]> git.openfabrics.org - ~shefty/rdma-dev.git/commitdiff
firewire: Implement compliant bus management.
authorKristian Høgsberg <krh@redhat.com>
Fri, 26 Jan 2007 05:38:45 +0000 (00:38 -0500)
committerStefan Richter <stefanr@s5r6.in-berlin.de>
Fri, 9 Mar 2007 21:02:48 +0000 (22:02 +0100)
Signed-off-by: Kristian Høgsberg <krh@redhat.com>
Signed-off-by: Stefan Richter <stefanr@s5r6.in-berlin.de>
drivers/firewire/fw-card.c
drivers/firewire/fw-device.c
drivers/firewire/fw-topology.c
drivers/firewire/fw-transaction.h

index 307c8b851382e25802fb4714dc7d9b31ca90ea03..f94874ce58af0194be47cd54d8d0e711e44805a1 100644 (file)
@@ -92,7 +92,7 @@ generate_config_rom (struct fw_card *card, size_t *config_rom_length)
                bib_generation(card->config_rom_generation++ % 14 + 2) |
                bib_max_rom(2) |
                bib_max_receive(card->max_receive) |
-               bib_isc | bib_cmc | bib_imc;
+               bib_bmc | bib_isc | bib_cmc | bib_imc;
        config_rom[3] = card->guid >> 32;
        config_rom[4] = card->guid;
 
@@ -190,48 +190,137 @@ static const char gap_count_table[] = {
        63, 5, 7, 8, 10, 13, 16, 18, 21, 24, 26, 29, 32, 35, 37, 40
 };
 
+struct bm_data {
+       struct fw_transaction t;
+       struct {
+               __be32 arg;
+               __be32 data;
+       } lock;
+       u32 old;
+       int rcode;
+       struct completion done;
+};
+
 static void
-fw_card_irm_work(struct work_struct *work)
+complete_bm_lock(struct fw_card *card, int rcode,
+                void *payload, size_t length, void *data)
+{
+       struct bm_data *bmd = data;
+
+       if (rcode == RCODE_COMPLETE)
+               bmd->old = be32_to_cpu(*(__be32 *) payload);
+       bmd->rcode = rcode;
+       complete(&bmd->done);
+}
+
+static void
+fw_card_bm_work(struct work_struct *work)
 {
        struct fw_card *card = container_of(work, struct fw_card, work.work);
        struct fw_device *root;
+       struct bm_data bmd;
        unsigned long flags;
-       int root_id, new_irm_id, gap_count, generation, do_reset = 0;
-
-       /* FIXME: This simple bus management unconditionally picks a
-        * cycle master if the current root can't do it.  We need to
-        * not do this if there is a bus manager already.  Also, some
-        * hubs set the contender bit, which is bogus, so we should
-        * probably do a little sanity check on the IRM (like, read
-        * the bandwidth register) if it's not us. */
+       int root_id, new_root_id, irm_id, gap_count, generation, grace;
+       int do_reset = 0;
 
        spin_lock_irqsave(&card->lock, flags);
 
        generation = card->generation;
        root = card->root_node->data;
        root_id = card->root_node->node_id;
+       grace = time_after(jiffies, card->reset_jiffies + DIV_ROUND_UP(HZ, 10));
+
+       if (card->bm_generation + 1 == generation ||
+           (card->bm_generation != generation && grace)) {
+               /* This first step is to figure out who is IRM and
+                * then try to become bus manager.  If the IRM is not
+                * well defined (e.g. does not have an active link
+                * layer or does not responds to our lock request, we
+                * will have to do a little vigilante bus management.
+                * In that case, we do a goto into the gap count logic
+                * so that when we do the reset, we still optimize the
+                * gap count.  That could well save a reset in the
+                * next generation. */
+
+               irm_id = card->irm_node->node_id;
+               if (!card->irm_node->link_on) {
+                       new_root_id = card->local_node->node_id;
+                       fw_notify("IRM has link off, making local node (%02x) root.\n",
+                                 new_root_id);
+                       goto pick_me;
+               }
+
+               bmd.lock.arg = cpu_to_be32(0x3f);
+               bmd.lock.data = cpu_to_be32(card->local_node->node_id);
+
+               spin_unlock_irqrestore(&card->lock, flags);
+
+               init_completion(&bmd.done);
+               fw_send_request(card, &bmd.t, TCODE_LOCK_COMPARE_SWAP,
+                               irm_id, generation,
+                               SCODE_100, CSR_REGISTER_BASE + CSR_BUS_MANAGER_ID,
+                               &bmd.lock, sizeof bmd.lock,
+                               complete_bm_lock, &bmd);
+               wait_for_completion(&bmd.done);
+
+               if (bmd.rcode == RCODE_GENERATION) {
+                       /* Another bus reset happened. Just return,
+                        * the BM work has been rescheduled. */
+                       return;
+               }
+
+               if (bmd.rcode == RCODE_COMPLETE && bmd.old != 0x3f)
+                       /* Somebody else is BM, let them do the work. */
+                       return;
+
+               spin_lock_irqsave(&card->lock, flags);
+               if (bmd.rcode != RCODE_COMPLETE) {
+                       /* The lock request failed, maybe the IRM
+                        * isn't really IRM capable after all. Let's
+                        * do a bus reset and pick the local node as
+                        * root, and thus, IRM. */
+                       new_root_id = card->local_node->node_id;
+                       fw_notify("BM lock failed, making local node (%02x) root.\n",
+                                 new_root_id);
+                       goto pick_me;
+               }
+       } else if (card->bm_generation != generation) {
+               /* OK, we weren't BM in the last generation, and it's
+                * less than 100ms since last bus reset. Reschedule
+                * this task 100ms from now. */
+               spin_unlock_irqrestore(&card->lock, flags);
+               schedule_delayed_work(&card->work, DIV_ROUND_UP(HZ, 10));
+               return;
+       }
+
+       /* We're bus manager for this generation, so next step is to
+        * make sure we have an active cycle master and do gap count
+        * optimization. */
+       card->bm_generation = generation;
 
        if (root == NULL) {
                /* Either link_on is false, or we failed to read the
                 * config rom.  In either case, pick another root. */
-               new_irm_id = card->local_node->node_id;
+               new_root_id = card->local_node->node_id;
        } else if (root->state != FW_DEVICE_RUNNING) {
                /* If we haven't probed this device yet, bail out now
                 * and let's try again once that's done. */
-               new_irm_id = root_id;
+               spin_unlock_irqrestore(&card->lock, flags);
+               return;
        } else if (root->config_rom[2] & bib_cmc) {
                /* FIXME: I suppose we should set the cmstr bit in the
                 * STATE_CLEAR register of this node, as described in
                 * 1394-1995, 8.4.2.6.  Also, send out a force root
                 * packet for this node. */
-               new_irm_id = root_id;
+               new_root_id = root_id;
        } else {
                /* Current root has an active link layer and we
                 * successfully read the config rom, but it's not
                 * cycle master capable. */
-               new_irm_id = card->local_node->node_id;
+               new_root_id = card->local_node->node_id;
        }
 
+ pick_me:
        /* Now figure out what gap count to set. */
        if (card->topology_type == FW_TOPOLOGY_A &&
            card->root_node->max_hops < ARRAY_SIZE(gap_count_table))
@@ -243,16 +332,16 @@ fw_card_irm_work(struct work_struct *work)
         * done less that 5 resets with the same physical topology and we
         * have either a new root or a new gap count setting, let's do it. */
 
-       if (card->irm_retries++ < 5 &&
-           (card->gap_count != gap_count || new_irm_id != root_id))
+       if (card->bm_retries++ < 5 &&
+           (card->gap_count != gap_count || new_root_id != root_id))
                do_reset = 1;
 
        spin_unlock_irqrestore(&card->lock, flags);
 
        if (do_reset) {
                fw_notify("phy config: card %d, new root=%x, gap_count=%d\n",
-                         card->index, new_irm_id, gap_count);
-               fw_send_phy_config(card, new_irm_id, generation, gap_count);
+                         card->index, new_root_id, gap_count);
+               fw_send_phy_config(card, new_root_id, generation, gap_count);
                fw_core_initiate_bus_reset(card, 1);
        }
 }
@@ -294,7 +383,7 @@ fw_card_initialize(struct fw_card *card, const struct fw_card_driver *driver,
 
        card->local_node = NULL;
 
-       INIT_DELAYED_WORK(&card->work, fw_card_irm_work);
+       INIT_DELAYED_WORK(&card->work, fw_card_bm_work);
 
        card->card_device.bus     = &fw_bus_type;
        card->card_device.release = release_card;
index d71824bb61cb0ad61d8d3ac50abb51d7babc9b72..4ea8d4d08a61d26f1123336a9354a373c8faff21 100644 (file)
@@ -449,6 +449,8 @@ static void fw_device_init(struct work_struct *work)
                } else {
                        fw_notify("giving up on config rom for node id %x\n",
                                  device->node_id);
+                       if (device->node == device->card->root_node)
+                               schedule_delayed_work(&device->card->work, 0);
                        fw_device_release(&device->device);
                }
                return;
index fc8f1e7d3b076093b1badd72e0e37ed6b896d653..27c6cb9c8367452265c018451e97a7de13d1ded4 100644 (file)
@@ -476,12 +476,13 @@ fw_core_handle_bus_reset(struct fw_card *card,
         * changed, either nodes were added or removed. In that case we
         * reset the IRM reset counter. */
        if (card->self_id_count != self_id_count)
-               card->irm_retries = 0;
+               card->bm_retries = 0;
 
        card->node_id = node_id;
        card->self_id_count = self_id_count;
        card->generation = generation;
        memcpy(card->self_ids, self_ids, self_id_count * 4);
+       card->reset_jiffies = jiffies;
 
        local_node = build_tree(card);
 
@@ -497,9 +498,7 @@ fw_core_handle_bus_reset(struct fw_card *card,
                update_tree(card, local_node);
        }
 
-       /* If we're not the root node, we may have to do some IRM work. */
-       if (card->local_node != card->root_node)
-               schedule_delayed_work(&card->work, 0);
+       schedule_delayed_work(&card->work, 0);
 
        spin_unlock_irqrestore(&card->lock, flags);
 }
index ad7ba32755eb8881c67baaad74bbc0313be1bb48..50ec2c683a318165af0d85505982f37d8b79c643 100644 (file)
@@ -276,6 +276,7 @@ struct fw_card {
        int current_tlabel, tlabel_mask;
        struct list_head transaction_list;
        struct timer_list flush_timer;
+       unsigned long reset_jiffies;
 
        unsigned long long guid;
        int max_receive;
@@ -301,9 +302,10 @@ struct fw_card {
 
        struct list_head link;
 
-       /* Work struct for IRM duties. */
+       /* Work struct for BM duties. */
        struct delayed_work work;
-       int irm_retries;
+       int bm_retries;
+       int bm_generation;
 };
 
 struct fw_card *fw_card_get(struct fw_card *card);