From: Adam Kupczyk <akupczyk@ibm.com>
Date: Thu, 27 Nov 2025 11:02:08 +0000 (+0000)
Subject: os/bluestore/bluefs: Make sure write buffer is aligned
X-Git-Url: http://git-server-git.apps.pok.os.sepia.ceph.com/?a=commitdiff_plain;h=dafa173733cc073597fa5e5be4c82a99c8730cbb;p=ceph.git

os/bluestore/bluefs: Make sure write buffer is aligned

In envelope mode header could cause write buffer to go unaligned.
Fixes it.

Fixes: https://tracker.ceph.com/issues/74010

Signed-off-by: Adam Kupczyk <akupczyk@ibm.com>
---

diff --git a/src/include/buffer.h b/src/include/buffer.h
index 1f1604e43bce..73b7309c15ad 100644
--- a/src/include/buffer.h
+++ b/src/include/buffer.h
@@ -890,6 +890,10 @@ struct error_code;
 	bl.obtain_contiguous_space(0);
       }
 
+      void refill() {
+        _refill(min_alloc);
+      }
+
       void append(const char* buf, size_t entire_len) {
 	 _append_common(entire_len,
 			[buf, this] (const size_t chunk_len) mutable {
diff --git a/src/os/bluestore/BlueFS.cc b/src/os/bluestore/BlueFS.cc
index 6d1ea61086b4..252c062771a1 100644
--- a/src/os/bluestore/BlueFS.cc
+++ b/src/os/bluestore/BlueFS.cc
@@ -3851,6 +3851,17 @@ ceph::bufferlist BlueFS::FileWriter::flush_buffer(
     // The alternative approach would be to place the entire tail and
     // padding on a dedicated, 4 KB long memory chunk. This shouldn't
     // trigger the rebuild while still being less expensive.
+    if (file->envelope_mode() &&
+      buffer.get_append_buffer_unused_tail_length() < tail + File::envelope_t::head_size()) {
+      // Envelope mode header must completely fit in single buffer::ptr,
+      // otherwise append_hole() will allocate new unaligned buffer.
+      // |4K  ......... ~6k ........  8K-1 | 8K .............................. 12K|
+      // <-----tail----><00000000000000000> <-----buffer_unused_tail_length------->
+      // <------to-be-flushed-to-disk-----> <----------tail-----------><env::head_size>
+      // Clearing the buffer is a way to force buffer_appender to allocate fresh
+      // pages. The size is min 2 * super.block_size so header will fit.
+      buffer.clear();
+    }
     buffer_appender.substr_of(bl, bl.length() - padding_len - tail, tail);
     buffer.splice(buffer.length() - tail, tail, &tail_block);
   } else {
@@ -4117,8 +4128,8 @@ void BlueFS::append_try_flush(FileWriter *h, const char* buf, size_t len)/*_WF_L
   {
     std::unique_lock hl(h->lock);
     if (h->file->envelope_mode() && h->get_buffer_length() == 0) {
-      uint32_t pos1 = h->get_effective_write_pos();
       h->envelope_head_filler = h->append_hole(File::envelope_t::head_size());
+      uint32_t pos1 = h->get_effective_write_pos() - File::envelope_t::head_size();
       uint32_t pos2 = reinterpret_cast<uintptr_t>(h->envelope_head_filler.c_str());
       ceph_assert(p2aligned(pos1 ^ pos2, CEPH_PAGE_SIZE));
     }
@@ -4724,7 +4735,7 @@ int BlueFS::open_for_write(
 
 BlueFS::FileWriter *BlueFS::_create_writer(FileRef f)
 {
-  FileWriter *w = new FileWriter(f);
+  FileWriter *w = new FileWriter(f, super.block_size);
   for (unsigned i = 0; i < MAX_BDEV; ++i) {
     if (bdev[i]) {
       w->iocv[i] = new IOContext(cct, NULL);
diff --git a/src/os/bluestore/BlueFS.h b/src/os/bluestore/BlueFS.h
index 8b766a5ef60d..450d6766acba 100644
--- a/src/os/bluestore/BlueFS.h
+++ b/src/os/bluestore/BlueFS.h
@@ -432,10 +432,11 @@ public:
     std::array<IOContext*,MAX_BDEV> iocv; ///< for each bdev
     std::array<bool, MAX_BDEV> dirty_devs;
 
-    FileWriter(FileRef f)
-      : file(std::move(f)),
-       buffer_appender(buffer.get_page_aligned_appender(
-                         g_conf()->bluefs_alloc_size / CEPH_PAGE_SIZE)), envelope_head_filler() {
+    FileWriter(FileRef f, unsigned super_block_size)
+      : file(std::move(f))
+      , buffer_appender(buffer.get_page_aligned_appender(
+        std::max<uint64_t>(g_conf()->bluefs_alloc_size, 2 * super_block_size) / CEPH_PAGE_SIZE))
+      , envelope_head_filler() {
       ++file->num_writers;
       iocv.fill(nullptr);
       dirty_devs.fill(false);
@@ -479,6 +480,10 @@ public:
     }
 
     bufferlist::contiguous_filler append_hole(uint64_t len) {
+      if (buffer.get_append_buffer_unused_tail_length() < len) {
+        ceph_assert(buffer.length() == 0);
+        buffer_appender.refill();
+      }
       return buffer.append_hole(len);
     }