[PATCH proxmox-backup v2 0/4] fix sync level updates for chunk store

[PATCH proxmox-backup v2 0/4] fix sync level updates for chunk store

[Christian Ebner]

Currently the sync level configured for a datastore is only
propagated to the chunk store on first lookup, when the chunk
store instance is created. Updating the sync level in the tuning
options does invalidate the cached datastore entry, does however
not re-instantiate the chunk store, to avoid dropping the locks
acquired via the process locker. This means however the sync level
on the chunk store is not updated.

Fixed by storing the chunk store state inside the mutex, already
present for syncing up concurrent access to the chunk store. This
also improves the code style and fixes a few smaller issues
encountered.

Changes since version 1 (thanks @Robert for feedback!):
- Fix lost sync level update, move implementation into chunk store
  instead
- Only try ensuring the sync level if it actually changed and fix race
  window by taking the mutex guard for the check.
- Avoid Arc for chunk store mutex, it can be shared by reference
- Fixed typo


proxmox-backup:

Christian Ebner (4):
  datastore: restrict chunk store mutex scope to crate only
  datastore: avoid useless double borrowing of datastore
  datastore: move try_ensure_sync_level() implementation to chunk store
  datastore: fix sync level update propagation to chunk store

 pbs-datastore/src/chunk_store.rs              | 82 ++++++++++++++-----
 pbs-datastore/src/datastore.rs                | 52 +++++++-----
 .../src/local_datastore_lru_cache.rs          | 12 +--
 3 files changed, 98 insertions(+), 48 deletions(-)


Summary over all repositories:
  3 files changed, 98 insertions(+), 48 deletions(-)

-- 
Generated by murpp 0.11.0

[PATCH proxmox-backup v2 1/4] datastore: restrict chunk store mutex scope to crate only

[Christian Ebner]

The mutex should never be used outside of the datastore
implementation, all operations on it go through other method calls.

Restrict the scope for the method to get the mutex.

Signed-off-by: Christian Ebner <c.ebner@proxmox.com>
---
changes since version 1:
- no changes

 pbs-datastore/src/chunk_store.rs | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/pbs-datastore/src/chunk_store.rs b/pbs-datastore/src/chunk_store.rs
index 68db88eab..2888dea39 100644
--- a/pbs-datastore/src/chunk_store.rs
+++ b/pbs-datastore/src/chunk_store.rs
@@ -423,7 +423,7 @@ impl ChunkStore {
         ProcessLocker::oldest_shared_lock(self.locker.clone().unwrap())
     }
 
-    pub fn mutex(&self) -> &std::sync::Mutex<()> {
+    pub(crate) fn mutex(&self) -> &std::sync::Mutex<()> {
         &self.mutex
     }
 
-- 
2.47.3

[PATCH proxmox-backup v2 2/4] datastore: avoid useless double borrowing of datastore

[Christian Ebner]

Getting the datastore entry from the cache already gives a reference,
no need to borrow again.

Signed-off-by: Christian Ebner <c.ebner@proxmox.com>
---
changes since version 1:
- no changes

 pbs-datastore/src/datastore.rs | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/pbs-datastore/src/datastore.rs b/pbs-datastore/src/datastore.rs
index 34efcd398..bc218cb94 100644
--- a/pbs-datastore/src/datastore.rs
+++ b/pbs-datastore/src/datastore.rs
@@ -588,7 +588,7 @@ impl DataStore {
         let entry = datastore_cache.get(lookup.name);
 
         // reuse chunk store so that we keep using the same process locker instance!
-        let chunk_store = if let Some(datastore) = &entry {
+        let chunk_store = if let Some(datastore) = entry {
             // Re-use DataStoreImpl
             if datastore.config_generation == gen_num && gen_num.is_some() {
                 update_active_operations(lookup.name, lookup.operation, 1)?;
-- 
2.47.3

[PATCH proxmox-backup v2 3/4] datastore: move try_ensure_sync_level() implementation to chunk store

[Christian Ebner]

Keep the public interface for the datastore in place, but move the
method to the chunk store, dropping the now unneeded state stored on
DataStoreImpl. The path for the syncfs call does not change by the
move as the previous implementation already passed the chunk store
base path.

While moving, adapt the log message to use the tracing macro.

This is in preparation for fixing the sync level updates not being
propagated to the chunk store.

Signed-off-by: Christian Ebner <c.ebner@proxmox.com>
---
changes since version 1:
- move implementation to chunk store, thereby fixing logic error in
  check.

 pbs-datastore/src/chunk_store.rs | 13 +++++++++++++
 pbs-datastore/src/datastore.rs   | 22 +++++-----------------
 2 files changed, 18 insertions(+), 17 deletions(-)

diff --git a/pbs-datastore/src/chunk_store.rs b/pbs-datastore/src/chunk_store.rs
index 2888dea39..e8c279a62 100644
--- a/pbs-datastore/src/chunk_store.rs
+++ b/pbs-datastore/src/chunk_store.rs
@@ -966,6 +966,19 @@ impl ChunkStore {
         }
         (chunk_path, counter)
     }
+
+    pub(super) fn try_ensure_sync_level(&self) -> Result<(), Error> {
+        if self.sync_level != DatastoreFSyncLevel::Filesystem {
+            return Ok(());
+        }
+        let file = std::fs::File::open(self.base_path())?;
+        let fd = file.as_raw_fd();
+        info!("syncing filesystem");
+        if unsafe { libc::syncfs(fd) } < 0 {
+            bail!("error during syncfs: {}", std::io::Error::last_os_error());
+        }
+        Ok(())
+    }
 }
 
 #[derive(PartialEq)]
diff --git a/pbs-datastore/src/datastore.rs b/pbs-datastore/src/datastore.rs
index bc218cb94..6338298a0 100644
--- a/pbs-datastore/src/datastore.rs
+++ b/pbs-datastore/src/datastore.rs
@@ -32,9 +32,9 @@ use proxmox_worker_task::WorkerTaskContext;
 
 use pbs_api_types::{
     ArchiveType, Authid, BackupGroupDeleteStats, BackupNamespace, BackupType, ChunkOrder,
-    DataStoreConfig, DatastoreBackendConfig, DatastoreBackendType, DatastoreFSyncLevel,
-    DatastoreTuning, GarbageCollectionCacheStats, GarbageCollectionStatus, MaintenanceMode,
-    MaintenanceType, Operation, S3Statistics, MAX_NAMESPACE_DEPTH, UPID,
+    DataStoreConfig, DatastoreBackendConfig, DatastoreBackendType, DatastoreTuning,
+    GarbageCollectionCacheStats, GarbageCollectionStatus, MaintenanceMode, MaintenanceType,
+    Operation, S3Statistics, MAX_NAMESPACE_DEPTH, UPID,
 };
 use pbs_config::s3::S3_CFG_TYPE_ID;
 use pbs_config::{BackupLockGuard, ConfigVersionCache};
@@ -204,7 +204,6 @@ pub struct DataStoreImpl {
     last_gc_status: Mutex<GarbageCollectionStatus>,
     verify_new: bool,
     chunk_order: ChunkOrder,
-    sync_level: DatastoreFSyncLevel,
     backend_config: DatastoreBackendConfig,
     lru_store_caching: Option<LocalDatastoreLruCache>,
     thread_settings: DatastoreThreadSettings,
@@ -225,7 +224,6 @@ impl DataStoreImpl {
             last_gc_status: Mutex::new(GarbageCollectionStatus::default()),
             verify_new: false,
             chunk_order: Default::default(),
-            sync_level: Default::default(),
             backend_config: Default::default(),
             lru_store_caching: None,
             thread_settings: Default::default(),
@@ -785,7 +783,6 @@ impl DataStore {
             last_gc_status: Mutex::new(gc_status),
             verify_new: config.verify_new.unwrap_or(false),
             chunk_order: tuning.chunk_order.unwrap_or_default(),
-            sync_level: tuning.sync_level.unwrap_or_default(),
             backend_config,
             lru_store_caching,
             thread_settings,
@@ -3012,19 +3009,10 @@ impl DataStore {
     }
     */
 
-    /// Syncs the filesystem of the datastore if 'sync_level' is set to
+    /// Syncs the filesystem of the chunk store base path if 'sync_level' is set to
     /// [`DatastoreFSyncLevel::Filesystem`]. Uses syncfs(2).
     pub fn try_ensure_sync_level(&self) -> Result<(), Error> {
-        if self.inner.sync_level != DatastoreFSyncLevel::Filesystem {
-            return Ok(());
-        }
-        let file = std::fs::File::open(self.base_path())?;
-        let fd = file.as_raw_fd();
-        log::info!("syncing filesystem");
-        if unsafe { libc::syncfs(fd) } < 0 {
-            bail!("error during syncfs: {}", std::io::Error::last_os_error());
-        }
-        Ok(())
+        self.inner.chunk_store.try_ensure_sync_level()
     }
 
     /// Destroy a datastore. This requires that there are no active operations on the datastore.
-- 
2.47.3

[PATCH proxmox-backup v2 4/4] datastore: fix sync level update propagation to chunk store

[Christian Ebner]

Changing the datastore tuning options triggers an invalidation of the
datastore cache entry, leading to re-instantiation with the new
config parameters on the next datastore lookup.
Since commit 0bd9c8701 ("datastore: lookup: reuse ChunkStore on stale
datastore re-open") this does however not lead to re-creation of the
chunk store instance in order to avoid dropping the process locker,
which would lead to loosing any existing shared lock. However, as a
consequence the sync level is not updated on the chunk store.

Fix this by:
- Storing the sync level as runtime properties of the chunk store as
  state within the mutex syncing concurrent modify access. It is held
  where needed anyways.
- Pass the mutex guard as additional parameters to the methods
  requiring the locked state. This encodes the requirement for the
  mutex guard directly into the function signature instead of
  labeling it as unsafe only.
- Assuring the previous sync level on config changes for consistency.
  To do this, extend and rename try_ensure_sync_level to also take the
  mutex guard and update the sync level if given.

Reported-by: Robert Obkircher <r.obkircher@proxmox.com>
Signed-off-by: Christian Ebner <c.ebner@proxmox.com>
---
changes since version 1:
- Only ensure sync level on config updates if changed in tuninig options
- Always perform checking/updating with held mutex guard
- Avoid using unneeded Arc for sharing mutex
- Fix typo in commit message and adapt

 pbs-datastore/src/chunk_store.rs              | 73 +++++++++++++------
 pbs-datastore/src/datastore.rs                | 30 +++++++-
 .../src/local_datastore_lru_cache.rs          | 12 +--
 3 files changed, 82 insertions(+), 33 deletions(-)

diff --git a/pbs-datastore/src/chunk_store.rs b/pbs-datastore/src/chunk_store.rs
index e8c279a62..6c97e31d7 100644
--- a/pbs-datastore/src/chunk_store.rs
+++ b/pbs-datastore/src/chunk_store.rs
@@ -1,7 +1,7 @@
 use std::os::unix::fs::MetadataExt;
 use std::os::unix::io::AsRawFd;
 use std::path::{Path, PathBuf};
-use std::sync::{Arc, Mutex};
+use std::sync::{Arc, Mutex, MutexGuard};
 use std::time::Duration;
 
 use anyhow::{bail, format_err, Context, Error};
@@ -27,14 +27,20 @@ use crate::{DataBlob, LocalDatastoreLruCache};
 
 const USING_MARKER_FILENAME_EXT: &str = "using";
 
+#[derive(Default)]
+/// Configurable runtime properties of a chunk store
+pub(crate) struct ChunkStoreProperties {
+    pub(crate) sync_level: DatastoreFSyncLevel,
+}
+
 /// File system based chunk store
 pub struct ChunkStore {
     name: String, // used for error reporting
     pub(crate) base: PathBuf,
     chunk_dir: PathBuf,
-    mutex: Mutex<()>,
+    // Mutex to sync chunk store access, including property updates
+    mutex: Arc<Mutex<ChunkStoreProperties>>,
     locker: Option<Arc<Mutex<ProcessLocker>>>,
-    sync_level: DatastoreFSyncLevel,
 }
 
 // TODO: what about sysctl setting vm.vfs_cache_pressure (0 - 100) ?
@@ -79,9 +85,8 @@ impl ChunkStore {
             name: String::new(),
             base: PathBuf::new(),
             chunk_dir: PathBuf::new(),
-            mutex: Mutex::new(()),
+            mutex: Arc::new(Mutex::new(Default::default())),
             locker: None,
-            sync_level: Default::default(),
         }
     }
 
@@ -204,16 +209,19 @@ impl ChunkStore {
             base,
             chunk_dir,
             locker: Some(locker),
-            mutex: Mutex::new(()),
-            sync_level,
+            mutex: Arc::new(Mutex::new(ChunkStoreProperties { sync_level })),
         })
     }
 
-    fn touch_chunk_no_lock(&self, digest: &[u8; 32]) -> Result<(), Error> {
+    fn touch_chunk_no_lock(
+        &self,
+        digest: &[u8; 32],
+        mutex_guard: MutexGuard<ChunkStoreProperties>,
+    ) -> Result<(), Error> {
         // unwrap: only `None` in unit tests
         assert!(self.locker.is_some());
 
-        self.cond_touch_chunk_no_lock(digest, true)?;
+        self.cond_touch_chunk_no_lock(digest, true, mutex_guard)?;
         Ok(())
     }
 
@@ -226,14 +234,15 @@ impl ChunkStore {
         digest: &[u8; 32],
         assert_exists: bool,
     ) -> Result<bool, Error> {
-        let _lock = self.mutex.lock();
-        self.cond_touch_chunk_no_lock(digest, assert_exists)
+        let lock = self.mutex.lock().unwrap();
+        self.cond_touch_chunk_no_lock(digest, assert_exists, lock)
     }
 
     fn cond_touch_chunk_no_lock(
         &self,
         digest: &[u8; 32],
         assert_exists: bool,
+        _mutex_guard: MutexGuard<ChunkStoreProperties>,
     ) -> Result<bool, Error> {
         // unwrap: only `None` in unit tests
         assert!(self.locker.is_some());
@@ -423,7 +432,11 @@ impl ChunkStore {
         ProcessLocker::oldest_shared_lock(self.locker.clone().unwrap())
     }
 
-    pub(crate) fn mutex(&self) -> &std::sync::Mutex<()> {
+    /// Mutex to lock chunk store for exclusive access.
+    ///
+    /// Must be held when modifying chunk store contents and allows to update
+    /// chunk store runtime properties.
+    pub(crate) fn mutex(&self) -> &Mutex<ChunkStoreProperties> {
         &self.mutex
     }
 
@@ -665,18 +678,17 @@ impl ChunkStore {
 
         //println!("DIGEST {}", hex::encode(digest));
 
-        let _lock = self.mutex.lock();
+        let lock = self.mutex.lock().unwrap();
 
-        // Safety: lock acquired above
-        unsafe { self.insert_chunk_nolock(chunk, digest, true) }
+        self.insert_chunk_nolock(chunk, digest, true, lock)
     }
 
-    /// Safety: requires holding the chunk store mutex!
-    pub(crate) unsafe fn insert_chunk_nolock(
+    pub(crate) fn insert_chunk_nolock(
         &self,
         chunk: &DataBlob,
         digest: &[u8; 32],
         warn_on_overwrite_empty: bool,
+        mutex_guard: MutexGuard<ChunkStoreProperties>,
     ) -> Result<(bool, u64), Error> {
         // unwrap: only `None` in unit tests
         assert!(self.locker.is_some());
@@ -694,7 +706,7 @@ impl ChunkStore {
             }
             let old_size = metadata.len();
             if encoded_size == old_size {
-                self.touch_chunk_no_lock(digest)?;
+                self.touch_chunk_no_lock(digest, mutex_guard)?;
                 return Ok((true, old_size));
             } else if old_size == 0 {
                 if warn_on_overwrite_empty {
@@ -721,11 +733,11 @@ impl ChunkStore {
                 // compressed, the size mismatch could be caused by different zstd versions
                 // so let's keep the one that was uploaded first, bit-rot is hopefully detected by
                 // verification at some point..
-                self.touch_chunk_no_lock(digest)?;
+                self.touch_chunk_no_lock(digest, mutex_guard)?;
                 return Ok((true, old_size));
             } else if old_size < encoded_size {
                 log::debug!("Got another copy of chunk with digest '{digest_str}', existing chunk is smaller, discarding uploaded one.");
-                self.touch_chunk_no_lock(digest)?;
+                self.touch_chunk_no_lock(digest, mutex_guard)?;
                 return Ok((true, old_size));
             } else {
                 log::debug!("Got another copy of chunk with digest '{digest_str}', existing chunk is bigger, replacing with uploaded one.");
@@ -742,17 +754,18 @@ impl ChunkStore {
             let gid = pbs_config::backup_group()?.gid;
             create_options = create_options.owner(uid).group(gid);
         }
+
         proxmox_sys::fs::replace_file(
             &chunk_path,
             raw_data,
             create_options,
-            self.sync_level == DatastoreFSyncLevel::File,
+            mutex_guard.sync_level == DatastoreFSyncLevel::File,
         )
         .map_err(|err| {
             format_err!("inserting chunk on store '{name}' failed for {digest_str} - {err}")
         })?;
 
-        if self.sync_level == DatastoreFSyncLevel::File {
+        if mutex_guard.sync_level == DatastoreFSyncLevel::File {
             // fsync dir handle to persist the tmp rename
             let dir = std::fs::File::open(chunk_dir_path)?;
             nix::unistd::fsync(dir.as_raw_fd())
@@ -967,10 +980,22 @@ impl ChunkStore {
         (chunk_path, counter)
     }
 
-    pub(super) fn try_ensure_sync_level(&self) -> Result<(), Error> {
-        if self.sync_level != DatastoreFSyncLevel::Filesystem {
+    pub(super) fn try_ensure_sync_level_with_update(
+        &self,
+        mut mutex_guard: MutexGuard<ChunkStoreProperties>,
+        update_sync_level: Option<DatastoreFSyncLevel>,
+    ) -> Result<(), Error> {
+        let assure_sync_level = mutex_guard.sync_level;
+        if let Some(sync_level) = update_sync_level {
+            mutex_guard.sync_level = sync_level;
+        }
+
+        drop(mutex_guard); // never hold during syncfs
+
+        if assure_sync_level != DatastoreFSyncLevel::Filesystem {
             return Ok(());
         }
+
         let file = std::fs::File::open(self.base_path())?;
         let fd = file.as_raw_fd();
         info!("syncing filesystem");
diff --git a/pbs-datastore/src/datastore.rs b/pbs-datastore/src/datastore.rs
index 6338298a0..adef4dfe3 100644
--- a/pbs-datastore/src/datastore.rs
+++ b/pbs-datastore/src/datastore.rs
@@ -595,12 +595,30 @@ impl DataStore {
                     operation: Some(lookup.operation),
                 }));
             }
+
+            let tuning: DatastoreTuning = serde_json::from_value(
+                DatastoreTuning::API_SCHEMA
+                    .parse_property_string(config.tuning.as_deref().unwrap_or(""))?,
+            )?;
+            let sync_level = tuning.sync_level.unwrap_or_default();
+
+            let mutex = datastore.chunk_store.mutex();
+            let mutex_guard = mutex.lock().unwrap();
+            if mutex_guard.sync_level != sync_level {
+                datastore
+                    .chunk_store
+                    .try_ensure_sync_level_with_update(mutex_guard, Some(sync_level))?;
+            } else {
+                drop(mutex_guard);
+            }
+
             Arc::clone(&datastore.chunk_store)
         } else {
             let tuning: DatastoreTuning = serde_json::from_value(
                 DatastoreTuning::API_SCHEMA
                     .parse_property_string(config.tuning.as_deref().unwrap_or(""))?,
             )?;
+
             Arc::new(ChunkStore::open(
                 lookup.name,
                 config.absolute_path(),
@@ -2483,7 +2501,8 @@ impl DataStore {
                         Ok(guard) => guard,
                         Err(_) => continue,
                     };
-                    let _guard = self.inner.chunk_store.mutex().lock().unwrap();
+                    let mutex = self.inner.chunk_store.mutex();
+                    let _guard = mutex.lock().unwrap();
 
                     // Check local markers (created or atime updated during phase1) and
                     // keep or delete chunk based on that.
@@ -3012,7 +3031,11 @@ impl DataStore {
     /// Syncs the filesystem of the chunk store base path if 'sync_level' is set to
     /// [`DatastoreFSyncLevel::Filesystem`]. Uses syncfs(2).
     pub fn try_ensure_sync_level(&self) -> Result<(), Error> {
-        self.inner.chunk_store.try_ensure_sync_level()
+        let mutex = self.inner.chunk_store.mutex();
+        let mutex_guard = mutex.lock().unwrap();
+        self.inner
+            .chunk_store
+            .try_ensure_sync_level_with_update(mutex_guard, None)
     }
 
     /// Destroy a datastore. This requires that there are no active operations on the datastore.
@@ -3434,7 +3457,8 @@ impl DataStore {
                 .chunk_store
                 .lock_chunk(digest, CHUNK_LOCK_TIMEOUT)?;
         }
-        let _lock = self.inner.chunk_store.mutex().lock().unwrap();
+        let mutex = self.inner.chunk_store.mutex();
+        let _lock = mutex.lock().unwrap();
 
         let (new_path, counter) = self.inner.chunk_store.next_bad_chunk_path(digest);
 
diff --git a/pbs-datastore/src/local_datastore_lru_cache.rs b/pbs-datastore/src/local_datastore_lru_cache.rs
index ac27d4637..cf38d4a57 100644
--- a/pbs-datastore/src/local_datastore_lru_cache.rs
+++ b/pbs-datastore/src/local_datastore_lru_cache.rs
@@ -34,12 +34,11 @@ impl LocalDatastoreLruCache {
     ///
     /// Fails if the chunk cannot be inserted successfully.
     pub fn insert(&self, digest: &[u8; 32], chunk: &DataBlob) -> Result<(), Error> {
-        let _lock = self.store.mutex().lock().unwrap();
+        let shared_mutex = self.store.mutex();
+        let lock = shared_mutex.lock().unwrap();
+
+        self.store.insert_chunk_nolock(chunk, digest, false, lock)?;
 
-        // Safety: lock acquire above
-        unsafe {
-            self.store.insert_chunk_nolock(chunk, digest, false)?;
-        }
         self.cache.insert(*digest, (), |digest| {
             // Safety: lock acquired above, this is executed inline!
             unsafe {
@@ -80,7 +79,8 @@ impl LocalDatastoreLruCache {
             Ok(mut file) => match DataBlob::load_from_reader(&mut file) {
                 // File was still cached with contents, load response from file
                 Ok(chunk) => {
-                    let _lock = self.store.mutex().lock().unwrap();
+                    let shared_mutex = self.store.mutex();
+                    let _lock = shared_mutex.lock().unwrap();
                     self.cache.insert(*digest, (), |digest| {
                         // Safety: lock acquired above, this is executed inline
                         unsafe {
-- 
2.47.3