ofproto is alive when we use it, > and the refcount part makes sure it lives longer enough. > > Also regarding a new patch filed recently, people are now making use > of RCU to protect ofproto: > > https://patchwork.ozlabs.org/project/openvswitch/patch/1638530715-44436-1-git-send-email-wangyunj...@huawei.com/ > > In this patch, I have merged guohongzhi's patch and mine, and fixes > accoring to the previous comments. > > Signed-off-by: Peng He <hepeng.0...@bytedance.com> > Signed-off-by: guohongzhi <guohongz...@huawei.com>
I don't know if keeping the history of the discussion is relevant on the git tree, but it was useful to get back the context. Maybe it could be made more concise before integration. The comments made in previous versions were addressed. I have been running an ofproto + rule deletion test for a few hours. By that point, usually I would see the crash happening. With this patch and https://patchwork.ozlabs.org/project/openvswitch/patch/1638530715-44436-1-git-send-email-wangyunj...@huawei.com/ applied, I do not see this crash anymore. I have a few comments below. > --- > ofproto/ofproto-dpif-xlate-cache.c | 2 + > ofproto/ofproto-dpif-xlate.c | 14 ++++--- > ofproto/ofproto-dpif.c | 24 +++++++----- > ofproto/ofproto-provider.h | 2 + > ofproto/ofproto.c | 62 +++++++++++++++++++++++++++--- > ofproto/ofproto.h | 4 ++ > 6 files changed, 87 insertions(+), 21 deletions(-) > > diff --git a/ofproto/ofproto-dpif-xlate-cache.c > b/ofproto/ofproto-dpif-xlate-cache.c > index dcc91cb38..9224ee2e6 100644 > --- a/ofproto/ofproto-dpif-xlate-cache.c > +++ b/ofproto/ofproto-dpif-xlate-cache.c > @@ -209,6 +209,7 @@ xlate_cache_clear_entry(struct xc_entry *entry) > { > switch (entry->type) { > case XC_TABLE: > + ofproto_unref(&(entry->table.ofproto->up)); > break; > case XC_RULE: > ofproto_rule_unref(&entry->rule->up); > @@ -231,6 +232,7 @@ xlate_cache_clear_entry(struct xc_entry *entry) > free(entry->learn.ofm); > break; > case XC_NORMAL: > + ofproto_unref(&(entry->normal.ofproto->up)); > break; > case XC_FIN_TIMEOUT: > /* 'u.fin.rule' is always already held as a XC_RULE, which > diff --git a/ofproto/ofproto-dpif-xlate.c b/ofproto/ofproto-dpif-xlate.c > index 6fb59e170..129cdf714 100644 > --- a/ofproto/ofproto-dpif-xlate.c > +++ b/ofproto/ofproto-dpif-xlate.c > @@ -3024,12 +3024,14 @@ xlate_normal(struct xlate_ctx *ctx) > struct xc_entry *entry; > > /* Save just enough info to update mac learning table later. */ > - entry = xlate_cache_add_entry(ctx->xin->xcache, XC_NORMAL); > - entry->normal.ofproto = ctx->xbridge->ofproto; > - entry->normal.in_port = flow->in_port.ofp_port; > - entry->normal.dl_src = flow->dl_src; > - entry->normal.vlan = vlan; > - entry->normal.is_gratuitous_arp = is_grat_arp; > + if (ofproto_try_ref(&ctx->xbridge->ofproto->up)) { > + entry = xlate_cache_add_entry(ctx->xin->xcache, XC_NORMAL); > + entry->normal.ofproto = ctx->xbridge->ofproto; > + entry->normal.in_port = flow->in_port.ofp_port; > + entry->normal.dl_src = flow->dl_src; > + entry->normal.vlan = vlan; > + entry->normal.is_gratuitous_arp = is_grat_arp; > + } At that point, if try_ref() failed, ofproto is not valid anymore. I am not sure it is worth reporting using xlate_report() like previous warnings, as the user might be aware already of the ofproto destruction. But shouldn't we just get out of the function at that point? Further processing is meaningless I think? > } > > /* Determine output bundle. */ > diff --git a/ofproto/ofproto-dpif.c b/ofproto/ofproto-dpif.c > index 8143dd965..c0a87456a 100644 > --- a/ofproto/ofproto-dpif.c > +++ b/ofproto/ofproto-dpif.c > @@ -4471,12 +4471,14 @@ rule_dpif_lookup_from_table(struct ofproto_dpif > *ofproto, > atomic_add_relaxed(&tbl->n_matched, stats->n_packets, > &orig); > } > if (xcache) { > - struct xc_entry *entry; > + if (ofproto_try_ref(&ofproto->up)) { > + struct xc_entry *entry; > > - entry = xlate_cache_add_entry(xcache, XC_TABLE); > - entry->table.ofproto = ofproto; > - entry->table.id = *table_id; > - entry->table.match = true; > + entry = xlate_cache_add_entry(xcache, XC_TABLE); > + entry->table.ofproto = ofproto; > + entry->table.id = *table_id; > + entry->table.match = true; > + } > } > return rule; > } > @@ -4507,12 +4509,14 @@ rule_dpif_lookup_from_table(struct ofproto_dpif > *ofproto, > stats->n_packets, &orig); > } > if (xcache) { > - struct xc_entry *entry; > + if (ofproto_try_ref(&ofproto->up)) { > + struct xc_entry *entry; > > - entry = xlate_cache_add_entry(xcache, XC_TABLE); > - entry->table.ofproto = ofproto; > - entry->table.id = next_id; > - entry->table.match = (rule != NULL); > + entry = xlate_cache_add_entry(xcache, XC_TABLE); > + entry->table.ofproto = ofproto; > + entry->table.id = next_id; > + entry->table.match = (rule != NULL); > + } > } > if (rule) { > goto out; /* Match. */ > diff --git a/ofproto/ofproto-provider.h b/ofproto/ofproto-provider.h > index 14b909973..ed10b8c76 100644 > --- a/ofproto/ofproto-provider.h > +++ b/ofproto/ofproto-provider.h > @@ -143,6 +143,8 @@ struct ofproto { > /* Variable length mf_field mapping. Stores all configured variable > length > * meta-flow fields (struct mf_field) in a switch. */ > struct vl_mff_map vl_mff_map; > + /* refcount to this ofproto, holds by rule/group/xlate_caches */ > + struct ovs_refcount refcount; > }; > > void ofproto_init_tables(struct ofproto *, int n_tables); > diff --git a/ofproto/ofproto.c b/ofproto/ofproto.c > index 56aeac720..10a22d9ec 100644 > --- a/ofproto/ofproto.c > +++ b/ofproto/ofproto.c > @@ -549,6 +549,7 @@ ofproto_create(const char *datapath_name, const > char *datapath_type, > > ovs_mutex_init(&ofproto->vl_mff_map.mutex); > cmap_init(&ofproto->vl_mff_map.cmap); > + ovs_refcount_init(&ofproto->refcount); > > error = ofproto->ofproto_class->construct(ofproto); > if (error) { > @@ -1695,9 +1696,24 @@ ofproto_destroy__(struct ofproto *ofproto) > ofproto->ofproto_class->dealloc(ofproto); > } > > -/* Destroying rules is doubly deferred, must have 'ofproto' around for them. > - * - 1st we defer the removal of the rules from the classifier > - * - 2nd we defer the actual destruction of the rules. */ > +/* We used to use defer function to wait for two grace periods > + * as we assume the rule that holds the ofproto pointer will > + * live at most two grace period. Howvever, we found at certain > + * cases, this assumption does not stand. > + * > + * destroying a rule may have to wait multiple grace periods: > + * remove_rules_postponed (one grace period) > + * -> remove_rule_rcu > + * -> remove_rule_rcu__ > + * -> ofproto_rule_unref -> ref count != 1 > + * -> ... more grace periods. > + * -> rule_destroy_cb (> 2 grace periods) > + * -> free > + * > + * So we have to check the refcount for sure all the rules > + * have been destroyed. > + * > + */ Referring to previous erroneous ideas of what had to be done is only misleading as a comment on this code. Now that the trigger for actual deletion is automatically, properly handled such comment is not useful anymore. > static void > ofproto_destroy_defer__(struct ofproto *ofproto) > OVS_EXCLUDED(ofproto_mutex) > @@ -1705,6 +1721,26 @@ ofproto_destroy_defer__(struct ofproto *ofproto) > ovsrcu_postpone(ofproto_destroy__, ofproto); > } > > +void > +ofproto_ref(struct ofproto *ofproto) > +{ > + ovs_refcount_ref(&ofproto->refcount); > +} > + > +bool > +ofproto_try_ref(struct ofproto *ofproto) > +{ > + return ovs_refcount_try_ref_rcu(&ofproto->refcount); > +} > + > +void > +ofproto_unref(struct ofproto *ofproto) > +{ > + if (ofproto && ovs_refcount_unref(&ofproto->refcount) == 1) { > + ovsrcu_postpone(ofproto_destroy_defer__, ofproto); > + } > +} > + > void > ofproto_destroy(struct ofproto *p, bool del) > OVS_EXCLUDED(ofproto_mutex) > @@ -1736,8 +1772,7 @@ ofproto_destroy(struct ofproto *p, bool del) > p->connmgr = NULL; > ovs_mutex_unlock(&ofproto_mutex); > > - /* Destroying rules is deferred, must have 'ofproto' around for them. */ > - ovsrcu_postpone(ofproto_destroy_defer__, p); > + ofproto_unref(p); > } > > /* Destroys the datapath with the respective 'name' and 'type'. With the > Linux > @@ -2929,6 +2964,10 @@ ofproto_rule_destroy__(struct rule *rule) > cls_rule_destroy(CONST_CAST(struct cls_rule *, &rule->cr)); > rule_actions_destroy(rule_get_actions(rule)); > ovs_mutex_destroy(&rule->mutex); > + /* we need to call ofproto_unref first, and thanks to rcu, ofproto is > alive > + * otherwise, group is freed, group->ofproto is invalid > + */ > + ofproto_unref(rule->ofproto); > rule->ofproto->ofproto_class->rule_dealloc(rule); > } > > @@ -3069,6 +3108,10 @@ group_destroy_cb(struct ofgroup *group) > &group->props)); > ofputil_bucket_list_destroy(CONST_CAST(struct ovs_list *, > &group->buckets)); > + /* we need to call ofproto_unref first, and thanks to rcu, ofproto is > alive > + * otherwise, group is freed, group->ofproto is invalid > + */ > + ofproto_unref(group->ofproto); > group->ofproto->ofproto_class->group_dealloc(group); > } > > @@ -5279,6 +5322,11 @@ ofproto_rule_create(struct ofproto *ofproto, > struct cls_rule *cr, > return OFPERR_OFPFMFC_UNKNOWN; > } > > + if (!ofproto_try_ref(ofproto)) { > + cls_rule_destroy(cr); > + return OFPERR_OFPFMFC_UNKNOWN; > + } > + If the ofproto referencing was done before rule allocation, it would not be necessary to destroy the rule there. > /* Initialize base state. */ > *CONST_CAST(struct ofproto **, &rule->ofproto) = ofproto; > cls_rule_move(CONST_CAST(struct cls_rule *, &rule->cr), cr); > @@ -7345,6 +7393,10 @@ init_group(struct ofproto *ofproto, const struct > ofputil_group_mod *gm, > return OFPERR_OFPGMFC_OUT_OF_GROUPS; > } > > + if (!ofproto_try_ref(ofproto)) { > + return OFPERR_OFPFMFC_UNKNOWN; > + } > + A few lines above this 'return': 7390 *ofgroup = ofproto->ofproto_class->group_alloc(); 7391 if (!*ofgroup) { 7392 VLOG_WARN_RL(&rl, "%s: failed to allocate group", ofproto->name); 7393 return OFPERR_OFPGMFC_OUT_OF_GROUPS; 7394 } So there is a resource leak here, the group should be deallocated. Alternatively, the ofproto try_ref could just be put before the allocation, simplifying the cancelling. Best regards, -- Gaetan Rivet _______________________________________________ dev mailing list d...@openvswitch.org https://mail.openvswitch.org/mailman/listinfo/ovs-dev
