Update to 3.4-final.
[linux-flexiantxendom0-3.2.10.git] / block / blk-timeout.c
index eba7335..7803548 100644 (file)
@@ -4,41 +4,80 @@
 #include <linux/kernel.h>
 #include <linux/module.h>
 #include <linux/blkdev.h>
+#include <linux/fault-inject.h>
 
 #include "blk.h"
 
+#ifdef CONFIG_FAIL_IO_TIMEOUT
+
+static DECLARE_FAULT_ATTR(fail_io_timeout);
+
+static int __init setup_fail_io_timeout(char *str)
+{
+       return setup_fault_attr(&fail_io_timeout, str);
+}
+__setup("fail_io_timeout=", setup_fail_io_timeout);
+
+int blk_should_fake_timeout(struct request_queue *q)
+{
+       if (!test_bit(QUEUE_FLAG_FAIL_IO, &q->queue_flags))
+               return 0;
+
+       return should_fail(&fail_io_timeout, 1);
+}
+
+static int __init fail_io_timeout_debugfs(void)
+{
+       struct dentry *dir = fault_create_debugfs_attr("fail_io_timeout",
+                                               NULL, &fail_io_timeout);
+
+       return IS_ERR(dir) ? PTR_ERR(dir) : 0;
+}
+
+late_initcall(fail_io_timeout_debugfs);
+
+ssize_t part_timeout_show(struct device *dev, struct device_attribute *attr,
+                         char *buf)
+{
+       struct gendisk *disk = dev_to_disk(dev);
+       int set = test_bit(QUEUE_FLAG_FAIL_IO, &disk->queue->queue_flags);
+
+       return sprintf(buf, "%d\n", set != 0);
+}
+
+ssize_t part_timeout_store(struct device *dev, struct device_attribute *attr,
+                          const char *buf, size_t count)
+{
+       struct gendisk *disk = dev_to_disk(dev);
+       int val;
+
+       if (count) {
+               struct request_queue *q = disk->queue;
+               char *p = (char *) buf;
+
+               val = simple_strtoul(p, &p, 10);
+               spin_lock_irq(q->queue_lock);
+               if (val)
+                       queue_flag_set(QUEUE_FLAG_FAIL_IO, q);
+               else
+                       queue_flag_clear(QUEUE_FLAG_FAIL_IO, q);
+               spin_unlock_irq(q->queue_lock);
+       }
+
+       return count;
+}
+
+#endif /* CONFIG_FAIL_IO_TIMEOUT */
+
 /*
  * blk_delete_timer - Delete/cancel timer for a given function.
  * @req:       request that we are canceling timer for
  *
- * Return value:
- *     1 if we were able to detach the timer.  0 if we blew it, and the
- *     timer function has already started to run. Caller must hold queue lock.
  */
-int blk_delete_timer(struct request *req)
+void blk_delete_timer(struct request *req)
 {
-       struct request_queue *q = req->q;
-
-       /*
-        * Nothing to detach
-        */
-       if (!q->rq_timed_out_fn || !req->deadline)
-               return 1;
-
-       /*
-        * Not on the list, must have already been scheduled (or never added)
-        */
-       if (list_empty(&req->timeout_list))
-               return 0;
-
        list_del_init(&req->timeout_list);
-
-       if (list_empty(&q->timeout_list))
-               del_timer(&q->timeout);
-
-       return 1;
 }
-EXPORT_SYMBOL_GPL(blk_delete_timer);
 
 static void blk_rq_timed_out(struct request *req)
 {
@@ -48,9 +87,10 @@ static void blk_rq_timed_out(struct request *req)
        ret = q->rq_timed_out_fn(req);
        switch (ret) {
        case BLK_EH_HANDLED:
-               blk_complete_request(req);
+               __blk_complete_request(req);
                break;
        case BLK_EH_RESET_TIMER:
+               blk_clear_rq_complete(req);
                blk_add_timer(req);
                break;
        case BLK_EH_NOT_HANDLED:
@@ -70,8 +110,9 @@ static void blk_rq_timed_out(struct request *req)
 void blk_rq_timed_out_timer(unsigned long data)
 {
        struct request_queue *q = (struct request_queue *) data;
-       unsigned long flags, uninitialized_var(next), next_set = 0;
+       unsigned long flags, next = 0;
        struct request *rq, *tmp;
+       int next_set = 0;
 
        spin_lock_irqsave(q->queue_lock, flags);
 
@@ -80,21 +121,19 @@ void blk_rq_timed_out_timer(unsigned long data)
                        list_del_init(&rq->timeout_list);
 
                        /*
-                        * if rq->bio is now NULL, then IO completion did
-                        * run on this request and we simply raced to get here
+                        * Check if we raced with end io completion
                         */
-                       if (rq->bio)
-                               blk_rq_timed_out(rq);
-               }
-               if (!next_set) {
+                       if (blk_mark_rq_complete(rq))
+                               continue;
+                       blk_rq_timed_out(rq);
+               } else if (!next_set || time_after(next, rq->deadline)) {
                        next = rq->deadline;
                        next_set = 1;
-               } else if (time_after(next, rq->deadline))
-                       next = rq->deadline;
+               }
        }
 
-       if (next_set && !list_empty(&q->timeout_list))
-               mod_timer(&q->timeout, round_jiffies(next));
+       if (next_set)
+               mod_timer(&q->timeout, round_jiffies_up(next));
 
        spin_unlock_irqrestore(q->queue_lock, flags);
 }
@@ -110,8 +149,9 @@ void blk_rq_timed_out_timer(unsigned long data)
  */
 void blk_abort_request(struct request *req)
 {
-       if (!blk_delete_timer(req))
+       if (blk_mark_rq_complete(req))
                return;
+       blk_delete_timer(req);
        blk_rq_timed_out(req);
 }
 EXPORT_SYMBOL_GPL(blk_abort_request);
@@ -129,28 +169,28 @@ void blk_add_timer(struct request *req)
        struct request_queue *q = req->q;
        unsigned long expiry;
 
+       if (!q->rq_timed_out_fn)
+               return;
+
        BUG_ON(!list_empty(&req->timeout_list));
+       BUG_ON(test_bit(REQ_ATOM_COMPLETE, &req->atomic_flags));
 
-       if (req->timeout)
-               req->deadline = jiffies + req->timeout;
-       else
-               req->deadline = jiffies + q->rq_timeout;
+       /*
+        * Some LLDs, like scsi, peek at the timeout to prevent a
+        * command from being retried forever.
+        */
+       if (!req->timeout)
+               req->timeout = q->rq_timeout;
 
+       req->deadline = jiffies + req->timeout;
        list_add_tail(&req->timeout_list, &q->timeout_list);
 
        /*
         * If the timer isn't already pending or this timeout is earlier
-        * than an existing one, modify the timer. Round to next nearest
+        * than an existing one, modify the timer. Round up to next nearest
         * second.
         */
-       expiry = round_jiffies(req->deadline);
-
-       /*
-        * We use ->deadline == 0 to detect whether a timer was added or
-        * not, so just increase to next jiffy for that specific case
-        */
-       if (unlikely(!req->deadline))
-               req->deadline = 1;
+       expiry = round_jiffies_up(req->deadline);
 
        if (!timer_pending(&q->timeout) ||
            time_before(expiry, q->timeout.expires))
@@ -166,14 +206,34 @@ void blk_abort_queue(struct request_queue *q)
 {
        unsigned long flags;
        struct request *rq, *tmp;
+       LIST_HEAD(list);
+
+       /*
+        * Not a request based block device, nothing to abort
+        */
+       if (!q->request_fn)
+               return;
 
        spin_lock_irqsave(q->queue_lock, flags);
 
        elv_abort_queue(q);
 
-       list_for_each_entry_safe(rq, tmp, &q->timeout_list, timeout_list)
+       /*
+        * Splice entries to local list, to avoid deadlocking if entries
+        * get readded to the timeout list by error handling
+        */
+       list_splice_init(&q->timeout_list, &list);
+
+       list_for_each_entry_safe(rq, tmp, &list, timeout_list)
                blk_abort_request(rq);
 
+       /*
+        * Occasionally, blk_abort_request() will return without
+        * deleting the element from the list. Make sure we add those back
+        * instead of leaving them on the local stack list.
+        */
+       list_splice(&list, &q->timeout_list);
+
        spin_unlock_irqrestore(q->queue_lock, flags);
 
 }