drbd: Fix locking across all resources
authorAndreas Gruenbacher <agruen@linbit.com>
Thu, 14 Aug 2014 16:33:30 +0000 (18:33 +0200)
committerJens Axboe <axboe@fb.com>
Wed, 25 Nov 2015 16:22:00 +0000 (09:22 -0700)
Instead of using a rwlock for synchronizing state changes across
resources, take the request locks of all resources for global state
changes.  Use resources_mutex to serialize global state changes.

This means that taking the request lock of a resource is now enough to
prevent changes of that resource.  (Previously, a read lock on the
global state lock was needed as well.)

Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
drivers/block/drbd/drbd_int.h
drivers/block/drbd/drbd_main.c
drivers/block/drbd/drbd_nl.c
drivers/block/drbd/drbd_state.c
drivers/block/drbd/drbd_state.h
drivers/block/drbd/drbd_worker.c

index 47d4b02103b84c17a83596640b07d35c90faf999..2c9ee223d54838a1fcd38535afc27e94b51c46d2 100644 (file)
@@ -292,6 +292,9 @@ struct drbd_device_work {
 
 extern int drbd_wait_misc(struct drbd_device *, struct drbd_interval *);
 
+extern void lock_all_resources(void);
+extern void unlock_all_resources(void);
+
 struct drbd_request {
        struct drbd_work w;
        struct drbd_device *device;
@@ -1418,7 +1421,7 @@ extern struct bio_set *drbd_md_io_bio_set;
 /* to allocate from that set */
 extern struct bio *bio_alloc_drbd(gfp_t gfp_mask);
 
-extern rwlock_t global_state_lock;
+extern struct mutex resources_mutex;
 
 extern int conn_lowest_minor(struct drbd_connection *connection);
 extern enum drbd_ret_code drbd_create_device(struct drbd_config_context *adm_ctx, unsigned int minor);
@@ -1688,19 +1691,6 @@ static inline int drbd_peer_req_has_active_page(struct drbd_peer_request *peer_r
        return 0;
 }
 
-static inline enum drbd_state_rv
-_drbd_set_state(struct drbd_device *device, union drbd_state ns,
-               enum chg_state_flags flags, struct completion *done)
-{
-       enum drbd_state_rv rv;
-
-       read_lock(&global_state_lock);
-       rv = __drbd_set_state(device, ns, flags, done);
-       read_unlock(&global_state_lock);
-
-       return rv;
-}
-
 static inline union drbd_state drbd_read_state(struct drbd_device *device)
 {
        struct drbd_resource *resource = device->resource;
index 3ee4a44cb225a750b01553613aa693dcfe60ab30..f66294db3b0806f1c2f358fad17ee2c096a3bbc8 100644 (file)
@@ -117,6 +117,7 @@ module_param_string(usermode_helper, usermode_helper, sizeof(usermode_helper), 0
  */
 struct idr drbd_devices;
 struct list_head drbd_resources;
+struct mutex resources_mutex;
 
 struct kmem_cache *drbd_request_cache;
 struct kmem_cache *drbd_ee_cache;      /* peer requests */
@@ -2923,7 +2924,7 @@ static int __init drbd_init(void)
        drbd_proc = NULL; /* play safe for drbd_cleanup */
        idr_init(&drbd_devices);
 
-       rwlock_init(&global_state_lock);
+       mutex_init(&resources_mutex);
        INIT_LIST_HEAD(&drbd_resources);
 
        err = drbd_genl_register();
@@ -3746,6 +3747,27 @@ int drbd_wait_misc(struct drbd_device *device, struct drbd_interval *i)
        return 0;
 }
 
+void lock_all_resources(void)
+{
+       struct drbd_resource *resource;
+       int __maybe_unused i = 0;
+
+       mutex_lock(&resources_mutex);
+       local_irq_disable();
+       for_each_resource(resource, &drbd_resources)
+               spin_lock_nested(&resource->req_lock, i++);
+}
+
+void unlock_all_resources(void)
+{
+       struct drbd_resource *resource;
+
+       for_each_resource(resource, &drbd_resources)
+               spin_unlock(&resource->req_lock);
+       local_irq_enable();
+       mutex_unlock(&resources_mutex);
+}
+
 #ifdef CONFIG_DRBD_FAULT_INJECTION
 /* Fault insertion support including random number generator shamelessly
  * stolen from kernel/rcutorture.c */
index 94e380ff61cae8cdcc5f78baeeecb50c784eaf95..d37c509e6a440914acfa57a927f59ec28decd2dc 100644 (file)
@@ -1389,13 +1389,13 @@ int drbd_adm_disk_opts(struct sk_buff *skb, struct genl_info *info)
                goto fail_unlock;
        }
 
-       write_lock_irq(&global_state_lock);
+       lock_all_resources();
        retcode = drbd_resync_after_valid(device, new_disk_conf->resync_after);
        if (retcode == NO_ERROR) {
                rcu_assign_pointer(device->ldev->disk_conf, new_disk_conf);
                drbd_resync_after_changed(device);
        }
-       write_unlock_irq(&global_state_lock);
+       unlock_all_resources();
 
        if (retcode != NO_ERROR)
                goto fail_unlock;
@@ -1539,18 +1539,13 @@ int drbd_adm_attach(struct sk_buff *skb, struct genl_info *info)
                goto fail;
        }
 
-       write_lock_irq(&global_state_lock);
-       retcode = drbd_resync_after_valid(device, new_disk_conf->resync_after);
-       if (retcode != NO_ERROR)
-               goto fail_unlock;
-
        rcu_read_lock();
        nc = rcu_dereference(connection->net_conf);
        if (nc) {
                if (new_disk_conf->fencing == FP_STONITH && nc->wire_protocol == DRBD_PROT_A) {
                        rcu_read_unlock();
                        retcode = ERR_STONITH_AND_PROT_A;
-                       goto fail_unlock;
+                       goto fail;
                }
        }
        rcu_read_unlock();
@@ -1561,7 +1556,7 @@ int drbd_adm_attach(struct sk_buff *skb, struct genl_info *info)
                drbd_err(device, "open(\"%s\") failed with %ld\n", new_disk_conf->backing_dev,
                        PTR_ERR(bdev));
                retcode = ERR_OPEN_DISK;
-               goto fail_unlock;
+               goto fail;
        }
        nbc->backing_bdev = bdev;
 
@@ -1581,7 +1576,7 @@ int drbd_adm_attach(struct sk_buff *skb, struct genl_info *info)
                drbd_err(device, "open(\"%s\") failed with %ld\n", new_disk_conf->meta_dev,
                        PTR_ERR(bdev));
                retcode = ERR_OPEN_MD_DISK;
-               goto fail_unlock;
+               goto fail;
        }
        nbc->md_bdev = bdev;
 
@@ -1589,7 +1584,7 @@ int drbd_adm_attach(struct sk_buff *skb, struct genl_info *info)
            (new_disk_conf->meta_dev_idx == DRBD_MD_INDEX_INTERNAL ||
             new_disk_conf->meta_dev_idx == DRBD_MD_INDEX_FLEX_INT)) {
                retcode = ERR_MD_IDX_INVALID;
-               goto fail_unlock;
+               goto fail;
        }
 
        resync_lru = lc_create("resync", drbd_bm_ext_cache,
@@ -1597,14 +1592,14 @@ int drbd_adm_attach(struct sk_buff *skb, struct genl_info *info)
                        offsetof(struct bm_extent, lce));
        if (!resync_lru) {
                retcode = ERR_NOMEM;
-               goto fail_unlock;
+               goto fail;
        }
 
        /* Read our meta data super block early.
         * This also sets other on-disk offsets. */
        retcode = drbd_md_read(device, nbc);
        if (retcode != NO_ERROR)
-               goto fail_unlock;
+               goto fail;
 
        if (new_disk_conf->al_extents < DRBD_AL_EXTENTS_MIN)
                new_disk_conf->al_extents = DRBD_AL_EXTENTS_MIN;
@@ -1616,7 +1611,7 @@ int drbd_adm_attach(struct sk_buff *skb, struct genl_info *info)
                        (unsigned long long) drbd_get_max_capacity(nbc),
                        (unsigned long long) new_disk_conf->disk_size);
                retcode = ERR_DISK_TOO_SMALL;
-               goto fail_unlock;
+               goto fail;
        }
 
        if (new_disk_conf->meta_dev_idx < 0) {
@@ -1633,7 +1628,7 @@ int drbd_adm_attach(struct sk_buff *skb, struct genl_info *info)
                drbd_warn(device, "refusing attach: md-device too small, "
                     "at least %llu sectors needed for this meta-disk type\n",
                     (unsigned long long) min_md_device_sectors);
-               goto fail_unlock;
+               goto fail;
        }
 
        /* Make sure the new disk is big enough
@@ -1641,7 +1636,7 @@ int drbd_adm_attach(struct sk_buff *skb, struct genl_info *info)
        if (drbd_get_max_capacity(nbc) <
            drbd_get_capacity(device->this_bdev)) {
                retcode = ERR_DISK_TOO_SMALL;
-               goto fail_unlock;
+               goto fail;
        }
 
        nbc->known_size = drbd_get_capacity(nbc->backing_bdev);
@@ -1671,7 +1666,7 @@ int drbd_adm_attach(struct sk_buff *skb, struct genl_info *info)
        retcode = rv;  /* FIXME: Type mismatch. */
        drbd_resume_io(device);
        if (rv < SS_SUCCESS)
-               goto fail_unlock;
+               goto fail;
 
        if (!get_ldev_if_state(device, D_ATTACHING))
                goto force_diskless;
@@ -1706,6 +1701,13 @@ int drbd_adm_attach(struct sk_buff *skb, struct genl_info *info)
                goto force_diskless_dec;
        }
 
+       lock_all_resources();
+       retcode = drbd_resync_after_valid(device, new_disk_conf->resync_after);
+       if (retcode != NO_ERROR) {
+               unlock_all_resources();
+               goto force_diskless_dec;
+       }
+
        /* Reset the "barriers don't work" bits here, then force meta data to
         * be written, to ensure we determine if barriers are supported. */
        if (new_disk_conf->md_flushes)
@@ -1728,6 +1730,7 @@ int drbd_adm_attach(struct sk_buff *skb, struct genl_info *info)
 
        drbd_resync_after_changed(device);
        drbd_bump_write_ordering(device->resource, device->ldev, WO_BDEV_FLUSH);
+       unlock_all_resources();
 
        if (drbd_md_test_flag(device->ldev, MDF_CRASHED_PRIMARY))
                set_bit(CRASHED_PRIMARY, &device->flags);
@@ -1850,8 +1853,6 @@ int drbd_adm_attach(struct sk_buff *skb, struct genl_info *info)
        if (rv < SS_SUCCESS)
                goto force_diskless_dec;
 
-       write_unlock(&global_state_lock);
-
        mod_timer(&device->request_timer, jiffies + HZ);
 
        if (device->state.role == R_PRIMARY)
@@ -1874,8 +1875,6 @@ int drbd_adm_attach(struct sk_buff *skb, struct genl_info *info)
  force_diskless:
        drbd_force_state(device, NS(disk, D_DISKLESS));
        drbd_md_sync(device);
- fail_unlock:
-       write_unlock_irq(&global_state_lock);
  fail:
        conn_reconfig_done(connection);
        if (nbc) {
@@ -3453,8 +3452,10 @@ int drbd_adm_new_resource(struct sk_buff *skb, struct genl_info *info)
        }
 
        /* not yet safe for genl_family.parallel_ops */
+       mutex_lock(&resources_mutex);
        if (!conn_create(adm_ctx.resource_name, &res_opts))
                retcode = ERR_NOMEM;
+       mutex_unlock(&resources_mutex);
 out:
        drbd_adm_finish(&adm_ctx, info, retcode);
        return 0;
@@ -3545,7 +3546,9 @@ static int adm_del_resource(struct drbd_resource *resource)
        if (!idr_is_empty(&resource->devices))
                return ERR_RES_IN_USE;
 
+       mutex_lock(&resources_mutex);
        list_del_rcu(&resource->resources);
+       mutex_unlock(&resources_mutex);
        /* Make sure all threads have actually stopped: state handling only
         * does drbd_thread_stop_nowait(). */
        list_for_each_entry(connection, &resource->connections, connections)
index 2d7dd269b6a82a41f94ce044f4baf4c87c92d473..535ae47f84c9bc3ac774568ea3c0e77a520bd731 100644 (file)
@@ -937,7 +937,7 @@ void drbd_resume_al(struct drbd_device *device)
                drbd_info(device, "Resumed AL updates\n");
 }
 
-/* helper for __drbd_set_state */
+/* helper for _drbd_set_state */
 static void set_ov_position(struct drbd_device *device, enum drbd_conns cs)
 {
        if (first_peer_device(device)->connection->agreed_pro_version < 90)
@@ -965,17 +965,17 @@ static void set_ov_position(struct drbd_device *device, enum drbd_conns cs)
 }
 
 /**
- * __drbd_set_state() - Set a new DRBD state
+ * _drbd_set_state() - Set a new DRBD state
  * @device:    DRBD device.
  * @ns:                new state.
  * @flags:     Flags
  * @done:      Optional completion, that will get completed after the after_state_ch() finished
  *
- * Caller needs to hold req_lock, and global_state_lock. Do not call directly.
+ * Caller needs to hold req_lock. Do not call directly.
  */
 enum drbd_state_rv
-__drbd_set_state(struct drbd_device *device, union drbd_state ns,
-                enum chg_state_flags flags, struct completion *done)
+_drbd_set_state(struct drbd_device *device, union drbd_state ns,
+               enum chg_state_flags flags, struct completion *done)
 {
        struct drbd_peer_device *peer_device = first_peer_device(device);
        struct drbd_connection *connection = peer_device ? peer_device->connection : NULL;
@@ -1444,7 +1444,7 @@ static void after_state_ch(struct drbd_device *device, union drbd_state os,
        if (os.disk != D_FAILED && ns.disk == D_FAILED) {
                enum drbd_io_error_p eh = EP_PASS_ON;
                int was_io_error = 0;
-               /* corresponding get_ldev was in __drbd_set_state, to serialize
+               /* corresponding get_ldev was in _drbd_set_state, to serialize
                 * our cleanup here with the transition to D_DISKLESS.
                 * But is is still not save to dreference ldev here, since
                 * we might come from an failed Attach before ldev was set. */
@@ -1759,7 +1759,7 @@ conn_set_state(struct drbd_connection *connection, union drbd_state mask, union
                if (flags & CS_IGN_OUTD_FAIL && ns.disk == D_OUTDATED && os.disk < D_OUTDATED)
                        ns.disk = os.disk;
 
-               rv = __drbd_set_state(device, ns, flags, NULL);
+               rv = _drbd_set_state(device, ns, flags, NULL);
                if (rv < SS_SUCCESS)
                        BUG();
 
index 7f53c40823cd581e28279b9ef046a07af54a974e..bd989536f888fc05af96942858795dc34713045f 100644 (file)
@@ -122,9 +122,9 @@ extern enum drbd_state_rv
 _drbd_request_state_holding_state_mutex(struct drbd_device *, union drbd_state,
                                        union drbd_state, enum chg_state_flags);
 
-extern enum drbd_state_rv __drbd_set_state(struct drbd_device *, union drbd_state,
-                                          enum chg_state_flags,
-                                          struct completion *done);
+extern enum drbd_state_rv _drbd_set_state(struct drbd_device *, union drbd_state,
+                                         enum chg_state_flags,
+                                         struct completion *done);
 extern void print_st_err(struct drbd_device *, union drbd_state,
                        union drbd_state, int);
 
index 5578c1477ba6610f27f80e4cffaa874939a05364..3b3d980ac8c7bf79f84f5a51870011673fdd4fc7 100644 (file)
@@ -55,13 +55,6 @@ static int make_resync_request(struct drbd_device *, int);
  *
  */
 
-
-/* About the global_state_lock
-   Each state transition on an device holds a read lock. In case we have
-   to evaluate the resync after dependencies, we grab a write lock, because
-   we need stable states on all devices for that.  */
-rwlock_t global_state_lock;
-
 /* used for synchronous meta data and bitmap IO
  * submitted by drbd_md_sync_page_io()
  */
@@ -1456,70 +1449,73 @@ static int _drbd_may_sync_now(struct drbd_device *device)
 }
 
 /**
- * _drbd_pause_after() - Pause resync on all devices that may not resync now
+ * drbd_pause_after() - Pause resync on all devices that may not resync now
  * @device:    DRBD device.
  *
  * Called from process context only (admin command and after_state_ch).
  */
-static int _drbd_pause_after(struct drbd_device *device)
+static bool drbd_pause_after(struct drbd_device *device)
 {
+       bool changed = false;
        struct drbd_device *odev;
-       int i, rv = 0;
+       int i;
 
        rcu_read_lock();
        idr_for_each_entry(&drbd_devices, odev, i) {
                if (odev->state.conn == C_STANDALONE && odev->state.disk == D_DISKLESS)
                        continue;
-               if (!_drbd_may_sync_now(odev))
-                       rv |= (__drbd_set_state(_NS(odev, aftr_isp, 1), CS_HARD, NULL)
-                              != SS_NOTHING_TO_DO);
+               if (!_drbd_may_sync_now(odev) &&
+                   _drbd_set_state(_NS(odev, aftr_isp, 1),
+                                   CS_HARD, NULL) != SS_NOTHING_TO_DO)
+                       changed = true;
        }
        rcu_read_unlock();
 
-       return rv;
+       return changed;
 }
 
 /**
- * _drbd_resume_next() - Resume resync on all devices that may resync now
+ * drbd_resume_next() - Resume resync on all devices that may resync now
  * @device:    DRBD device.
  *
  * Called from process context only (admin command and worker).
  */
-static int _drbd_resume_next(struct drbd_device *device)
+static bool drbd_resume_next(struct drbd_device *device)
 {
+       bool changed = false;
        struct drbd_device *odev;
-       int i, rv = 0;
+       int i;
 
        rcu_read_lock();
        idr_for_each_entry(&drbd_devices, odev, i) {
                if (odev->state.conn == C_STANDALONE && odev->state.disk == D_DISKLESS)
                        continue;
                if (odev->state.aftr_isp) {
-                       if (_drbd_may_sync_now(odev))
-                               rv |= (__drbd_set_state(_NS(odev, aftr_isp, 0),
-                                                       CS_HARD, NULL)
-                                      != SS_NOTHING_TO_DO) ;
+                       if (_drbd_may_sync_now(odev) &&
+                           _drbd_set_state(_NS(odev, aftr_isp, 0),
+                                           CS_HARD, NULL) != SS_NOTHING_TO_DO)
+                               changed = true;
                }
        }
        rcu_read_unlock();
-       return rv;
+       return changed;
 }
 
 void resume_next_sg(struct drbd_device *device)
 {
-       write_lock_irq(&global_state_lock);
-       _drbd_resume_next(device);
-       write_unlock_irq(&global_state_lock);
+       lock_all_resources();
+       drbd_resume_next(device);
+       unlock_all_resources();
 }
 
 void suspend_other_sg(struct drbd_device *device)
 {
-       write_lock_irq(&global_state_lock);
-       _drbd_pause_after(device);
-       write_unlock_irq(&global_state_lock);
+       lock_all_resources();
+       drbd_pause_after(device);
+       unlock_all_resources();
 }
 
-/* caller must hold global_state_lock */
+/* caller must lock_all_resources() */
 enum drbd_ret_code drbd_resync_after_valid(struct drbd_device *device, int o_minor)
 {
        struct drbd_device *odev;
@@ -1557,15 +1553,15 @@ enum drbd_ret_code drbd_resync_after_valid(struct drbd_device *device, int o_min
        }
 }
 
-/* caller must hold global_state_lock */
+/* caller must lock_all_resources() */
 void drbd_resync_after_changed(struct drbd_device *device)
 {
-       int changes;
+       int changed;
 
        do {
-               changes  = _drbd_pause_after(device);
-               changes |= _drbd_resume_next(device);
-       } while (changes);
+               changed  = drbd_pause_after(device);
+               changed |= drbd_resume_next(device);
+       } while (changed);
 }
 
 void drbd_rs_controller_reset(struct drbd_device *device)
@@ -1685,19 +1681,14 @@ void drbd_start_resync(struct drbd_device *device, enum drbd_conns side)
        } else {
                mutex_lock(device->state_mutex);
        }
-       clear_bit(B_RS_H_DONE, &device->flags);
 
-       /* req_lock: serialize with drbd_send_and_submit() and others
-        * global_state_lock: for stable sync-after dependencies */
-       spin_lock_irq(&device->resource->req_lock);
-       write_lock(&global_state_lock);
+       lock_all_resources();
+       clear_bit(B_RS_H_DONE, &device->flags);
        /* Did some connection breakage or IO error race with us? */
        if (device->state.conn < C_CONNECTED
        || !get_ldev_if_state(device, D_NEGOTIATING)) {
-               write_unlock(&global_state_lock);
-               spin_unlock_irq(&device->resource->req_lock);
-               mutex_unlock(device->state_mutex);
-               return;
+               unlock_all_resources();
+               goto out;
        }
 
        ns = drbd_read_state(device);
@@ -1711,7 +1702,7 @@ void drbd_start_resync(struct drbd_device *device, enum drbd_conns side)
        else /* side == C_SYNC_SOURCE */
                ns.pdsk = D_INCONSISTENT;
 
-       r = __drbd_set_state(device, ns, CS_VERBOSE, NULL);
+       r = _drbd_set_state(device, ns, CS_VERBOSE, NULL);
        ns = drbd_read_state(device);
 
        if (ns.conn < C_CONNECTED)
@@ -1732,7 +1723,7 @@ void drbd_start_resync(struct drbd_device *device, enum drbd_conns side)
                        device->rs_mark_left[i] = tw;
                        device->rs_mark_time[i] = now;
                }
-               _drbd_pause_after(device);
+               drbd_pause_after(device);
                /* Forget potentially stale cached per resync extent bit-counts.
                 * Open coded drbd_rs_cancel_all(device), we already have IRQs
                 * disabled, and know the disk state is ok. */
@@ -1742,8 +1733,7 @@ void drbd_start_resync(struct drbd_device *device, enum drbd_conns side)
                device->resync_wenr = LC_FREE;
                spin_unlock(&device->al_lock);
        }
-       write_unlock(&global_state_lock);
-       spin_unlock_irq(&device->resource->req_lock);
+       unlock_all_resources();
 
        if (r == SS_SUCCESS) {
                wake_up(&device->al_wait); /* for lc_reset() above */
@@ -1807,6 +1797,7 @@ void drbd_start_resync(struct drbd_device *device, enum drbd_conns side)
                drbd_md_sync(device);
        }
        put_ldev(device);
+out:
        mutex_unlock(device->state_mutex);
 }