qcow2: add discard-no-unref option

When we for example have a sparse qcow2 image and discard: unmap is enabled,
there can be a lot of fragmentation in the image after some time. Especially on VM's
that do a lot of writes/deletes.
This causes the qcow2 image to grow even over 110% of its virtual size,
because the free gaps in the image get too small to allocate new
continuous clusters. So it allocates new space at the end of the image.

Disabling discard is not an option, as discard is needed to keep the
incremental backup size as low as possible. Without discard, the
incremental backups would become large, as qemu thinks it's just dirty
blocks but it doesn't know the blocks are unneeded.
So we need to avoid fragmentation but also 'empty' the unneeded blocks in
the image to have a small incremental backup.

In addition, we also want to send the discards further down the stack, so
the underlying blocks are still discarded.

Therefor we introduce a new qcow2 option "discard-no-unref".
When setting this option to true, discards will no longer have the qcow2
driver relinquish cluster allocations. Other than that, the request is
handled as normal: All clusters in range are marked as zero, and, if
pass-discard-request is true, it is passed further down the stack.
The only difference is that the now-zero clusters are preallocated
instead of being unallocated.
This will avoid fragmentation on the qcow2 image.

Fixes: https://gitlab.com/qemu-project/qemu/-/issues/1621
Signed-off-by: Jean-Louis Dupond <jean-louis@dupond.be>
Message-Id: <20230605084523.34134-2-jean-louis@dupond.be>
Reviewed-by: Hanna Czenczek <hreitz@redhat.com>
Signed-off-by: Hanna Czenczek <hreitz@redhat.com>

3 files changed, 49 insertions, 4 deletions

diff --git a/block/qcow2-cluster.c b/block/qcow2-cluster.c
index 39cda7f..2e76de0 100644
--- a/block/qcow2-cluster.c
+++ b/block/qcow2-cluster.c
@@ -1925,6 +1925,10 @@ static int discard_in_l2_slice(BlockDriverState *bs, uint64_t offset,
         uint64_t new_l2_bitmap = old_l2_bitmap;
         QCow2ClusterType cluster_type =
             qcow2_get_cluster_type(bs, old_l2_entry);
+        bool keep_reference = (cluster_type != QCOW2_CLUSTER_COMPRESSED) &&
+                              !full_discard &&
+                              (s->discard_no_unref &&
+                               type == QCOW2_DISCARD_REQUEST);
 
         /*
          * If full_discard is true, the cluster should not read back as zeroes,
@@ -1943,10 +1947,22 @@ static int discard_in_l2_slice(BlockDriverState *bs, uint64_t offset,
             new_l2_entry = new_l2_bitmap = 0;
         } else if (bs->backing || qcow2_cluster_is_allocated(cluster_type)) {
             if (has_subclusters(s)) {
-                new_l2_entry = 0;
+                if (keep_reference) {
+                    new_l2_entry = old_l2_entry;
+                } else {
+                    new_l2_entry = 0;
+                }
                 new_l2_bitmap = QCOW_L2_BITMAP_ALL_ZEROES;
             } else {
-                new_l2_entry = s->qcow_version >= 3 ? QCOW_OFLAG_ZERO : 0;
+                if (s->qcow_version >= 3) {
+                    if (keep_reference) {
+                        new_l2_entry |= QCOW_OFLAG_ZERO;
+                    } else {
+                        new_l2_entry = QCOW_OFLAG_ZERO;
+                    }
+                } else {
+                    new_l2_entry = 0;
+                }
             }
         }
 
@@ -1960,8 +1976,16 @@ static int discard_in_l2_slice(BlockDriverState *bs, uint64_t offset,
         if (has_subclusters(s)) {
             set_l2_bitmap(s, l2_slice, l2_index + i, new_l2_bitmap);
         }
-        /* Then decrease the refcount */
-        qcow2_free_any_cluster(bs, old_l2_entry, type);
+        if (!keep_reference) {
+            /* Then decrease the refcount */
+            qcow2_free_any_cluster(bs, old_l2_entry, type);
+        } else if (s->discard_passthrough[type] &&
+                   (cluster_type == QCOW2_CLUSTER_NORMAL ||
+                    cluster_type == QCOW2_CLUSTER_ZERO_ALLOC)) {
+            /* If we keep the reference, pass on the discard still */
+            bdrv_pdiscard(s->data_file, old_l2_entry & L2E_OFFSET_MASK,
+                          s->cluster_size);
+       }
     }
 
     qcow2_cache_put(s->l2_table_cache, (void **) &l2_slice);
diff --git a/block/qcow2.c b/block/qcow2.c
index 7f39483..e23edd4 100644
--- a/block/qcow2.c
+++ b/block/qcow2.c
@@ -682,6 +682,7 @@ static const char *const mutable_opts[] = {
     QCOW2_OPT_DISCARD_REQUEST,
     QCOW2_OPT_DISCARD_SNAPSHOT,
     QCOW2_OPT_DISCARD_OTHER,
+    QCOW2_OPT_DISCARD_NO_UNREF,
     QCOW2_OPT_OVERLAP,
     QCOW2_OPT_OVERLAP_TEMPLATE,
     QCOW2_OPT_OVERLAP_MAIN_HEADER,
@@ -727,6 +728,11 @@ static QemuOptsList qcow2_runtime_opts = {
             .help = "Generate discard requests when other clusters are freed",
         },
         {
+            .name = QCOW2_OPT_DISCARD_NO_UNREF,
+            .type = QEMU_OPT_BOOL,
+            .help = "Do not unreference discarded clusters",
+        },
+        {
             .name = QCOW2_OPT_OVERLAP,
             .type = QEMU_OPT_STRING,
             .help = "Selects which overlap checks to perform from a range of "
@@ -969,6 +975,7 @@ typedef struct Qcow2ReopenState {
     bool use_lazy_refcounts;
     int overlap_check;
     bool discard_passthrough[QCOW2_DISCARD_MAX];
+    bool discard_no_unref;
     uint64_t cache_clean_interval;
     QCryptoBlockOpenOptions *crypto_opts; /* Disk encryption runtime options */
 } Qcow2ReopenState;
@@ -1140,6 +1147,15 @@ static int qcow2_update_options_prepare(BlockDriverState *bs,
     r->discard_passthrough[QCOW2_DISCARD_OTHER] =
         qemu_opt_get_bool(opts, QCOW2_OPT_DISCARD_OTHER, false);
 
+    r->discard_no_unref = qemu_opt_get_bool(opts, QCOW2_OPT_DISCARD_NO_UNREF,
+                                            false);
+    if (r->discard_no_unref && s->qcow_version < 3) {
+        error_setg(errp,
+                   "discard-no-unref is only supported since qcow2 version 3");
+        ret = -EINVAL;
+        goto fail;
+    }
+
     switch (s->crypt_method_header) {
     case QCOW_CRYPT_NONE:
         if (encryptfmt) {
@@ -1220,6 +1236,8 @@ static void qcow2_update_options_commit(BlockDriverState *bs,
         s->discard_passthrough[i] = r->discard_passthrough[i];
     }
 
+    s->discard_no_unref = r->discard_no_unref;
+
     if (s->cache_clean_interval != r->cache_clean_interval) {
         cache_clean_timer_del(bs);
         s->cache_clean_interval = r->cache_clean_interval;
diff --git a/block/qcow2.h b/block/qcow2.h
index 4f67eb9..ea9adb5 100644
--- a/block/qcow2.h
+++ b/block/qcow2.h
@@ -133,6 +133,7 @@
 #define QCOW2_OPT_DISCARD_REQUEST "pass-discard-request"
 #define QCOW2_OPT_DISCARD_SNAPSHOT "pass-discard-snapshot"
 #define QCOW2_OPT_DISCARD_OTHER "pass-discard-other"
+#define QCOW2_OPT_DISCARD_NO_UNREF "discard-no-unref"
 #define QCOW2_OPT_OVERLAP "overlap-check"
 #define QCOW2_OPT_OVERLAP_TEMPLATE "overlap-check.template"
 #define QCOW2_OPT_OVERLAP_MAIN_HEADER "overlap-check.main-header"
@@ -385,6 +386,8 @@ typedef struct BDRVQcow2State {
 
     bool discard_passthrough[QCOW2_DISCARD_MAX];
 
+    bool discard_no_unref;
+
     int overlap_check; /* bitmask of Qcow2MetadataOverlap values */
     bool signaled_corruption;