net/smc: separate locks for SMCD and SMCR link group lists
authorUrsula Braun <ubraun@linux.ibm.com>
Wed, 9 Oct 2019 08:07:44 +0000 (10:07 +0200)
committerJakub Kicinski <jakub.kicinski@netronome.com>
Thu, 10 Oct 2019 02:45:44 +0000 (19:45 -0700)
This patch introduces separate locks for the split SMCD and SMCR
link group lists.

Signed-off-by: Ursula Braun <ubraun@linux.ibm.com>
Signed-off-by: Karsten Graul <kgraul@linux.ibm.com>
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
include/net/smc.h
net/smc/smc_core.c
net/smc/smc_ism.c

index c08e8c4..438bb02 100644 (file)
@@ -76,6 +76,7 @@ struct smcd_dev {
        u8 pnetid[SMC_MAX_PNETID_LEN];
        bool pnetid_by_user;
        struct list_head lgr_list;
+       spinlock_t lgr_lock;
 };
 
 struct smcd_dev *smcd_alloc_dev(struct device *parent, const char *name,
index 46d6795..949b191 100644 (file)
@@ -42,6 +42,19 @@ static struct smc_lgr_list smc_lgr_list = {  /* established link groups */
 static void smc_buf_free(struct smc_link_group *lgr, bool is_rmb,
                         struct smc_buf_desc *buf_desc);
 
+/* return head of link group list and its lock for a given link group */
+static inline struct list_head *smc_lgr_list_head(struct smc_link_group *lgr,
+                                                 spinlock_t **lgr_lock)
+{
+       if (lgr->is_smcd) {
+               *lgr_lock = &lgr->smcd->lgr_lock;
+               return &lgr->smcd->lgr_list;
+       }
+
+       *lgr_lock = &smc_lgr_list.lock;
+       return &smc_lgr_list.list;
+}
+
 static void smc_lgr_schedule_free_work(struct smc_link_group *lgr)
 {
        /* client link group creation always follows the server link group
@@ -157,19 +170,21 @@ static void smc_lgr_free_work(struct work_struct *work)
        struct smc_link_group *lgr = container_of(to_delayed_work(work),
                                                  struct smc_link_group,
                                                  free_work);
+       spinlock_t *lgr_lock;
        bool conns;
 
-       spin_lock_bh(&smc_lgr_list.lock);
+       smc_lgr_list_head(lgr, &lgr_lock);
+       spin_lock_bh(lgr_lock);
        read_lock_bh(&lgr->conns_lock);
        conns = RB_EMPTY_ROOT(&lgr->conns_all);
        read_unlock_bh(&lgr->conns_lock);
        if (!conns) { /* number of lgr connections is no longer zero */
-               spin_unlock_bh(&smc_lgr_list.lock);
+               spin_unlock_bh(lgr_lock);
                return;
        }
        if (!list_empty(&lgr->list))
                list_del_init(&lgr->list); /* remove from smc_lgr_list */
-       spin_unlock_bh(&smc_lgr_list.lock);
+       spin_unlock_bh(lgr_lock);
 
        if (!lgr->is_smcd && !lgr->terminating) {
                struct smc_link *lnk = &lgr->lnk[SMC_SINGLE_LINK];
@@ -200,6 +215,7 @@ static int smc_lgr_create(struct smc_sock *smc, struct smc_init_info *ini)
        struct smc_link_group *lgr;
        struct list_head *lgr_list;
        struct smc_link *lnk;
+       spinlock_t *lgr_lock;
        u8 rndvec[3];
        int rc = 0;
        int i;
@@ -235,6 +251,7 @@ static int smc_lgr_create(struct smc_sock *smc, struct smc_init_info *ini)
                lgr->peer_gid = ini->ism_gid;
                lgr->smcd = ini->ism_dev;
                lgr_list = &ini->ism_dev->lgr_list;
+               lgr_lock = &lgr->smcd->lgr_lock;
        } else {
                /* SMC-R specific settings */
                lgr->role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
@@ -248,6 +265,7 @@ static int smc_lgr_create(struct smc_sock *smc, struct smc_init_info *ini)
                lnk->smcibdev = ini->ib_dev;
                lnk->ibport = ini->ib_port;
                lgr_list = &smc_lgr_list.list;
+               lgr_lock = &smc_lgr_list.lock;
                lnk->path_mtu =
                        ini->ib_dev->pattr[ini->ib_port - 1].active_mtu;
                if (!ini->ib_dev->initialized)
@@ -277,9 +295,9 @@ static int smc_lgr_create(struct smc_sock *smc, struct smc_init_info *ini)
                        goto destroy_qp;
        }
        smc->conn.lgr = lgr;
-       spin_lock_bh(&smc_lgr_list.lock);
+       spin_lock_bh(lgr_lock);
        list_add(&lgr->list, lgr_list);
-       spin_unlock_bh(&smc_lgr_list.lock);
+       spin_unlock_bh(lgr_lock);
        return 0;
 
 destroy_qp:
@@ -442,11 +460,15 @@ static void smc_lgr_free(struct smc_link_group *lgr)
 
 void smc_lgr_forget(struct smc_link_group *lgr)
 {
-       spin_lock_bh(&smc_lgr_list.lock);
+       struct list_head *lgr_list;
+       spinlock_t *lgr_lock;
+
+       lgr_list = smc_lgr_list_head(lgr, &lgr_lock);
+       spin_lock_bh(lgr_lock);
        /* do not use this link group for new connections */
-       if (!list_empty(&lgr->list))
-               list_del_init(&lgr->list);
-       spin_unlock_bh(&smc_lgr_list.lock);
+       if (!list_empty(lgr_list))
+               list_del_init(lgr_list);
+       spin_unlock_bh(lgr_lock);
 }
 
 /* terminate linkgroup abnormally */
@@ -487,9 +509,12 @@ static void __smc_lgr_terminate(struct smc_link_group *lgr)
 
 void smc_lgr_terminate(struct smc_link_group *lgr)
 {
-       spin_lock_bh(&smc_lgr_list.lock);
+       spinlock_t *lgr_lock;
+
+       smc_lgr_list_head(lgr, &lgr_lock);
+       spin_lock_bh(lgr_lock);
        __smc_lgr_terminate(lgr);
-       spin_unlock_bh(&smc_lgr_list.lock);
+       spin_unlock_bh(lgr_lock);
 }
 
 /* Called when IB port is terminated */
@@ -514,7 +539,7 @@ void smc_smcd_terminate(struct smcd_dev *dev, u64 peer_gid, unsigned short vlan)
        LIST_HEAD(lgr_free_list);
 
        /* run common cleanup function and build free list */
-       spin_lock_bh(&smc_lgr_list.lock);
+       spin_lock_bh(&dev->lgr_lock);
        list_for_each_entry_safe(lgr, l, &dev->lgr_list, list) {
                if ((!peer_gid || lgr->peer_gid == peer_gid) &&
                    (vlan == VLAN_VID_MASK || lgr->vlan_id == vlan)) {
@@ -522,7 +547,7 @@ void smc_smcd_terminate(struct smcd_dev *dev, u64 peer_gid, unsigned short vlan)
                        list_move(&lgr->list, &lgr_free_list);
                }
        }
-       spin_unlock_bh(&smc_lgr_list.lock);
+       spin_unlock_bh(&dev->lgr_lock);
 
        /* cancel the regular free workers and actually free lgrs */
        list_for_each_entry_safe(lgr, l, &lgr_free_list, list) {
@@ -609,9 +634,11 @@ int smc_conn_create(struct smc_sock *smc, struct smc_init_info *ini)
        struct list_head *lgr_list;
        struct smc_link_group *lgr;
        enum smc_lgr_role role;
+       spinlock_t *lgr_lock;
        int rc = 0;
 
        lgr_list = ini->is_smcd ? &ini->ism_dev->lgr_list : &smc_lgr_list.list;
+       lgr_lock = ini->is_smcd ? &ini->ism_dev->lgr_lock : &smc_lgr_list.lock;
        ini->cln_first_contact = SMC_FIRST_CONTACT;
        role = smc->listen_smc ? SMC_SERV : SMC_CLNT;
        if (role == SMC_CLNT && ini->srv_first_contact)
@@ -619,7 +646,7 @@ int smc_conn_create(struct smc_sock *smc, struct smc_init_info *ini)
                goto create;
 
        /* determine if an existing link group can be reused */
-       spin_lock_bh(&smc_lgr_list.lock);
+       spin_lock_bh(lgr_lock);
        list_for_each_entry(lgr, lgr_list, list) {
                write_lock_bh(&lgr->conns_lock);
                if ((ini->is_smcd ?
@@ -640,7 +667,7 @@ int smc_conn_create(struct smc_sock *smc, struct smc_init_info *ini)
                }
                write_unlock_bh(&lgr->conns_lock);
        }
-       spin_unlock_bh(&smc_lgr_list.lock);
+       spin_unlock_bh(lgr_lock);
 
        if (role == SMC_CLNT && !ini->srv_first_contact &&
            ini->cln_first_contact == SMC_FIRST_CONTACT) {
index 674eb5a..34dc619 100644 (file)
@@ -286,6 +286,7 @@ struct smcd_dev *smcd_alloc_dev(struct device *parent, const char *name,
        smc_pnetid_by_dev_port(parent, 0, smcd->pnetid);
 
        spin_lock_init(&smcd->lock);
+       spin_lock_init(&smcd->lgr_lock);
        INIT_LIST_HEAD(&smcd->vlan);
        INIT_LIST_HEAD(&smcd->lgr_list);
        smcd->event_wq = alloc_ordered_workqueue("ism_evt_wq-%s)",