os/bluestore: simplify and fix blob invalidation

We can't invalidate based on any particular lextent we drop, as there may
be other lextents (in this onode or in another one) that reference the
same data.  This can break even a single multi-write transaction that
relies on the buffer cache to fill out blocks.

Fix by just doing the unallocated version unconditionally.

Signed-off-by: Sage Weil <sage@redhat.com>

diff --git a/src/os/bluestore/BlueStore.cc b/src/os/bluestore/BlueStore.cc
index 8b13695e6a5009eeaddd03e5ba4c42367d5c6008..30125306aed7c4d8b5599529621fdaab1a597a64 100644 (file)
--- a/src/os/bluestore/BlueStore.cc
+++ b/src/os/bluestore/BlueStore.cc
@@ -5919,19 +5919,13 @@ void BlueStore::_wctx_finish(
     vector<bluestore_pextent_t> r;
     bool compressed = b->blob.is_compressed();
     b->blob.put_ref(l.offset, l.length, min_alloc_size, &r);
-    if (l.blob > 0) {
-      // blob is owned by the onode; invalidate buffers as *we* drop
-      // logical refs.
-      b->bc.discard(l.offset, l.length);
-    } else {
-      // blob is shared.  we can't invalidate our logical extents as
-      // we drop them because other onodes may still reference them.
-      // but we can throw out anything that is no longer allocated.
-      // Note that this will leave behind edge bits that are no
-      // longer referenced but not deallocated (until they age out
-      // of the cache naturally).
-      b->discard_unallocated();
-    }
+    // we can't invalidate our logical extents as we drop them because
+    // other lextents (either in our onode or others) may still
+    // reference them.  but we can throw out anything that is no
+    // longer allocated.  Note that this will leave behind edge bits
+    // that are no longer referenced but not deallocated (until they
+    // age out of the cache naturally).
+    b->discard_unallocated();
     txc->statfs_delta.stored() -= l.length;
     if (compressed) {
       txc->statfs_delta.compressed_original() -= l.length;