Merge pull request ceph#56744 from benhanokh/ncb_async_discard_fix

os/BlueStore: NCB fix for leaked space when bdev_async_discard is ena…

Reviewed-by: Igor Fedotov <[email protected]>
Reviewed-by: Ronen Friedman <[email protected]>
Reviewed-by: Joshua Baergen <[email protected]>

Author: yuriw

src/blk/BlockDevice.h

@@ -286,7 +286,7 @@ class BlockDevice {
   virtual int flush() = 0;
   virtual bool try_discard(interval_set<uint64_t> &to_release, bool async=true) { return false; }
   virtual void discard_drain() { return; }
-
+  virtual void swap_discard_queued(interval_set<uint64_t>& other)  { other.clear(); }
   // for managing buffered readers/writers
   virtual int invalidate_cache(uint64_t off, uint64_t len) = 0;
   virtual int open(const std::string& path) = 0;

src/blk/kernel/KernelDevice.cc

@@ -591,7 +591,7 @@ void KernelDevice::discard_drain()
 {
   dout(10) << __func__ << dendl;
   std::unique_lock l(discard_lock);
-  while (!discard_queued.empty() || discard_running) {
+  while (!discard_queued.empty() || (discard_running > 0)) {
     discard_cond.wait(l);
   }
 }
@@ -731,6 +731,12 @@ void KernelDevice::_aio_thread()
   dout(10) << __func__ << " end" << dendl;
 }
 
+void KernelDevice::swap_discard_queued(interval_set<uint64_t>& other)
+{
+  std::unique_lock l(discard_lock);
+  discard_queued.swap(other);
+}
+
 void KernelDevice::_discard_thread(uint64_t tid)
 {
   dout(10) << __func__ << " thread " << tid << " start" << dendl;
@@ -755,13 +761,21 @@ void KernelDevice::_discard_thread(uint64_t tid)
       discard_cond.wait(l);
       dout(20) << __func__ << " wake" << dendl;
     } else {
-      // Swap the queued discards for a local list we'll process here
-      // without caring about thread fairness.  This allows the current
-      // thread to wait on the discard running while other threads pick
-      // up the next-in-queue, and do the same, ultimately issuing more
-      // discards in parallel, which is the goal.
-      discard_processing.swap(discard_queued);
-      discard_running = true;
+      // Limit local processing to MAX_LOCAL_DISCARD items.
+      // This will allow threads to work in parallel
+      //      instead of a single thread taking over the whole discard_queued.
+      // It will also allow threads to finish in a timely manner.
+      constexpr unsigned MAX_LOCAL_DISCARD = 32;
+      unsigned count = 0;
+      for (auto p = discard_queued.begin();
+	   p != discard_queued.end() && count < MAX_LOCAL_DISCARD;
+	   ++p, ++count) {
+	discard_processing.insert(p.get_start(), p.get_len());
+	discard_queued.erase(p);
+      }
+
+      // there are multiple active threads -> must use a counter instead of a flag
+      discard_running ++;
       l.unlock();
       dout(20) << __func__ << " finishing" << dendl;
       for (auto p = discard_processing.begin(); p != discard_processing.end(); ++p) {
@@ -771,7 +785,8 @@ void KernelDevice::_discard_thread(uint64_t tid)
       discard_callback(discard_callback_priv, static_cast<void*>(&discard_processing));
       discard_processing.clear();
       l.lock();
-      discard_running = false;
+      discard_running --;
+      ceph_assert(discard_running >= 0);
     }
   }
 

src/blk/kernel/KernelDevice.h

@@ -55,7 +55,7 @@ class KernelDevice : public BlockDevice,
 
   ceph::mutex discard_lock = ceph::make_mutex("KernelDevice::discard_lock");
   ceph::condition_variable discard_cond;
-  bool discard_running = false;
+  int discard_running = 0;
   interval_set<uint64_t> discard_queued;
 
   struct AioCompletionThread : public Thread {
@@ -124,7 +124,7 @@ class KernelDevice : public BlockDevice,
 
   void aio_submit(IOContext *ioc) override;
   void discard_drain() override;
-
+  void swap_discard_queued(interval_set<uint64_t>& other) override;
   int collect_metadata(const std::string& prefix, std::map<std::string,std::string> *pm) const override;
   int get_devname(std::string *s) const override {
     if (devname.empty()) {

src/os/bluestore/BlueStore.cc

@@ -7788,9 +7788,26 @@ void BlueStore::_close_db()
   db = nullptr;
 
   if (do_destage && fm && fm->is_null_manager()) {
+    if (cct->_conf->osd_fast_shutdown) {
+      interval_set<uint64_t> discard_queued;
+      bdev->swap_discard_queued(discard_queued);
+      dout(10) << __func__ << "::discard_drain: size=" << discard_queued.size()
+	       << " num_intervals=" << discard_queued.num_intervals() << dendl;
+      // copy discard_queued to the allocator before storing it
+      for (auto p = discard_queued.begin(); p != discard_queued.end(); ++p) {
+	dout(20) << __func__ << "::discarded-extent=[" << p.get_start()
+		 << ", " << p.get_len() << "]" << dendl;
+	alloc->init_add_free(p.get_start(), p.get_len());
+      }
+    }
+
+    // When we reach here it is either a graceful shutdown (so can drain the full discards-queue)
+    //   or it was a fast shutdown, but we already moved the main discards-queue to the allocator
+    //   and only need to wait for the threads local discard_processing queues to drain
+    bdev->discard_drain();
     int ret = store_allocator(alloc);
-    if (ret != 0) {
-      derr << __func__ << "::NCB::store_allocator() failed (continue with bitmapFreelistManager)" << dendl;
+    if (unlikely(ret != 0)) {
+      derr << __func__ << "::NCB::store_allocator() failed (we will need to rebuild it on startup)" << dendl;
     }
   }