From 02d907dde78cddb0889f65f4d76793fd1c407315 Mon Sep 17 00:00:00 2001
From: Adam Kupczyk <akupczyk@ibm.com>
Date: Mon, 7 Oct 2024 11:45:41 +0000
Subject: [PATCH] os/bluestore: Make truncate() drop unused allocations

Now when truncate() drops unused allocations.
Modified Close() in BlueRocksEnv to unconditionally call truncate.

Fixes: https://tracker.ceph.com/issues/68488

Signed-off-by: Adam Kupczyk <akupczyk@ibm.com>
(cherry picked from commit 9fc65f160cd3764a68fb3697d067c358761fc837)

Conflicts:
	src/os/bluestore/BlueFS.cc, trivial
---
 src/os/bluestore/BlueFS.cc       | 65 +++++++++++++++++++++++++-------
 src/os/bluestore/BlueRocksEnv.cc | 14 ++-----
 2 files changed, 56 insertions(+), 23 deletions(-)

diff --git a/src/os/bluestore/BlueFS.cc b/src/os/bluestore/BlueFS.cc
index 60d07bfb51ca0..cb81a99308b04 100644
--- a/src/os/bluestore/BlueFS.cc
+++ b/src/os/bluestore/BlueFS.cc
@@ -3660,15 +3660,16 @@ uint64_t BlueFS::_flush_special(FileWriter *h)
 int BlueFS::truncate(FileWriter *h, uint64_t offset)/*_WF_L*/
 {
   std::lock_guard hl(h->lock);
+  auto& fnode = h->file->fnode;
   dout(10) << __func__ << " 0x" << std::hex << offset << std::dec
-           << " file " << h->file->fnode << dendl;
+           << " file " << fnode << dendl;
   if (h->file->deleted) {
     dout(10) << __func__ << "  deleted, no-op" << dendl;
     return 0;
   }
 
   // we never truncate internal log files
-  ceph_assert(h->file->fnode.ino > 1);
+  ceph_assert(fnode.ino > 1);
 
   // truncate off unflushed data?
   if (h->pos < offset &&
@@ -3682,20 +3683,58 @@ int BlueFS::truncate(FileWriter *h, uint64_t offset)/*_WF_L*/
     if (r < 0)
       return r;
   }
-  if (offset == h->file->fnode.size) {
-    return 0;  // no-op!
-  }
-  if (offset > h->file->fnode.size) {
+  if (offset > fnode.size) {
     ceph_abort_msg("truncate up not supported");
   }
-  ceph_assert(h->file->fnode.size >= offset);
+  ceph_assert(offset <= fnode.size);
   _flush_bdev(h);
-
-  std::lock_guard ll(log.lock);
-  vselector->sub_usage(h->file->vselector_hint, h->file->fnode.size - offset);
-  h->file->fnode.size = offset;
-  h->file->is_dirty = true;
-  log.t.op_file_update_inc(h->file->fnode);
+  {
+    std::lock_guard ll(log.lock);
+    std::lock_guard dl(dirty.lock);
+    bool changed_extents = false;
+    vselector->sub_usage(h->file->vselector_hint, fnode);
+    uint64_t x_off = 0;
+    auto p = fnode.seek(offset, &x_off);
+    uint64_t cut_off =
+      (p == fnode.extents.end()) ? 0 : p2roundup(x_off, alloc_size[p->bdev]);
+    uint64_t new_allocated;
+    if (0 == cut_off) {
+      // whole pextent to remove
+      changed_extents = true;
+      new_allocated = offset;
+    } else if (cut_off < p->length) {
+      dirty.pending_release[p->bdev].insert(p->offset + cut_off, p->length - cut_off);
+      new_allocated = (offset - x_off) + cut_off;
+      p->length = cut_off;
+      changed_extents = true;
+      ++p;
+    } else {
+      ceph_assert(cut_off >= p->length);
+      new_allocated  = (offset - x_off) + p->length;
+      // just leave it here
+      ++p;
+    }
+    while (p != fnode.extents.end()) {
+      dirty.pending_release[p->bdev].insert(p->offset, p->length);
+      p = fnode.extents.erase(p);
+      changed_extents = true;
+    }
+    if (changed_extents) {
+      fnode.size = offset;
+      fnode.allocated = new_allocated;
+      fnode.reset_delta();
+      log.t.op_file_update(fnode);
+      // sad, but is_dirty must be set to signal flushing of the log
+      h->file->is_dirty = true;
+    } else {
+      if (offset != fnode.size) {
+        fnode.size = offset;
+        //skipping log.t.op_file_update_inc, it will be done by flush()
+        h->file->is_dirty = true;
+      }
+    }
+    vselector->add_usage(h->file->vselector_hint, fnode);
+  }
   return 0;
 }
 
diff --git a/src/os/bluestore/BlueRocksEnv.cc b/src/os/bluestore/BlueRocksEnv.cc
index 68040af428280..7cbe0a1d12146 100644
--- a/src/os/bluestore/BlueRocksEnv.cc
+++ b/src/os/bluestore/BlueRocksEnv.cc
@@ -221,18 +221,12 @@ class BlueRocksWritableFile : public rocksdb::WritableFile {
   }
 
   rocksdb::Status Close() override {
-    fs->fsync(h);
 
-    // mimic posix env, here.  shrug.
-    size_t block_size;
-    size_t last_allocated_block;
-    GetPreallocationStatus(&block_size, &last_allocated_block);
-    if (last_allocated_block > 0) {
-      int r = fs->truncate(h, h->pos);
-      if (r < 0)
-	return err_to_status(r);
+    int r = fs->truncate(h, h->pos);
+    if (r < 0) {
+      return err_to_status(r);
     }
-
+    fs->fsync(h);
     return rocksdb::Status::OK();
   }
 
-- 
2.39.5