mlx4_en: Fix a kernel panic when waking tx queue
[safe/jmp/linux-2.6] / drivers / firewire / fw-device.c
index a834221..a47e212 100644 (file)
@@ -1,6 +1,5 @@
-/*                                             -*- c-basic-offset: 8 -*-
- *
- * fw-device.c - Device probing and sysfs code.
+/*
+ * Device probing and sysfs code.
  *
  * Copyright (C) 2005-2006  Kristian Hoegsberg <krh@bitplanet.net>
  *
  * Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
  */
 
-#include <linux/module.h>
-#include <linux/wait.h>
-#include <linux/errno.h>
-#include <linux/kthread.h>
-#include <linux/device.h>
+#include <linux/ctype.h>
 #include <linux/delay.h>
+#include <linux/device.h>
+#include <linux/errno.h>
 #include <linux/idr.h>
+#include <linux/jiffies.h>
+#include <linux/kobject.h>
+#include <linux/list.h>
+#include <linux/mutex.h>
 #include <linux/rwsem.h>
-#include <asm/semaphore.h>
-#include "fw-transaction.h"
-#include "fw-topology.h"
+#include <linux/semaphore.h>
+#include <linux/spinlock.h>
+#include <linux/string.h>
+#include <linux/workqueue.h>
+
+#include <asm/system.h>
+
 #include "fw-device.h"
+#include "fw-topology.h"
+#include "fw-transaction.h"
 
 void fw_csr_iterator_init(struct fw_csr_iterator *ci, u32 * p)
 {
@@ -129,78 +136,172 @@ static int get_modalias(struct fw_unit *unit, char *buffer, size_t buffer_size)
                        vendor, model, specifier_id, version);
 }
 
-static int
-fw_unit_uevent(struct device *dev, char **envp, int num_envp,
-              char *buffer, int buffer_size)
+static int fw_unit_uevent(struct device *dev, struct kobj_uevent_env *env)
 {
        struct fw_unit *unit = fw_unit(dev);
        char modalias[64];
-       int length = 0;
-       int i = 0;
-
-       if (!is_fw_unit(dev))
-               goto out;
 
-       get_modalias(unit, modalias, sizeof modalias);
+       get_modalias(unit, modalias, sizeof(modalias));
 
-       if (add_uevent_var(envp, num_envp, &i,
-                          buffer, buffer_size, &length,
-                          "MODALIAS=%s", modalias))
+       if (add_uevent_var(env, "MODALIAS=%s", modalias))
                return -ENOMEM;
 
- out:
-       envp[i] = NULL;
-
        return 0;
 }
 
 struct bus_type fw_bus_type = {
        .name = "firewire",
        .match = fw_unit_match,
-       .uevent = fw_unit_uevent,
 };
 EXPORT_SYMBOL(fw_bus_type);
 
-extern struct fw_device *fw_device_get(struct fw_device *device)
+int fw_device_enable_phys_dma(struct fw_device *device)
 {
-       get_device(&device->device);
+       int generation = device->generation;
 
-       return device;
+       /* device->node_id, accessed below, must not be older than generation */
+       smp_rmb();
+
+       return device->card->driver->enable_phys_dma(device->card,
+                                                    device->node_id,
+                                                    generation);
 }
+EXPORT_SYMBOL(fw_device_enable_phys_dma);
+
+struct config_rom_attribute {
+       struct device_attribute attr;
+       u32 key;
+};
 
-extern void fw_device_put(struct fw_device *device)
+static ssize_t show_immediate(struct device *dev,
+                             struct device_attribute *dattr, char *buf)
 {
-       put_device(&device->device);
+       struct config_rom_attribute *attr =
+               container_of(dattr, struct config_rom_attribute, attr);
+       struct fw_csr_iterator ci;
+       u32 *dir;
+       int key, value, ret = -ENOENT;
+
+       down_read(&fw_device_rwsem);
+
+       if (is_fw_unit(dev))
+               dir = fw_unit(dev)->directory;
+       else
+               dir = fw_device(dev)->config_rom + 5;
+
+       fw_csr_iterator_init(&ci, dir);
+       while (fw_csr_iterator_next(&ci, &key, &value))
+               if (attr->key == key) {
+                       ret = snprintf(buf, buf ? PAGE_SIZE : 0,
+                                      "0x%06x\n", value);
+                       break;
+               }
+
+       up_read(&fw_device_rwsem);
+
+       return ret;
 }
 
-static void fw_device_release(struct device *dev)
+#define IMMEDIATE_ATTR(name, key)                              \
+       { __ATTR(name, S_IRUGO, show_immediate, NULL), key }
+
+static ssize_t show_text_leaf(struct device *dev,
+                             struct device_attribute *dattr, char *buf)
 {
-       struct fw_device *device = fw_device(dev);
-       unsigned long flags;
+       struct config_rom_attribute *attr =
+               container_of(dattr, struct config_rom_attribute, attr);
+       struct fw_csr_iterator ci;
+       u32 *dir, *block = NULL, *p, *end;
+       int length, key, value, last_key = 0, ret = -ENOENT;
+       char *b;
 
-       /* Take the card lock so we don't set this to NULL while a
-        * FW_NODE_UPDATED callback is being handled. */
-       spin_lock_irqsave(&device->card->lock, flags);
-       device->node->data = NULL;
-       spin_unlock_irqrestore(&device->card->lock, flags);
+       down_read(&fw_device_rwsem);
 
-       fw_node_put(device->node);
-       fw_card_put(device->card);
-       kfree(device->config_rom);
-       kfree(device);
+       if (is_fw_unit(dev))
+               dir = fw_unit(dev)->directory;
+       else
+               dir = fw_device(dev)->config_rom + 5;
+
+       fw_csr_iterator_init(&ci, dir);
+       while (fw_csr_iterator_next(&ci, &key, &value)) {
+               if (attr->key == last_key &&
+                   key == (CSR_DESCRIPTOR | CSR_LEAF))
+                       block = ci.p - 1 + value;
+               last_key = key;
+       }
+
+       if (block == NULL)
+               goto out;
+
+       length = min(block[0] >> 16, 256U);
+       if (length < 3)
+               goto out;
+
+       if (block[1] != 0 || block[2] != 0)
+               /* Unknown encoding. */
+               goto out;
+
+       if (buf == NULL) {
+               ret = length * 4;
+               goto out;
+       }
+
+       b = buf;
+       end = &block[length + 1];
+       for (p = &block[3]; p < end; p++, b += 4)
+               * (u32 *) b = (__force u32) __cpu_to_be32(*p);
+
+       /* Strip trailing whitespace and add newline. */
+       while (b--, (isspace(*b) || *b == '\0') && b > buf);
+       strcpy(b + 1, "\n");
+       ret = b + 2 - buf;
+ out:
+       up_read(&fw_device_rwsem);
+
+       return ret;
 }
 
-int fw_device_enable_phys_dma(struct fw_device *device)
+#define TEXT_LEAF_ATTR(name, key)                              \
+       { __ATTR(name, S_IRUGO, show_text_leaf, NULL), key }
+
+static struct config_rom_attribute config_rom_attributes[] = {
+       IMMEDIATE_ATTR(vendor, CSR_VENDOR),
+       IMMEDIATE_ATTR(hardware_version, CSR_HARDWARE_VERSION),
+       IMMEDIATE_ATTR(specifier_id, CSR_SPECIFIER_ID),
+       IMMEDIATE_ATTR(version, CSR_VERSION),
+       IMMEDIATE_ATTR(model, CSR_MODEL),
+       TEXT_LEAF_ATTR(vendor_name, CSR_VENDOR),
+       TEXT_LEAF_ATTR(model_name, CSR_MODEL),
+       TEXT_LEAF_ATTR(hardware_version_name, CSR_HARDWARE_VERSION),
+};
+
+static void init_fw_attribute_group(struct device *dev,
+                                   struct device_attribute *attrs,
+                                   struct fw_attribute_group *group)
 {
-       return device->card->driver->enable_phys_dma(device->card,
-                                                    device->node_id,
-                                                    device->generation);
+       struct device_attribute *attr;
+       int i, j;
+
+       for (j = 0; attrs[j].attr.name != NULL; j++)
+               group->attrs[j] = &attrs[j].attr;
+
+       for (i = 0; i < ARRAY_SIZE(config_rom_attributes); i++) {
+               attr = &config_rom_attributes[i].attr;
+               if (attr->show(dev, attr, NULL) < 0)
+                       continue;
+               group->attrs[j++] = &attr->attr;
+       }
+
+       BUG_ON(j >= ARRAY_SIZE(group->attrs));
+       group->attrs[j++] = NULL;
+       group->groups[0] = &group->group;
+       group->groups[1] = NULL;
+       group->group.attrs = group->attrs;
+       dev->groups = group->groups;
 }
-EXPORT_SYMBOL(fw_device_enable_phys_dma);
 
-static ssize_t
-show_modalias_attribute(struct device *dev,
-                       struct device_attribute *attr, char *buf)
+static ssize_t modalias_show(struct device *dev,
+                            struct device_attribute *attr, char *buf)
 {
        struct fw_unit *unit = fw_unit(dev);
        int length;
@@ -211,142 +312,193 @@ show_modalias_attribute(struct device *dev,
        return length + 1;
 }
 
-static struct device_attribute modalias_attribute = {
-       .attr = { .name = "modalias", .mode = S_IRUGO, },
-       .show = show_modalias_attribute,
+static ssize_t rom_index_show(struct device *dev,
+                             struct device_attribute *attr, char *buf)
+{
+       struct fw_device *device = fw_device(dev->parent);
+       struct fw_unit *unit = fw_unit(dev);
+
+       return snprintf(buf, PAGE_SIZE, "%d\n",
+                       (int)(unit->directory - device->config_rom));
+}
+
+static struct device_attribute fw_unit_attributes[] = {
+       __ATTR_RO(modalias),
+       __ATTR_RO(rom_index),
+       __ATTR_NULL,
 };
 
-static ssize_t
-show_config_rom_attribute(struct device *dev,
-                         struct device_attribute *attr, char *buf)
+static ssize_t config_rom_show(struct device *dev,
+                              struct device_attribute *attr, char *buf)
 {
        struct fw_device *device = fw_device(dev);
+       size_t length;
 
-       memcpy(buf, device->config_rom, device->config_rom_length * 4);
+       down_read(&fw_device_rwsem);
+       length = device->config_rom_length * 4;
+       memcpy(buf, device->config_rom, length);
+       up_read(&fw_device_rwsem);
 
-       return device->config_rom_length * 4;
+       return length;
 }
 
-static struct device_attribute config_rom_attribute = {
-       .attr = {.name = "config_rom", .mode = S_IRUGO,},
-       .show = show_config_rom_attribute,
-};
-
-static ssize_t
-show_rom_index_attribute(struct device *dev,
+static ssize_t guid_show(struct device *dev,
                         struct device_attribute *attr, char *buf)
 {
-       struct fw_device *device = fw_device(dev->parent);
-       struct fw_unit *unit = fw_unit(dev);
+       struct fw_device *device = fw_device(dev);
+       int ret;
 
-       return snprintf(buf, PAGE_SIZE, "%d\n",
-                       (int)(unit->directory - device->config_rom));
-}
+       down_read(&fw_device_rwsem);
+       ret = snprintf(buf, PAGE_SIZE, "0x%08x%08x\n",
+                      device->config_rom[3], device->config_rom[4]);
+       up_read(&fw_device_rwsem);
 
-static struct device_attribute rom_index_attribute = {
-       .attr = { .name = "rom_index", .mode = S_IRUGO, },
-       .show = show_rom_index_attribute,
-};
+       return ret;
+}
 
-struct read_quadlet_callback_data {
-       struct completion done;
-       int rcode;
-       u32 data;
+static struct device_attribute fw_device_attributes[] = {
+       __ATTR_RO(config_rom),
+       __ATTR_RO(guid),
+       __ATTR_NULL,
 };
 
-static void
-complete_transaction(struct fw_card *card, int rcode,
-                    void *payload, size_t length, void *data)
+static int read_rom(struct fw_device *device,
+                   int generation, int index, u32 *data)
 {
-       struct read_quadlet_callback_data *callback_data = data;
+       int rcode;
 
-       if (rcode == RCODE_COMPLETE)
-               callback_data->data = be32_to_cpu(*(__be32 *)payload);
-       callback_data->rcode = rcode;
-       complete(&callback_data->done);
-}
+       /* device->node_id, accessed below, must not be older than generation */
+       smp_rmb();
 
-static int read_rom(struct fw_device *device, int index, u32 * data)
-{
-       struct read_quadlet_callback_data callback_data;
-       struct fw_transaction t;
-       u64 offset;
+       rcode = fw_run_transaction(device->card, TCODE_READ_QUADLET_REQUEST,
+                       device->node_id, generation, device->max_speed,
+                       (CSR_REGISTER_BASE | CSR_CONFIG_ROM) + index * 4,
+                       data, 4);
+       be32_to_cpus(data);
 
-       init_completion(&callback_data.done);
+       return rcode;
+}
 
-       offset = 0xfffff0000400ULL + index * 4;
-       fw_send_request(device->card, &t, TCODE_READ_QUADLET_REQUEST,
-                       device->node_id,
-                       device->generation, SCODE_100,
-                       offset, NULL, 4, complete_transaction, &callback_data);
+#define READ_BIB_ROM_SIZE      256
+#define READ_BIB_STACK_SIZE    16
 
-       wait_for_completion(&callback_data.done);
+/*
+ * Read the bus info block, perform a speed probe, and read all of the rest of
+ * the config ROM.  We do all this with a cached bus generation.  If the bus
+ * generation changes under us, read_bus_info_block will fail and get retried.
+ * It's better to start all over in this case because the node from which we
+ * are reading the ROM may have changed the ROM during the reset.
+ */
+static int read_bus_info_block(struct fw_device *device, int generation)
+{
+       u32 *rom, *stack, *old_rom, *new_rom;
+       u32 sp, key;
+       int i, end, length, ret = -1;
 
-       *data = callback_data.data;
+       rom = kmalloc(sizeof(*rom) * READ_BIB_ROM_SIZE +
+                     sizeof(*stack) * READ_BIB_STACK_SIZE, GFP_KERNEL);
+       if (rom == NULL)
+               return -ENOMEM;
 
-       return callback_data.rcode;
-}
+       stack = &rom[READ_BIB_ROM_SIZE];
 
-static int read_bus_info_block(struct fw_device *device)
-{
-       static u32 rom[256];
-       u32 stack[16], sp, key;
-       int i, end, length;
+       device->max_speed = SCODE_100;
 
        /* First read the bus info block. */
        for (i = 0; i < 5; i++) {
-               if (read_rom(device, i, &rom[i]) != RCODE_COMPLETE)
-                       return -1;
-               /* As per IEEE1212 7.2, during power-up, devices can
+               if (read_rom(device, generation, i, &rom[i]) != RCODE_COMPLETE)
+                       goto out;
+               /*
+                * As per IEEE1212 7.2, during power-up, devices can
                 * reply with a 0 for the first quadlet of the config
                 * rom to indicate that they are booting (for example,
                 * if the firmware is on the disk of a external
                 * harddisk).  In that case we just fail, and the
-                * retry mechanism will try again later. */
+                * retry mechanism will try again later.
+                */
                if (i == 0 && rom[i] == 0)
-                       return -1;
+                       goto out;
+       }
+
+       device->max_speed = device->node->max_speed;
+
+       /*
+        * Determine the speed of
+        *   - devices with link speed less than PHY speed,
+        *   - devices with 1394b PHY (unless only connected to 1394a PHYs),
+        *   - all devices if there are 1394b repeaters.
+        * Note, we cannot use the bus info block's link_spd as starting point
+        * because some buggy firmwares set it lower than necessary and because
+        * 1394-1995 nodes do not have the field.
+        */
+       if ((rom[2] & 0x7) < device->max_speed ||
+           device->max_speed == SCODE_BETA ||
+           device->card->beta_repeaters_present) {
+               u32 dummy;
+
+               /* for S1600 and S3200 */
+               if (device->max_speed == SCODE_BETA)
+                       device->max_speed = device->card->link_speed;
+
+               while (device->max_speed > SCODE_100) {
+                       if (read_rom(device, generation, 0, &dummy) ==
+                           RCODE_COMPLETE)
+                               break;
+                       device->max_speed--;
+               }
        }
 
-       /* Now parse the config rom.  The config rom is a recursive
+       /*
+        * Now parse the config rom.  The config rom is a recursive
         * directory structure so we parse it using a stack of
         * references to the blocks that make up the structure.  We
         * push a reference to the root directory on the stack to
-        * start things off. */
+        * start things off.
+        */
        length = i;
        sp = 0;
        stack[sp++] = 0xc0000005;
        while (sp > 0) {
-               /* Pop the next block reference of the stack.  The
+               /*
+                * Pop the next block reference of the stack.  The
                 * lower 24 bits is the offset into the config rom,
                 * the upper 8 bits are the type of the reference the
-                * block. */
+                * block.
+                */
                key = stack[--sp];
                i = key & 0xffffff;
-               if (i >= ARRAY_SIZE(rom))
-                       /* The reference points outside the standard
-                        * config rom area, something's fishy. */
-                       return -1;
+               if (i >= READ_BIB_ROM_SIZE)
+                       /*
+                        * The reference points outside the standard
+                        * config rom area, something's fishy.
+                        */
+                       goto out;
 
                /* Read header quadlet for the block to get the length. */
-               if (read_rom(device, i, &rom[i]) != RCODE_COMPLETE)
-                       return -1;
+               if (read_rom(device, generation, i, &rom[i]) != RCODE_COMPLETE)
+                       goto out;
                end = i + (rom[i] >> 16) + 1;
                i++;
-               if (end > ARRAY_SIZE(rom))
-                       /* This block extends outside standard config
+               if (end > READ_BIB_ROM_SIZE)
+                       /*
+                        * This block extends outside standard config
                         * area (and the array we're reading it
                         * into).  That's broken, so ignore this
-                        * device. */
-                       return -1;
+                        * device.
+                        */
+                       goto out;
 
-               /* Now read in the block.  If this is a directory
+               /*
+                * Now read in the block.  If this is a directory
                 * block, check the entries as we read them to see if
-                * it references another block, and push it in that case. */
+                * it references another block, and push it in that case.
+                */
                while (i < end) {
-                       if (read_rom(device, i, &rom[i]) != RCODE_COMPLETE)
-                               return -1;
+                       if (read_rom(device, generation, i, &rom[i]) !=
+                           RCODE_COMPLETE)
+                               goto out;
                        if ((key >> 30) == 3 && (rom[i] >> 30) > 1 &&
-                           sp < ARRAY_SIZE(stack))
+                           sp < READ_BIB_STACK_SIZE)
                                stack[sp++] = i + rom[i];
                        i++;
                }
@@ -354,13 +506,23 @@ static int read_bus_info_block(struct fw_device *device)
                        length = i;
        }
 
-       device->config_rom = kmalloc(length * 4, GFP_KERNEL);
-       if (device->config_rom == NULL)
-               return -1;
-       memcpy(device->config_rom, rom, length * 4);
+       old_rom = device->config_rom;
+       new_rom = kmemdup(rom, length * 4, GFP_KERNEL);
+       if (new_rom == NULL)
+               goto out;
+
+       down_write(&fw_device_rwsem);
+       device->config_rom = new_rom;
        device->config_rom_length = length;
+       up_write(&fw_device_rwsem);
 
-       return 0;
+       kfree(old_rom);
+       ret = 0;
+       device->cmc = rom[2] >> 30 & 1;
+ out:
+       kfree(rom);
+
+       return ret;
 }
 
 static void fw_unit_release(struct device *dev)
@@ -370,9 +532,14 @@ static void fw_unit_release(struct device *dev)
        kfree(unit);
 }
 
+static struct device_type fw_unit_type = {
+       .uevent         = fw_unit_uevent,
+       .release        = fw_unit_release,
+};
+
 static int is_fw_unit(struct device *dev)
 {
-       return dev->release == fw_unit_release;
+       return dev->type == &fw_unit_type;
 }
 
 static void create_units(struct fw_device *device)
@@ -387,9 +554,11 @@ static void create_units(struct fw_device *device)
                if (key != (CSR_UNIT | CSR_DIRECTORY))
                        continue;
 
-               /* Get the address of the unit directory and try to
-                * match the drivers id_tables against it. */
-               unit = kzalloc(sizeof *unit, GFP_KERNEL);
+               /*
+                * Get the address of the unit directory and try to
+                * match the drivers id_tables against it.
+                */
+               unit = kzalloc(sizeof(*unit), GFP_KERNEL);
                if (unit == NULL) {
                        fw_error("failed to allocate memory for unit\n");
                        continue;
@@ -397,189 +566,451 @@ static void create_units(struct fw_device *device)
 
                unit->directory = ci.p + value - 1;
                unit->device.bus = &fw_bus_type;
-               unit->device.release = fw_unit_release;
+               unit->device.type = &fw_unit_type;
                unit->device.parent = &device->device;
-               snprintf(unit->device.bus_id, sizeof unit->device.bus_id,
-                        "%s.%d", device->device.bus_id, i++);
+               dev_set_name(&unit->device, "%s.%d", dev_name(&device->device), i++);
 
-               if (device_register(&unit->device) < 0) {
-                       kfree(unit);
-                       continue;
-               }
+               init_fw_attribute_group(&unit->device,
+                                       fw_unit_attributes,
+                                       &unit->attribute_group);
+               if (device_register(&unit->device) < 0)
+                       goto skip_unit;
 
-               if (device_create_file(&unit->device, &modalias_attribute) < 0) {
-                       device_unregister(&unit->device);
-                       kfree(unit);
-               }
+               continue;
 
-               if (device_create_file(&unit->device, &rom_index_attribute) < 0) {
-                       device_unregister(&unit->device);
-                       kfree(unit);
-               }
+       skip_unit:
+               kfree(unit);
        }
 }
 
 static int shutdown_unit(struct device *device, void *data)
 {
-       struct fw_unit *unit = fw_unit(device);
-
-       if (is_fw_unit(device)) {
-               device_remove_file(&unit->device, &modalias_attribute);
-               device_unregister(&unit->device);
-       }
+       device_unregister(device);
 
        return 0;
 }
 
-static DEFINE_IDR(fw_device_idr);
+/*
+ * fw_device_rwsem acts as dual purpose mutex:
+ *   - serializes accesses to fw_device_idr,
+ *   - serializes accesses to fw_device.config_rom/.config_rom_length and
+ *     fw_unit.directory, unless those accesses happen at safe occasions
+ */
+DECLARE_RWSEM(fw_device_rwsem);
+
+DEFINE_IDR(fw_device_idr);
 int fw_cdev_major;
 
-struct fw_device *fw_device_from_devt(dev_t devt)
+struct fw_device *fw_device_get_by_devt(dev_t devt)
 {
        struct fw_device *device;
 
-       down_read(&fw_bus_type.subsys.rwsem);
+       down_read(&fw_device_rwsem);
        device = idr_find(&fw_device_idr, MINOR(devt));
-       up_read(&fw_bus_type.subsys.rwsem);
+       if (device)
+               fw_device_get(device);
+       up_read(&fw_device_rwsem);
 
        return device;
 }
 
+/*
+ * These defines control the retry behavior for reading the config
+ * rom.  It shouldn't be necessary to tweak these; if the device
+ * doesn't respond to a config rom read within 10 seconds, it's not
+ * going to respond at all.  As for the initial delay, a lot of
+ * devices will be able to respond within half a second after bus
+ * reset.  On the other hand, it's not really worth being more
+ * aggressive than that, since it scales pretty well; if 10 devices
+ * are plugged in, they're all getting read within one second.
+ */
+
+#define MAX_RETRIES    10
+#define RETRY_DELAY    (3 * HZ)
+#define INITIAL_DELAY  (HZ / 2)
+#define SHUTDOWN_DELAY (2 * HZ)
+
 static void fw_device_shutdown(struct work_struct *work)
 {
        struct fw_device *device =
                container_of(work, struct fw_device, work.work);
        int minor = MINOR(device->device.devt);
 
-       down_write(&fw_bus_type.subsys.rwsem);
-       idr_remove(&fw_device_idr, minor);
-       up_write(&fw_bus_type.subsys.rwsem);
+       if (time_is_after_jiffies(device->card->reset_jiffies + SHUTDOWN_DELAY)
+           && !list_empty(&device->card->link)) {
+               schedule_delayed_work(&device->work, SHUTDOWN_DELAY);
+               return;
+       }
+
+       if (atomic_cmpxchg(&device->state,
+                          FW_DEVICE_GONE,
+                          FW_DEVICE_SHUTDOWN) != FW_DEVICE_GONE)
+               return;
 
        fw_device_cdev_remove(device);
-       device_remove_file(&device->device, &config_rom_attribute);
        device_for_each_child(&device->device, NULL, shutdown_unit);
        device_unregister(&device->device);
+
+       down_write(&fw_device_rwsem);
+       idr_remove(&fw_device_idr, minor);
+       up_write(&fw_device_rwsem);
+
+       fw_device_put(device);
 }
 
-/* These defines control the retry behavior for reading the config
- * rom.  It shouldn't be necessary to tweak these; if the device
- * doesn't respond to a config rom read within 10 seconds, it's not
- * going to respond at all.  As for the initial delay, a lot of
- * devices will be able to respond within half a second after bus
- * reset.  On the other hand, it's not really worth being more
- * aggressive than that, since it scales pretty well; if 10 devices
- * are plugged in, they're all getting read within one second. */
+static void fw_device_release(struct device *dev)
+{
+       struct fw_device *device = fw_device(dev);
+       struct fw_card *card = device->card;
+       unsigned long flags;
 
-#define MAX_RETRIES    5
-#define RETRY_DELAY    (2 * HZ)
-#define INITIAL_DELAY  (HZ / 2)
+       /*
+        * Take the card lock so we don't set this to NULL while a
+        * FW_NODE_UPDATED callback is being handled or while the
+        * bus manager work looks at this node.
+        */
+       spin_lock_irqsave(&card->lock, flags);
+       device->node->data = NULL;
+       spin_unlock_irqrestore(&card->lock, flags);
+
+       fw_node_put(device->node);
+       kfree(device->config_rom);
+       kfree(device);
+       fw_card_put(card);
+}
+
+static struct device_type fw_device_type = {
+       .release = fw_device_release,
+};
+
+static int update_unit(struct device *dev, void *data)
+{
+       struct fw_unit *unit = fw_unit(dev);
+       struct fw_driver *driver = (struct fw_driver *)dev->driver;
+
+       if (is_fw_unit(dev) && driver != NULL && driver->update != NULL) {
+               down(&dev->sem);
+               driver->update(unit);
+               up(&dev->sem);
+       }
+
+       return 0;
+}
+
+static void fw_device_update(struct work_struct *work)
+{
+       struct fw_device *device =
+               container_of(work, struct fw_device, work.work);
+
+       fw_device_cdev_update(device);
+       device_for_each_child(&device->device, NULL, update_unit);
+}
+
+/*
+ * If a device was pending for deletion because its node went away but its
+ * bus info block and root directory header matches that of a newly discovered
+ * device, revive the existing fw_device.
+ * The newly allocated fw_device becomes obsolete instead.
+ */
+static int lookup_existing_device(struct device *dev, void *data)
+{
+       struct fw_device *old = fw_device(dev);
+       struct fw_device *new = data;
+       struct fw_card *card = new->card;
+       int match = 0;
+
+       down_read(&fw_device_rwsem); /* serialize config_rom access */
+       spin_lock_irq(&card->lock);  /* serialize node access */
+
+       if (memcmp(old->config_rom, new->config_rom, 6 * 4) == 0 &&
+           atomic_cmpxchg(&old->state,
+                          FW_DEVICE_GONE,
+                          FW_DEVICE_RUNNING) == FW_DEVICE_GONE) {
+               struct fw_node *current_node = new->node;
+               struct fw_node *obsolete_node = old->node;
+
+               new->node = obsolete_node;
+               new->node->data = new;
+               old->node = current_node;
+               old->node->data = old;
+
+               old->max_speed = new->max_speed;
+               old->node_id = current_node->node_id;
+               smp_wmb();  /* update node_id before generation */
+               old->generation = card->generation;
+               old->config_rom_retries = 0;
+               fw_notify("rediscovered device %s\n", dev_name(dev));
+
+               PREPARE_DELAYED_WORK(&old->work, fw_device_update);
+               schedule_delayed_work(&old->work, 0);
+
+               if (current_node == card->root_node)
+                       fw_schedule_bm_work(card, 0);
+
+               match = 1;
+       }
+
+       spin_unlock_irq(&card->lock);
+       up_read(&fw_device_rwsem);
+
+       return match;
+}
+
+enum { BC_UNKNOWN = 0, BC_UNIMPLEMENTED, BC_IMPLEMENTED, };
+
+void fw_device_set_broadcast_channel(struct fw_device *device, int generation)
+{
+       struct fw_card *card = device->card;
+       __be32 data;
+       int rcode;
+
+       if (!card->broadcast_channel_allocated)
+               return;
+
+       if (device->bc_implemented == BC_UNKNOWN) {
+               rcode = fw_run_transaction(card, TCODE_READ_QUADLET_REQUEST,
+                               device->node_id, generation, device->max_speed,
+                               CSR_REGISTER_BASE + CSR_BROADCAST_CHANNEL,
+                               &data, 4);
+               switch (rcode) {
+               case RCODE_COMPLETE:
+                       if (data & cpu_to_be32(1 << 31)) {
+                               device->bc_implemented = BC_IMPLEMENTED;
+                               break;
+                       }
+                       /* else fall through to case address error */
+               case RCODE_ADDRESS_ERROR:
+                       device->bc_implemented = BC_UNIMPLEMENTED;
+               }
+       }
+
+       if (device->bc_implemented == BC_IMPLEMENTED) {
+               data = cpu_to_be32(BROADCAST_CHANNEL_INITIAL |
+                                  BROADCAST_CHANNEL_VALID);
+               fw_run_transaction(card, TCODE_WRITE_QUADLET_REQUEST,
+                               device->node_id, generation, device->max_speed,
+                               CSR_REGISTER_BASE + CSR_BROADCAST_CHANNEL,
+                               &data, 4);
+       }
+}
 
 static void fw_device_init(struct work_struct *work)
 {
        struct fw_device *device =
                container_of(work, struct fw_device, work.work);
-       int minor, err;
+       struct device *revived_dev;
+       int minor, ret;
 
-       /* All failure paths here set node->data to NULL, so that we
+       /*
+        * All failure paths here set node->data to NULL, so that we
         * don't try to do device_for_each_child() on a kfree()'d
-        * device. */
+        * device.
+        */
 
-       if (read_bus_info_block(device) < 0) {
-               if (device->config_rom_retries < MAX_RETRIES) {
+       if (read_bus_info_block(device, device->generation) < 0) {
+               if (device->config_rom_retries < MAX_RETRIES &&
+                   atomic_read(&device->state) == FW_DEVICE_INITIALIZING) {
                        device->config_rom_retries++;
                        schedule_delayed_work(&device->work, RETRY_DELAY);
                } 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_schedule_bm_work(device->card, 0);
                        fw_device_release(&device->device);
                }
                return;
        }
 
-       err = -ENOMEM;
-       down_write(&fw_bus_type.subsys.rwsem);
-       if (idr_pre_get(&fw_device_idr, GFP_KERNEL))
-               err = idr_get_new(&fw_device_idr, device, &minor);
-       up_write(&fw_bus_type.subsys.rwsem);
-       if (err < 0)
+       revived_dev = device_find_child(device->card->device,
+                                       device, lookup_existing_device);
+       if (revived_dev) {
+               put_device(revived_dev);
+               fw_device_release(&device->device);
+
+               return;
+       }
+
+       device_initialize(&device->device);
+
+       fw_device_get(device);
+       down_write(&fw_device_rwsem);
+       ret = idr_pre_get(&fw_device_idr, GFP_KERNEL) ?
+             idr_get_new(&fw_device_idr, device, &minor) :
+             -ENOMEM;
+       up_write(&fw_device_rwsem);
+
+       if (ret < 0)
                goto error;
 
        device->device.bus = &fw_bus_type;
-       device->device.release = fw_device_release;
+       device->device.type = &fw_device_type;
        device->device.parent = device->card->device;
        device->device.devt = MKDEV(fw_cdev_major, minor);
-       snprintf(device->device.bus_id, sizeof device->device.bus_id,
-                "fw%d", minor);
+       dev_set_name(&device->device, "fw%d", minor);
 
+       init_fw_attribute_group(&device->device,
+                               fw_device_attributes,
+                               &device->attribute_group);
        if (device_add(&device->device)) {
                fw_error("Failed to add device.\n");
                goto error_with_cdev;
        }
 
-       if (device_create_file(&device->device, &config_rom_attribute) < 0) {
-               fw_error("Failed to create config rom file.\n");
-               goto error_with_device;
-       }
-
        create_units(device);
 
-       /* Transition the device to running state.  If it got pulled
+       /*
+        * Transition the device to running state.  If it got pulled
         * out from under us while we did the intialization work, we
         * have to shut down the device again here.  Normally, though,
         * fw_node_event will be responsible for shutting it down when
         * necessary.  We have to use the atomic cmpxchg here to avoid
         * racing with the FW_NODE_DESTROYED case in
-        * fw_node_event(). */
+        * fw_node_event().
+        */
        if (atomic_cmpxchg(&device->state,
-                   FW_DEVICE_INITIALIZING,
-                   FW_DEVICE_RUNNING) == FW_DEVICE_SHUTDOWN)
-               fw_device_shutdown(&device->work.work);
-       else
-               fw_notify("created new fw device %s (%d config rom retries)\n",
-                         device->device.bus_id, device->config_rom_retries);
+                          FW_DEVICE_INITIALIZING,
+                          FW_DEVICE_RUNNING) == FW_DEVICE_GONE) {
+               PREPARE_DELAYED_WORK(&device->work, fw_device_shutdown);
+               schedule_delayed_work(&device->work, SHUTDOWN_DELAY);
+       } else {
+               if (device->config_rom_retries)
+                       fw_notify("created device %s: GUID %08x%08x, S%d00, "
+                                 "%d config ROM retries\n",
+                                 dev_name(&device->device),
+                                 device->config_rom[3], device->config_rom[4],
+                                 1 << device->max_speed,
+                                 device->config_rom_retries);
+               else
+                       fw_notify("created device %s: GUID %08x%08x, S%d00\n",
+                                 dev_name(&device->device),
+                                 device->config_rom[3], device->config_rom[4],
+                                 1 << device->max_speed);
+               device->config_rom_retries = 0;
+
+               fw_device_set_broadcast_channel(device, device->generation);
+       }
 
-       /* Reschedule the IRM work if we just finished reading the
+       /*
+        * Reschedule the IRM work if we just finished reading the
         * root node config rom.  If this races with a bus reset we
         * just end up running the IRM work a couple of extra times -
-        * pretty harmless. */
+        * pretty harmless.
+        */
        if (device->node == device->card->root_node)
-               schedule_delayed_work(&device->card->work, 0);
+               fw_schedule_bm_work(device->card, 0);
 
        return;
 
- error_with_device:
-       device_del(&device->device);
  error_with_cdev:
-       down_write(&fw_bus_type.subsys.rwsem);
+       down_write(&fw_device_rwsem);
        idr_remove(&fw_device_idr, minor);
-       up_write(&fw_bus_type.subsys.rwsem);
+       up_write(&fw_device_rwsem);
  error:
-       put_device(&device->device);
+       fw_device_put(device);          /* fw_device_idr's reference */
+
+       put_device(&device->device);    /* our reference */
 }
 
-static int update_unit(struct device *dev, void *data)
+enum {
+       REREAD_BIB_ERROR,
+       REREAD_BIB_GONE,
+       REREAD_BIB_UNCHANGED,
+       REREAD_BIB_CHANGED,
+};
+
+/* Reread and compare bus info block and header of root directory */
+static int reread_bus_info_block(struct fw_device *device, int generation)
 {
-       struct fw_unit *unit = fw_unit(dev);
-       struct fw_driver *driver = (struct fw_driver *)dev->driver;
+       u32 q;
+       int i;
 
-       if (is_fw_unit(dev) && driver != NULL && driver->update != NULL) {
-               down(&dev->sem);
-               driver->update(unit);
-               up(&dev->sem);
+       for (i = 0; i < 6; i++) {
+               if (read_rom(device, generation, i, &q) != RCODE_COMPLETE)
+                       return REREAD_BIB_ERROR;
+
+               if (i == 0 && q == 0)
+                       return REREAD_BIB_GONE;
+
+               if (q != device->config_rom[i])
+                       return REREAD_BIB_CHANGED;
        }
 
-       return 0;
+       return REREAD_BIB_UNCHANGED;
 }
 
-static void fw_device_update(struct work_struct *work)
+static void fw_device_refresh(struct work_struct *work)
 {
        struct fw_device *device =
                container_of(work, struct fw_device, work.work);
+       struct fw_card *card = device->card;
+       int node_id = device->node_id;
 
-       fw_device_cdev_update(device);
-       device_for_each_child(&device->device, NULL, update_unit);
+       switch (reread_bus_info_block(device, device->generation)) {
+       case REREAD_BIB_ERROR:
+               if (device->config_rom_retries < MAX_RETRIES / 2 &&
+                   atomic_read(&device->state) == FW_DEVICE_INITIALIZING) {
+                       device->config_rom_retries++;
+                       schedule_delayed_work(&device->work, RETRY_DELAY / 2);
+
+                       return;
+               }
+               goto give_up;
+
+       case REREAD_BIB_GONE:
+               goto gone;
+
+       case REREAD_BIB_UNCHANGED:
+               if (atomic_cmpxchg(&device->state,
+                                  FW_DEVICE_INITIALIZING,
+                                  FW_DEVICE_RUNNING) == FW_DEVICE_GONE)
+                       goto gone;
+
+               fw_device_update(work);
+               device->config_rom_retries = 0;
+               goto out;
+
+       case REREAD_BIB_CHANGED:
+               break;
+       }
+
+       /*
+        * Something changed.  We keep things simple and don't investigate
+        * further.  We just destroy all previous units and create new ones.
+        */
+       device_for_each_child(&device->device, NULL, shutdown_unit);
+
+       if (read_bus_info_block(device, device->generation) < 0) {
+               if (device->config_rom_retries < MAX_RETRIES &&
+                   atomic_read(&device->state) == FW_DEVICE_INITIALIZING) {
+                       device->config_rom_retries++;
+                       schedule_delayed_work(&device->work, RETRY_DELAY);
+
+                       return;
+               }
+               goto give_up;
+       }
+
+       create_units(device);
+
+       if (atomic_cmpxchg(&device->state,
+                          FW_DEVICE_INITIALIZING,
+                          FW_DEVICE_RUNNING) == FW_DEVICE_GONE)
+               goto gone;
+
+       fw_notify("refreshed device %s\n", dev_name(&device->device));
+       device->config_rom_retries = 0;
+       goto out;
+
+ give_up:
+       fw_notify("giving up on refresh of device %s\n", dev_name(&device->device));
+ gone:
+       atomic_set(&device->state, FW_DEVICE_GONE);
+       PREPARE_DELAYED_WORK(&device->work, fw_device_shutdown);
+       schedule_delayed_work(&device->work, SHUTDOWN_DELAY);
+ out:
+       if (node_id == card->root_node->node_id)
+               fw_schedule_bm_work(card, 0);
 }
 
 void fw_node_event(struct fw_card *card, struct fw_node *node, int event)
@@ -591,44 +1022,70 @@ void fw_node_event(struct fw_card *card, struct fw_node *node, int event)
        case FW_NODE_LINK_ON:
                if (!node->link_on)
                        break;
-
+ create:
                device = kzalloc(sizeof(*device), GFP_ATOMIC);
                if (device == NULL)
                        break;
 
-               /* Do minimal intialization of the device here, the
-                * rest will happen in fw_device_init().  We need the
-                * card and node so we can read the config rom and we
-                * need to do device_initialize() now so
-                * device_for_each_child() in FW_NODE_UPDATED is
-                * doesn't freak out. */
-               device_initialize(&device->device);
+               /*
+                * Do minimal intialization of the device here, the
+                * rest will happen in fw_device_init().
+                *
+                * Attention:  A lot of things, even fw_device_get(),
+                * cannot be done before fw_device_init() finished!
+                * You can basically just check device->state and
+                * schedule work until then, but only while holding
+                * card->lock.
+                */
                atomic_set(&device->state, FW_DEVICE_INITIALIZING);
                device->card = fw_card_get(card);
                device->node = fw_node_get(node);
                device->node_id = node->node_id;
                device->generation = card->generation;
+               mutex_init(&device->client_list_mutex);
                INIT_LIST_HEAD(&device->client_list);
 
-               /* Set the node data to point back to this device so
+               /*
+                * Set the node data to point back to this device so
                 * FW_NODE_UPDATED callbacks can update the node_id
-                * and generation for the device. */
+                * and generation for the device.
+                */
                node->data = device;
 
-               /* Many devices are slow to respond after bus resets,
+               /*
+                * Many devices are slow to respond after bus resets,
                 * especially if they are bus powered and go through
                 * power-up after getting plugged in.  We schedule the
-                * first config rom scan half a second after bus reset. */
+                * first config rom scan half a second after bus reset.
+                */
                INIT_DELAYED_WORK(&device->work, fw_device_init);
                schedule_delayed_work(&device->work, INITIAL_DELAY);
                break;
 
+       case FW_NODE_INITIATED_RESET:
+               device = node->data;
+               if (device == NULL)
+                       goto create;
+
+               device->node_id = node->node_id;
+               smp_wmb();  /* update node_id before generation */
+               device->generation = card->generation;
+               if (atomic_cmpxchg(&device->state,
+                           FW_DEVICE_RUNNING,
+                           FW_DEVICE_INITIALIZING) == FW_DEVICE_RUNNING) {
+                       PREPARE_DELAYED_WORK(&device->work, fw_device_refresh);
+                       schedule_delayed_work(&device->work,
+                               node == card->local_node ? 0 : INITIAL_DELAY);
+               }
+               break;
+
        case FW_NODE_UPDATED:
                if (!node->link_on || node->data == NULL)
                        break;
 
                device = node->data;
                device->node_id = node->node_id;
+               smp_wmb();  /* update node_id before generation */
                device->generation = card->generation;
                if (atomic_read(&device->state) == FW_DEVICE_RUNNING) {
                        PREPARE_DELAYED_WORK(&device->work, fw_device_update);
@@ -641,7 +1098,8 @@ void fw_node_event(struct fw_card *card, struct fw_node *node, int event)
                if (!node->data)
                        break;
 
-               /* Destroy the device associated with the node.  There
+               /*
+                * Destroy the device associated with the node.  There
                 * are two cases here: either the device is fully
                 * initialized (FW_DEVICE_RUNNING) or we're in the
                 * process of reading its config rom
@@ -650,12 +1108,14 @@ void fw_node_event(struct fw_card *card, struct fw_node *node, int event)
                 * full fw_device_shutdown().  If not, there's work
                 * scheduled to read it's config rom, and we just put
                 * the device in shutdown state to have that code fail
-                * to create the device. */
+                * to create the device.
+                */
                device = node->data;
                if (atomic_xchg(&device->state,
-                               FW_DEVICE_SHUTDOWN) == FW_DEVICE_RUNNING) {
+                               FW_DEVICE_GONE) == FW_DEVICE_RUNNING) {
                        PREPARE_DELAYED_WORK(&device->work, fw_device_shutdown);
-                       schedule_delayed_work(&device->work, 0);
+                       schedule_delayed_work(&device->work,
+                               list_empty(&card->link) ? 0 : SHUTDOWN_DELAY);
                }
                break;
        }