Patchwork [RFC,v2,01/23] qcow2: Handle dependencies earlier

login
register
mail settings
Submitter Kevin Wolf
Date Feb. 13, 2013, 1:21 p.m.
Message ID <1360761733-25347-2-git-send-email-kwolf@redhat.com>
Download mbox | patch
Permalink /patch/220129/
State New
Headers show

Comments

Kevin Wolf - Feb. 13, 2013, 1:21 p.m.
Handling overlapping allocations isn't just a detail of cluster
allocation. It is rather one of three ways to get the host cluster
offset for a write request:

1. If a request overlaps an in-flight allocations, the cluster offset
   can be taken from there (this is what handle_dependencies will evolve
   into) or the request must just wait until the allocation has
   completed. Accessing the L2 is not valid in this case, it has
   outdated information.

2. Outside overlapping areas, check the clusters that can be written to
   as they are, with no COW involved.

3. If a COW is required, allocate new clusters

Changing the code to reflect this doesn't change the behaviour because
overlaps cannot exist for clusters that are kept in step 2. It does
however make it easier for later patches to work on clusters that belong
to an allocation that is still in flight.

Signed-off-by: Kevin Wolf <kwolf@redhat.com>
---
 block/qcow2-cluster.c |   59 +++++++++++++++++++++++++++++++++++-------------
 block/qcow2.h         |    5 ++++
 2 files changed, 48 insertions(+), 16 deletions(-)
Stefan Hajnoczi - Feb. 14, 2013, 2:15 p.m.
On Wed, Feb 13, 2013 at 02:21:51PM +0100, Kevin Wolf wrote:
> @@ -882,20 +876,55 @@ int qcow2_alloc_cluster_offset(BlockDriverState *bs, uint64_t offset,
>      trace_qcow2_alloc_clusters_offset(qemu_coroutine_self(), offset,
>                                        n_start, n_end);
>  
> -    /* Find L2 entry for the first involved cluster */
>  again:
> -    ret = get_cluster_table(bs, offset, &l2_table, &l2_index);
> -    if (ret < 0) {
> -        return ret;
> -    }
> -
>      /*
>       * Calculate the number of clusters to look for. We stop at L2 table
>       * boundaries to keep things simple.
>       */
> +    l2_index = offset_to_l2_index(s, offset);
>      nb_clusters = MIN(size_to_clusters(s, n_end << BDRV_SECTOR_BITS),
>                        s->l2_size - l2_index);
>  
> +    /*
> +     * Now start gathering as many contiguous clusters as possible:
> +     *
> +     * 1. Check for overlaps with in-flight allocations
> +     *
> +     *      a) Overlap not in the first cluster -> shorten this request and let
> +     *         the caller handle the rest in its next loop iteration.
> +     *
> +     *      b) Real overlaps of two requests. Yield and restart the search for
> +     *         contiguous clusters (the situation could have changed while we
> +     *         were sleeping)
> +     *
> +     *      c) TODO: Request starts in the same cluster as the in-flight
> +     *         allocation ends. Shorten the COW of the in-fight allocation, set
> +     *         cluster_offset to write to the same cluster and set up the right
> +     *         synchronisation between the in-flight request and the new one.
> +     *
> +     * 2. Count contiguous COPIED clusters.
> +     *    TODO: Consider cluster_offset if set in step 1c.
> +     *
> +     * 3. If the request still hasn't completed, allocate new clusters,
> +     *    considering any cluster_offset of steps 1c or 2.
> +     */
> +    ret = handle_dependencies(bs, offset, &nb_clusters);
> +    if (ret == -EAGAIN) {
> +        goto again;

Now we might be able to replace the again label with a while loop.

do {
    l2_index = offset_to_l2_index(s, offset);
    nb_clusters = MIN(size_to_clusters(s, n_end << BDRV_SECTOR_BITS),
                      s->l2_size - l2_index);
} while ((ret = handle_dependencies(bs, offset, &nb_clusters)) == -EAGAIN);

Not essential, but perhaps clearer.

Patch

diff --git a/block/qcow2-cluster.c b/block/qcow2-cluster.c
index 56fccf9..c3d57e0 100644
--- a/block/qcow2-cluster.c
+++ b/block/qcow2-cluster.c
@@ -821,16 +821,10 @@  static int do_alloc_cluster_offset(BlockDriverState *bs, uint64_t guest_offset,
     uint64_t *host_offset, unsigned int *nb_clusters)
 {
     BDRVQcowState *s = bs->opaque;
-    int ret;
 
     trace_qcow2_do_alloc_clusters_offset(qemu_coroutine_self(), guest_offset,
                                          *host_offset, *nb_clusters);
 
-    ret = handle_dependencies(bs, guest_offset, nb_clusters);
-    if (ret < 0) {
-        return ret;
-    }
-
     /* Allocate new clusters */
     trace_qcow2_cluster_alloc_phys(qemu_coroutine_self());
     if (*host_offset == 0) {
@@ -842,7 +836,7 @@  static int do_alloc_cluster_offset(BlockDriverState *bs, uint64_t guest_offset,
         *host_offset = cluster_offset;
         return 0;
     } else {
-        ret = qcow2_alloc_clusters_at(bs, *host_offset, *nb_clusters);
+        int ret = qcow2_alloc_clusters_at(bs, *host_offset, *nb_clusters);
         if (ret < 0) {
             return ret;
         }
@@ -882,20 +876,55 @@  int qcow2_alloc_cluster_offset(BlockDriverState *bs, uint64_t offset,
     trace_qcow2_alloc_clusters_offset(qemu_coroutine_self(), offset,
                                       n_start, n_end);
 
-    /* Find L2 entry for the first involved cluster */
 again:
-    ret = get_cluster_table(bs, offset, &l2_table, &l2_index);
-    if (ret < 0) {
-        return ret;
-    }
-
     /*
      * Calculate the number of clusters to look for. We stop at L2 table
      * boundaries to keep things simple.
      */
+    l2_index = offset_to_l2_index(s, offset);
     nb_clusters = MIN(size_to_clusters(s, n_end << BDRV_SECTOR_BITS),
                       s->l2_size - l2_index);
 
+    /*
+     * Now start gathering as many contiguous clusters as possible:
+     *
+     * 1. Check for overlaps with in-flight allocations
+     *
+     *      a) Overlap not in the first cluster -> shorten this request and let
+     *         the caller handle the rest in its next loop iteration.
+     *
+     *      b) Real overlaps of two requests. Yield and restart the search for
+     *         contiguous clusters (the situation could have changed while we
+     *         were sleeping)
+     *
+     *      c) TODO: Request starts in the same cluster as the in-flight
+     *         allocation ends. Shorten the COW of the in-fight allocation, set
+     *         cluster_offset to write to the same cluster and set up the right
+     *         synchronisation between the in-flight request and the new one.
+     *
+     * 2. Count contiguous COPIED clusters.
+     *    TODO: Consider cluster_offset if set in step 1c.
+     *
+     * 3. If the request still hasn't completed, allocate new clusters,
+     *    considering any cluster_offset of steps 1c or 2.
+     */
+    ret = handle_dependencies(bs, offset, &nb_clusters);
+    if (ret == -EAGAIN) {
+        goto again;
+    } else if (ret < 0) {
+        return ret;
+    } else {
+        /* handle_dependencies() may have decreased cur_bytes (shortened
+         * the allocations below) so that the next dependency is processed
+         * correctly during the next loop iteration. */
+    }
+
+    /* Find L2 entry for the first involved cluster */
+    ret = get_cluster_table(bs, offset, &l2_table, &l2_index);
+    if (ret < 0) {
+        return ret;
+    }
+
     cluster_offset = be64_to_cpu(l2_table[l2_index]);
 
     /*
@@ -960,9 +989,7 @@  again:
         /* Allocate, if necessary at a given offset in the image file */
         ret = do_alloc_cluster_offset(bs, alloc_offset, &alloc_cluster_offset,
                                       &nb_clusters);
-        if (ret == -EAGAIN) {
-            goto again;
-        } else if (ret < 0) {
+        if (ret < 0) {
             goto fail;
         }
 
diff --git a/block/qcow2.h b/block/qcow2.h
index 718b52b..1322012 100644
--- a/block/qcow2.h
+++ b/block/qcow2.h
@@ -273,6 +273,11 @@  static inline int size_to_l1(BDRVQcowState *s, int64_t size)
     return (size + (1ULL << shift) - 1) >> shift;
 }
 
+static inline int offset_to_l2_index(BDRVQcowState *s, int64_t offset)
+{
+    return (offset >> s->cluster_bits) & (s->l2_size - 1);
+}
+
 static inline int64_t align_offset(int64_t offset, int n)
 {
     offset = (offset + n - 1) & ~(n - 1);