md: common infrastructure for running operations with raid5_run_ops
authorDan Williams <dan.j.williams@intel.com>
Tue, 2 Jan 2007 20:52:30 +0000 (13:52 -0700)
committerDan Williams <dan.j.williams@intel.com>
Fri, 13 Jul 2007 15:06:16 +0000 (08:06 -0700)
All the handle_stripe operations that are to be transitioned to use
raid5_run_ops need a method to coherently gather work under the stripe-lock
and hand that work off to raid5_run_ops.  The 'get_stripe_work' routine
runs under the lock to read all the bits in sh->ops.pending that do not
have the corresponding bit set in sh->ops.ack.  This modified 'pending'
bitmap is then passed to raid5_run_ops for processing.

The transition from 'ack' to 'completion' does not need similar protection
as the existing release_stripe infrastructure will guarantee that
handle_stripe will run again after a completion bit is set, and
handle_stripe can tolerate a sh->ops.completed bit being set while the lock
is held.

A call to async_tx_issue_pending_all() is added to raid5d to kick the
offload engines once all pending stripe operations work has been submitted.
This enables batching of the submission and completion of operations.

Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Acked-By: NeilBrown <neilb@suse.de>
drivers/md/raid5.c

index 0b7002479655008e028cb1f80ab354b424460125..d89a25e7c17bb0d50b0bc78bb290877ddb31ea8a 100644 (file)
@@ -141,6 +141,7 @@ static void __release_stripe(raid5_conf_t *conf, struct stripe_head *sh)
                        }
                        md_wakeup_thread(conf->mddev->thread);
                } else {
+                       BUG_ON(sh->ops.pending);
                        if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
                                atomic_dec(&conf->preread_active_stripes);
                                if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
@@ -242,7 +243,8 @@ static void init_stripe(struct stripe_head *sh, sector_t sector, int pd_idx, int
 
        BUG_ON(atomic_read(&sh->count) != 0);
        BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
-       
+       BUG_ON(sh->ops.pending || sh->ops.ack || sh->ops.complete);
+
        CHECK_DEVLOCK();
        pr_debug("init_stripe called, stripe %llu\n",
                (unsigned long long)sh->sector);
@@ -258,11 +260,11 @@ static void init_stripe(struct stripe_head *sh, sector_t sector, int pd_idx, int
        for (i = sh->disks; i--; ) {
                struct r5dev *dev = &sh->dev[i];
 
-               if (dev->toread || dev->towrite || dev->written ||
+               if (dev->toread || dev->read || dev->towrite || dev->written ||
                    test_bit(R5_LOCKED, &dev->flags)) {
-                       printk("sector=%llx i=%d %p %p %p %d\n",
+                       printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
                               (unsigned long long)sh->sector, i, dev->toread,
-                              dev->towrite, dev->written,
+                              dev->read, dev->towrite, dev->written,
                               test_bit(R5_LOCKED, &dev->flags));
                        BUG();
                }
@@ -342,6 +344,44 @@ static struct stripe_head *get_active_stripe(raid5_conf_t *conf, sector_t sector
        return sh;
 }
 
+/* test_and_ack_op() ensures that we only dequeue an operation once */
+#define test_and_ack_op(op, pend) \
+do {                                                   \
+       if (test_bit(op, &sh->ops.pending) &&           \
+               !test_bit(op, &sh->ops.complete)) {     \
+               if (test_and_set_bit(op, &sh->ops.ack)) \
+                       clear_bit(op, &pend);           \
+               else                                    \
+                       ack++;                          \
+       } else                                          \
+               clear_bit(op, &pend);                   \
+} while (0)
+
+/* find new work to run, do not resubmit work that is already
+ * in flight
+ */
+static unsigned long get_stripe_work(struct stripe_head *sh)
+{
+       unsigned long pending;
+       int ack = 0;
+
+       pending = sh->ops.pending;
+
+       test_and_ack_op(STRIPE_OP_BIOFILL, pending);
+       test_and_ack_op(STRIPE_OP_COMPUTE_BLK, pending);
+       test_and_ack_op(STRIPE_OP_PREXOR, pending);
+       test_and_ack_op(STRIPE_OP_BIODRAIN, pending);
+       test_and_ack_op(STRIPE_OP_POSTXOR, pending);
+       test_and_ack_op(STRIPE_OP_CHECK, pending);
+       if (test_and_clear_bit(STRIPE_OP_IO, &sh->ops.pending))
+               ack++;
+
+       sh->ops.count -= ack;
+       BUG_ON(sh->ops.count < 0);
+
+       return pending;
+}
+
 static int
 raid5_end_read_request(struct bio *bi, unsigned int bytes_done, int error);
 static int
@@ -2494,7 +2534,6 @@ static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh,
  *    schedule a write of some buffers
  *    return confirmation of parity correctness
  *
- * Parity calculations are done inside the stripe lock
  * buffers are taken off read_list or write_list, and bh_cache buffers
  * get BH_Lock set before the stripe lock is released.
  *
@@ -2507,11 +2546,13 @@ static void handle_stripe5(struct stripe_head *sh)
        struct bio *return_bi = NULL;
        struct stripe_head_state s;
        struct r5dev *dev;
+       unsigned long pending = 0;
 
        memset(&s, 0, sizeof(s));
-       pr_debug("handling stripe %llu, cnt=%d, pd_idx=%d\n",
-               (unsigned long long)sh->sector, atomic_read(&sh->count),
-               sh->pd_idx);
+       pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d "
+               "ops=%lx:%lx:%lx\n", (unsigned long long)sh->sector, sh->state,
+               atomic_read(&sh->count), sh->pd_idx,
+               sh->ops.pending, sh->ops.ack, sh->ops.complete);
 
        spin_lock(&sh->lock);
        clear_bit(STRIPE_HANDLE, &sh->state);
@@ -2674,8 +2715,14 @@ static void handle_stripe5(struct stripe_head *sh)
        if (s.expanding && s.locked == 0)
                handle_stripe_expansion(conf, sh, NULL);
 
+       if (sh->ops.count)
+               pending = get_stripe_work(sh);
+
        spin_unlock(&sh->lock);
 
+       if (pending)
+               raid5_run_ops(sh, pending);
+
        return_io(return_bi);
 
        for (i=disks; i-- ;) {
@@ -3798,8 +3845,10 @@ static void raid5d (mddev_t *mddev)
                        handled++;
                }
 
-               if (list_empty(&conf->handle_list))
+               if (list_empty(&conf->handle_list)) {
+                       async_tx_issue_pending_all();
                        break;
+               }
 
                first = conf->handle_list.next;
                sh = list_entry(first, struct stripe_head, lru);