On graceful shutdown we will wait for discard queue to drain before storing the allocator.

ON fast shutdown we will simply copy the discard queue entries to the allocator

Signed-off-by: Gabriel BenHanokh <[email protected]>

Author: benhanokh

src/blk/BlockDevice.h

@@ -285,8 +285,8 @@ class BlockDevice {
     int write_hint = WRITE_LIFE_NOT_SET) = 0;
   virtual int flush() = 0;
   virtual bool try_discard(interval_set<uint64_t> &to_release, bool async=true) { return false; }
-  virtual int discard_drain(uint32_t timeout_msec = 0) { return 0; }
-
+  virtual void discard_drain() { return; }
+  virtual const interval_set<uint64_t>* get_discard_queued() { return nullptr;}
   // 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

@@ -587,35 +587,13 @@ bool KernelDevice::_discard_started()
   return !discard_threads.empty();
 }
 
-int KernelDevice::discard_drain(uint32_t timeout_msec = 0)
+void KernelDevice::discard_drain()
 {
   dout(10) << __func__ << dendl;
-  bool check_timeout = false;
-  utime_t end_time;
-  if (timeout_msec) {
-    check_timeout = true;
-    uint32_t timeout_sec = 0;
-    if (timeout_msec >= 1000) {
-      timeout_sec = (timeout_msec / 1000);
-      timeout_msec = (timeout_msec % 1000);
-    }
-    end_time = ceph_clock_now();
-    // add the timeout after converting from msec to nsec
-    end_time.tv.tv_nsec += (timeout_msec * (1000*1000));
-    if (end_time.tv.tv_nsec > (1000*1000*1000)) {
-      end_time.tv.tv_nsec -= (1000*1000*1000);
-      end_time.tv.tv_sec += 1;
-    }
-    end_time.tv.tv_sec += timeout_sec;
-  }
   std::unique_lock l(discard_lock);
   while (!discard_queued.empty() || discard_running) {
     discard_cond.wait(l);
-    if (check_timeout && ceph_clock_now() > end_time) {
-      return -1;
-    }
   }
-  return 0;
 }
 
 static bool is_expected_ioerr(const int r)

src/blk/kernel/KernelDevice.h

@@ -123,8 +123,8 @@ class KernelDevice : public BlockDevice,
   ~KernelDevice();
 
   void aio_submit(IOContext *ioc) override;
-  int discard_drain(uint32_t timeout_msec) override;
-
+  void discard_drain() override;
+  const interval_set<uint64_t>* get_discard_queued() override { return &discard_queued;}
   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

@@ -7762,17 +7762,24 @@ void BlueStore::_close_db()
   db = nullptr;
 
   if (do_destage && fm && fm->is_null_manager()) {
-    // force all backgrounds discards to be committed before storing allocator
-    // set timeout to 500msec
-    int ret = bdev->discard_drain(500);
-    if (ret == 0) {
-      ret = store_allocator(alloc);
-      if (unlikely(ret != 0)) {
-	derr << __func__ << "::NCB::store_allocator() failed (we will need to rebuild it on startup)" << dendl;
+    if (cct->_conf->osd_fast_shutdown == false) {
+      // graceful shutdown -> commit backgrounds discards before storing allocator
+      bdev->discard_drain();
+    }
+
+    auto discard_queued = bdev->get_discard_queued();
+    if (discard_queued && (discard_queued->num_intervals() > 0)) {
+      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());
       }
     }
-    else {
-      derr << __func__ << "::NCB::discard_drain() exceeded timeout (abort!)" << dendl;
+    int ret = store_allocator(alloc);
+    if (unlikely(ret != 0)) {
+      derr << __func__ << "::NCB::store_allocator() failed (we will need to rebuild it on startup)" << dendl;
     }
   }