From: Sage Weil <sage@newdream.net>
Date: Tue, 7 Sep 2021 18:53:06 +0000 (-0500)
Subject: os/bluestore: fix startup vs device write pointers
X-Git-Tag: v17.1.0~535^2~39
X-Git-Url: http://git-server-git.apps.pok.os.sepia.ceph.com/?a=commitdiff_plain;h=7f74551b7bdf13afb0f0d31b18bdafff91df3d82;p=ceph.git

os/bluestore: fix startup vs device write pointers

Compare freelist write pointers to the device on startup.

Signed-off-by: Sage Weil <sage@newdream.net>
---

diff --git a/src/os/bluestore/BlueStore.cc b/src/os/bluestore/BlueStore.cc
index 2004f7286457..459f7070b1bc 100644
--- a/src/os/bluestore/BlueStore.cc
+++ b/src/os/bluestore/BlueStore.cc
@@ -5592,7 +5592,36 @@ int BlueStore::_init_alloc()
     ceph_assert(a);
     auto f = dynamic_cast<ZonedFreelistManager*>(fm);
     ceph_assert(f);
-    a->init_from_zone_pointers(f->get_zone_states(db),
+    vector<uint64_t> wp = bdev->get_zones();
+    vector<zone_state_t> zones = f->get_zone_states(db);
+    ceph_assert(wp.size() == zones.size());
+
+    // reconcile zone state
+    auto num_zones = bdev->get_size() / zone_size;
+    for (unsigned i = first_sequential_zone; i < num_zones; ++i) {
+      ceph_assert(wp[i] >= i * zone_size);
+      ceph_assert(wp[i] <= (i + 1) * zone_size); // pos might be at start of next zone
+      uint64_t p = wp[i] - i * zone_size;
+      if (zones[i].write_pointer > p) {
+	derr << __func__ << " zone 0x" << std::hex << i
+	     << " bluestore write pointer 0x" << zones[i].write_pointer
+	     << " > device write pointer 0x" << p
+	     << std::dec << dendl;
+	ceph_abort("bad write pointer");
+      } else if (zones[i].write_pointer < p) {
+	// this is "normal" in that it can happen after any crash (if we have a
+	// write in flight but did not manage to commit the transaction)
+	auto delta = p - zones[i].write_pointer;
+	dout(1) << __func__ << " zone 0x" << std::hex << i
+		 << " device write pointer 0x" << p
+		 << " > bluestore pointer 0x" << zones[i].write_pointer
+		 << ", advancing 0x" << delta << std::dec << dendl;
+	zones[i].num_dead_bytes += delta;
+	zones[i].write_pointer = p;
+      }
+    }
+
+    a->init_from_zone_pointers(zones,
 			       &zoned_cleaner_lock,
 			       &zoned_cleaner_cond);
     dout(1) << __func__
diff --git a/src/os/bluestore/ZonedAllocator.cc b/src/os/bluestore/ZonedAllocator.cc
index 0ec8000e17b6..03326d2e5fc5 100644
--- a/src/os/bluestore/ZonedAllocator.cc
+++ b/src/os/bluestore/ZonedAllocator.cc
@@ -151,7 +151,7 @@ void ZonedAllocator::dump(std::function<void(uint64_t offset,
 }
 
 void ZonedAllocator::init_from_zone_pointers(
-  std::vector<zone_state_t> &&_zone_states,
+  std::vector<zone_state_t> _zone_states,
   ceph::mutex *_cleaner_lock,
   ceph::condition_variable *_cleaner_cond)
 {
diff --git a/src/os/bluestore/ZonedAllocator.h b/src/os/bluestore/ZonedAllocator.h
index 73697e4cfd5f..903b62fd1a09 100644
--- a/src/os/bluestore/ZonedAllocator.h
+++ b/src/os/bluestore/ZonedAllocator.h
@@ -96,7 +96,7 @@ public:
   void mark_zones_to_clean_free(void);
 
   void init_from_zone_pointers(
-    std::vector<zone_state_t> &&_zone_states,
+    std::vector<zone_state_t> _zone_states,
     ceph::mutex *_cleaner_lock,
     ceph::condition_variable *_cleaner_cond);
   void init_add_free(uint64_t offset, uint64_t length) override {}
diff --git a/src/os/bluestore/zoned_types.h b/src/os/bluestore/zoned_types.h
index c92f2ada0996..93d3138399a0 100644
--- a/src/os/bluestore/zoned_types.h
+++ b/src/os/bluestore/zoned_types.h
@@ -14,8 +14,8 @@
 // We use the same struct for an on-disk and in-memory representation of the
 // state.
 struct zone_state_t {
-  uint64_t num_dead_bytes = 0;
-  uint64_t write_pointer = 0;
+  uint64_t num_dead_bytes = 0;  ///< dead bytes deallocated (behind the write pointer)
+  uint64_t write_pointer = 0;   ///< relative offset within the zone
 
   void encode(ceph::buffer::list &bl) const {
     using ceph::encode;