ACPI / IPMI: Add reference counting for ACPI IPMI transfers
authorLv Zheng <lv.zheng@intel.com>
Fri, 13 Sep 2013 05:14:11 +0000 (13:14 +0800)
committerRafael J. Wysocki <rafael.j.wysocki@intel.com>
Mon, 30 Sep 2013 17:46:12 +0000 (19:46 +0200)
This patch adds reference counting for ACPI IPMI transfers to tune the
locking granularity of tx_msg_lock.

This patch also makes the whole acpi_ipmi module's coding style consistent
by using reference counting for all its objects (i.e., acpi_ipmi_device and
acpi_ipmi_msg).

The acpi_ipmi_msg handling is re-designed using referece counting.
 1. tx_msg is always unlinked before complete(), so that it is safe to put
    complete() out side of tx_msg_lock.
 2. tx_msg reference counters are incremented before calling
    ipmi_request_settime() and tx_msg_lock protection is added to
    ipmi_cancel_tx_msg() so that a complete() can be safely called in
    parellel with tx_msg unlinking in failure cases.
 3. tx_msg holds a reference to acpi_ipmi_device so that it can be flushed
    and freed in the contexts other than acpi_ipmi_space_handler().

The lockdep_chains shows all acpi_ipmi locks are leaf locks after the
tuning:
 1. ipmi_lock is always leaf:
    irq_context: 0
    [ffffffff81a943f8] smi_watchers_mutex
    [ffffffffa06eca60] driver_data.ipmi_lock
    irq_context: 0
    [ffffffff82767b40] &buffer->mutex
    [ffffffffa00a6678] s_active#103
    [ffffffffa06eca60] driver_data.ipmi_lock
 2. without this patch applied, lock used by complete() is held after
    holding tx_msg_lock:
    irq_context: 0
    [ffffffff82767b40] &buffer->mutex
    [ffffffffa00a6678] s_active#103
    [ffffffffa06ecce8] &(&ipmi_device->tx_msg_lock)->rlock
    irq_context: 1
    [ffffffffa06ecce8] &(&ipmi_device->tx_msg_lock)->rlock
    irq_context: 1
    [ffffffffa06ecce8] &(&ipmi_device->tx_msg_lock)->rlock
    [ffffffffa06eccf0] &x->wait#25
    irq_context: 1
    [ffffffffa06ecce8] &(&ipmi_device->tx_msg_lock)->rlock
    [ffffffffa06eccf0] &x->wait#25
    [ffffffff81e36620] &p->pi_lock
    irq_context: 1
    [ffffffffa06ecce8] &(&ipmi_device->tx_msg_lock)->rlock
    [ffffffffa06eccf0] &x->wait#25
    [ffffffff81e36620] &p->pi_lock
    [ffffffff81e5d0a8] &rq->lock
 3. with this patch applied, tx_msg_lock is always leaf:
    irq_context: 0
    [ffffffff82767b40] &buffer->mutex
    [ffffffffa00a66d8] s_active#107
    [ffffffffa07ecdc8] &(&ipmi_device->tx_msg_lock)->rlock
    irq_context: 1
    [ffffffffa07ecdc8] &(&ipmi_device->tx_msg_lock)->rlock

Signed-off-by: Lv Zheng <lv.zheng@intel.com>
Reviewed-by: Huang Ying <ying.huang@intel.com>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
drivers/acpi/acpi_ipmi.c

index 7ec4cd1..b9da5ef 100644 (file)
@@ -104,6 +104,7 @@ struct acpi_ipmi_msg {
        u8      data[ACPI_IPMI_MAX_MSG_LENGTH];
        u8      rx_len;
        struct acpi_ipmi_device *device;
+       struct kref     kref;
 };
 
 /* IPMI request/response buffer per ACPI 4.0, sec 5.5.2.4.3.2 */
@@ -208,16 +209,20 @@ static void acpi_ipmi_dev_put(struct acpi_ipmi_device *ipmi_device)
        kref_put(&ipmi_device->kref, ipmi_dev_release_kref);
 }
 
-static struct acpi_ipmi_msg *acpi_alloc_ipmi_msg(struct acpi_ipmi_device *ipmi)
+static struct acpi_ipmi_msg *ipmi_msg_alloc(void)
 {
+       struct acpi_ipmi_device *ipmi;
        struct acpi_ipmi_msg *ipmi_msg;
-       struct pnp_dev *pnp_dev = ipmi->pnp_dev;
 
+       ipmi = acpi_ipmi_dev_get();
+       if (!ipmi)
+               return NULL;
        ipmi_msg = kzalloc(sizeof(struct acpi_ipmi_msg), GFP_KERNEL);
-       if (!ipmi_msg)  {
-               dev_warn(&pnp_dev->dev, "Can't allocate memory for ipmi_msg\n");
+       if (!ipmi_msg) {
+               acpi_ipmi_dev_put(ipmi);
                return NULL;
        }
+       kref_init(&ipmi_msg->kref);
        init_completion(&ipmi_msg->tx_complete);
        INIT_LIST_HEAD(&ipmi_msg->head);
        ipmi_msg->device = ipmi;
@@ -225,6 +230,32 @@ static struct acpi_ipmi_msg *acpi_alloc_ipmi_msg(struct acpi_ipmi_device *ipmi)
        return ipmi_msg;
 }
 
+static void ipmi_msg_release(struct acpi_ipmi_msg *tx_msg)
+{
+       acpi_ipmi_dev_put(tx_msg->device);
+       kfree(tx_msg);
+}
+
+static void ipmi_msg_release_kref(struct kref *kref)
+{
+       struct acpi_ipmi_msg *tx_msg =
+               container_of(kref, struct acpi_ipmi_msg, kref);
+
+       ipmi_msg_release(tx_msg);
+}
+
+static struct acpi_ipmi_msg *acpi_ipmi_msg_get(struct acpi_ipmi_msg *tx_msg)
+{
+       kref_get(&tx_msg->kref);
+
+       return tx_msg;
+}
+
+static void acpi_ipmi_msg_put(struct acpi_ipmi_msg *tx_msg)
+{
+       kref_put(&tx_msg->kref, ipmi_msg_release_kref);
+}
+
 #define                IPMI_OP_RGN_NETFN(offset)       ((offset >> 8) & 0xff)
 #define                IPMI_OP_RGN_CMD(offset)         (offset & 0xff)
 static int acpi_format_ipmi_request(struct acpi_ipmi_msg *tx_msg,
@@ -305,7 +336,7 @@ static void acpi_format_ipmi_response(struct acpi_ipmi_msg *msg,
 
 static void ipmi_flush_tx_msg(struct acpi_ipmi_device *ipmi)
 {
-       struct acpi_ipmi_msg *tx_msg, *temp;
+       struct acpi_ipmi_msg *tx_msg;
        unsigned long flags;
 
        /*
@@ -317,18 +348,47 @@ static void ipmi_flush_tx_msg(struct acpi_ipmi_device *ipmi)
         * ipmi_recv_msg(s) are freed after invoking ipmi_destroy_user().
         */
        spin_lock_irqsave(&ipmi->tx_msg_lock, flags);
-       list_for_each_entry_safe(tx_msg, temp, &ipmi->tx_msg_list, head) {
+       while (!list_empty(&ipmi->tx_msg_list)) {
+               tx_msg = list_first_entry(&ipmi->tx_msg_list,
+                                         struct acpi_ipmi_msg,
+                                         head);
+               list_del(&tx_msg->head);
+               spin_unlock_irqrestore(&ipmi->tx_msg_lock, flags);
+
                /* wake up the sleep thread on the Tx msg */
                complete(&tx_msg->tx_complete);
+               acpi_ipmi_msg_put(tx_msg);
+               spin_lock_irqsave(&ipmi->tx_msg_lock, flags);
+       }
+       spin_unlock_irqrestore(&ipmi->tx_msg_lock, flags);
+}
+
+static void ipmi_cancel_tx_msg(struct acpi_ipmi_device *ipmi,
+                              struct acpi_ipmi_msg *msg)
+{
+       struct acpi_ipmi_msg *tx_msg, *temp;
+       bool msg_found = false;
+       unsigned long flags;
+
+       spin_lock_irqsave(&ipmi->tx_msg_lock, flags);
+       list_for_each_entry_safe(tx_msg, temp, &ipmi->tx_msg_list, head) {
+               if (msg == tx_msg) {
+                       msg_found = true;
+                       list_del(&tx_msg->head);
+                       break;
+               }
        }
        spin_unlock_irqrestore(&ipmi->tx_msg_lock, flags);
+
+       if (msg_found)
+               acpi_ipmi_msg_put(tx_msg);
 }
 
 static void ipmi_msg_handler(struct ipmi_recv_msg *msg, void *user_msg_data)
 {
        struct acpi_ipmi_device *ipmi_device = user_msg_data;
-       int msg_found = 0;
-       struct acpi_ipmi_msg *tx_msg;
+       bool msg_found = false;
+       struct acpi_ipmi_msg *tx_msg, *temp;
        struct pnp_dev *pnp_dev = ipmi_device->pnp_dev;
        unsigned long flags;
 
@@ -339,17 +399,19 @@ static void ipmi_msg_handler(struct ipmi_recv_msg *msg, void *user_msg_data)
                goto out_msg;
        }
        spin_lock_irqsave(&ipmi_device->tx_msg_lock, flags);
-       list_for_each_entry(tx_msg, &ipmi_device->tx_msg_list, head) {
+       list_for_each_entry_safe(tx_msg, temp, &ipmi_device->tx_msg_list, head) {
                if (msg->msgid == tx_msg->tx_msgid) {
-                       msg_found = 1;
+                       msg_found = true;
+                       list_del(&tx_msg->head);
                        break;
                }
        }
+       spin_unlock_irqrestore(&ipmi_device->tx_msg_lock, flags);
 
        if (!msg_found) {
                dev_warn(&pnp_dev->dev, "Unexpected response (msg id %ld) is "
                        "returned.\n", msg->msgid);
-               goto out_lock;
+               goto out_msg;
        }
 
        /* copy the response data to Rx_data buffer */
@@ -375,8 +437,7 @@ static void ipmi_msg_handler(struct ipmi_recv_msg *msg, void *user_msg_data)
        tx_msg->msg_done = ACPI_IPMI_OK;
 out_comp:
        complete(&tx_msg->tx_complete);
-out_lock:
-       spin_unlock_irqrestore(&ipmi_device->tx_msg_lock, flags);
+       acpi_ipmi_msg_put(tx_msg);
 out_msg:
        ipmi_free_recv_msg(msg);
 };
@@ -491,26 +552,23 @@ acpi_ipmi_space_handler(u32 function, acpi_physical_address address,
        if ((function & ACPI_IO_MASK) == ACPI_READ)
                return AE_TYPE;
 
-       ipmi_device = acpi_ipmi_dev_get();
-       if (!ipmi_device)
+       tx_msg = ipmi_msg_alloc();
+       if (!tx_msg)
                return AE_NOT_EXIST;
 
-       tx_msg = acpi_alloc_ipmi_msg(ipmi_device);
-       if (!tx_msg) {
-               status = AE_NO_MEMORY;
-               goto out_ref;
-       }
+       ipmi_device = tx_msg->device;
 
        if (acpi_format_ipmi_request(tx_msg, address, value) != 0) {
-               status = AE_TYPE;
-               goto out_msg;
+               ipmi_msg_release(tx_msg);
+               return AE_TYPE;
        }
+       acpi_ipmi_msg_get(tx_msg);
        mutex_lock(&driver_data.ipmi_lock);
        /* Do not add a tx_msg that can not be flushed. */
        if (ipmi_device->dead) {
-               status = AE_NOT_EXIST;
                mutex_unlock(&driver_data.ipmi_lock);
-               goto out_msg;
+               ipmi_msg_release(tx_msg);
+               return AE_NOT_EXIST;
        }
        spin_lock_irqsave(&ipmi_device->tx_msg_lock, flags);
        list_add_tail(&tx_msg->head, &ipmi_device->tx_msg_list);
@@ -523,20 +581,15 @@ acpi_ipmi_space_handler(u32 function, acpi_physical_address address,
                                        NULL, 0, 0, IPMI_TIMEOUT);
        if (err) {
                status = AE_ERROR;
-               goto out_list;
+               goto out_msg;
        }
        wait_for_completion(&tx_msg->tx_complete);
        acpi_format_ipmi_response(tx_msg, value);
        status = AE_OK;
 
-out_list:
-       spin_lock_irqsave(&ipmi_device->tx_msg_lock, flags);
-       list_del(&tx_msg->head);
-       spin_unlock_irqrestore(&ipmi_device->tx_msg_lock, flags);
 out_msg:
-       kfree(tx_msg);
-out_ref:
-       acpi_ipmi_dev_put(ipmi_device);
+       ipmi_cancel_tx_msg(ipmi_device, tx_msg);
+       acpi_ipmi_msg_put(tx_msg);
        return status;
 }