From: Abutalib Aghayev <agayev@cs.cmu.edu>
Date: Fri, 12 Jun 2020 14:51:59 +0000 (-0400)
Subject: os/bluestore: Add missing punch_hole call so that object offsets are updated.
X-Git-Tag: v16.1.0~1998^2~1
X-Git-Url: http://git-server-git.apps.pok.os.sepia.ceph.com/?a=commitdiff_plain;h=6317feae56614529cd42553a6a7dba6f52999b5f;p=ceph.git

os/bluestore: Add missing punch_hole call so that object offsets are updated.

Rearranged an HM-SMR related code in _do_write_small so that it reuses
previously calculated offsets and, most importantly, added the missing
onode->extent_map.punch_hole call that releases overwritten blobs and updates
the location of the object to point to the latest copy.  Without it the OSD
could not restart becase a stale copy of osd_superblock object was being read.

Signed-off-by: Abutalib Aghayev <agayev@cs.cmu.edu>
---

diff --git a/src/os/bluestore/BlueStore.cc b/src/os/bluestore/BlueStore.cc
index 7e7488f47934c..ddb730f7efa0d 100644
--- a/src/os/bluestore/BlueStore.cc
+++ b/src/os/bluestore/BlueStore.cc
@@ -13035,22 +13035,6 @@ void BlueStore::_do_write_small(
 	   << std::dec << dendl;
   ceph_assert(length < min_alloc_size);
 
-  // On zoned devices, the first goal is to support non-overwrite workloads,
-  // such as RGW, with large, aligned objects.  Therefore, for user writes
-  // _do_write_small should not trigger.  OSDs, however, write and update a tiny
-  // amount of metadata, such as OSD maps, to disk.  For those cases, we
-  // temporarily just pad them to min_alloc_size and write them to a new place
-  // on every update.
-  if (bdev->is_smr()) {
-    BlobRef b = c->new_blob();
-    uint64_t b_off = 0, b_off0 = 0;
-    bufferlist l;
-    blp.copy(length, l);
-    _pad_zeros(&l, &b_off0, min_alloc_size);
-    wctx->write(offset, b, min_alloc_size, b_off0, l, b_off, length, false, true);
-    return;
-  }
-
   uint64_t end_offs = offset + length;
 
   logger->inc(l_bluestore_write_small);
@@ -13073,6 +13057,22 @@ void BlueStore::_do_write_small(
   // than 'offset' only).
   o->extent_map.fault_range(db, min_off, offset + max_bsize - min_off);
 
+  // On zoned devices, the first goal is to support non-overwrite workloads,
+  // such as RGW, with large, aligned objects.  Therefore, for user writes
+  // _do_write_small should not trigger.  OSDs, however, write and update a tiny
+  // amount of metadata, such as OSD maps, to disk.  For those cases, we
+  // temporarily just pad them to min_alloc_size and write them to a new place
+  // on every update.
+  if (bdev->is_smr()) {
+    BlobRef b = c->new_blob();
+    uint64_t b_off = p2phase<uint64_t>(offset, alloc_len);
+    uint64_t b_off0 = b_off;
+    _pad_zeros(&bl, &b_off0, min_alloc_size);
+    o->extent_map.punch_hole(c, offset, length, &wctx->old_extents);
+    wctx->write(offset, b, alloc_len, b_off0, bl, b_off, length, false, true);
+    return;
+  }
+
   // Look for an existing mutable blob we can use.
   auto begin = o->extent_map.extent_map.begin();
   auto end = o->extent_map.extent_map.end();