public inbox for pbs-devel@lists.proxmox.com
 help / color / mirror / Atom feed
From: Christian Ebner <c.ebner@proxmox.com>
To: pbs-devel@lists.proxmox.com
Subject: [pbs-devel] [PATCH proxmox-backup v2 10/12] api: chunk upload: fix race between chunk backend upload and insert
Date: Wed,  8 Oct 2025 17:21:23 +0200	[thread overview]
Message-ID: <20251008152125.849216-11-c.ebner@proxmox.com> (raw)
In-Reply-To: <20251008152125.849216-1-c.ebner@proxmox.com>

Chunk are first uploaded to the object store for S3 backed
datastores, only then inserted into the local datastore cache.
By this, it is assured that the chunk is only ever considered as
valid after successful upload. Garbage collection does however rely
on the local marker file to be present, only then considering the chunk
as in-use. While this marker is created if not present during phase 1
of garbage collection, this only happens for chunks which are already
referenced by a complete index file.

Therefore, there remains a race window between garbage collection
listing the chunks (upload completed) and lookup of the local marker
file being present (chunk cache insert after upload). This can lead to
chunks which just finished upload, but were not yet inserted into the
local cache store to be removed again.

To close this race window, mark chunks which are currently being
uploaded to the backend by an additional marker file, checked by the
garbage collection as well if the regular marker is not found, and
only removed after successful chunk insert.

Concurrent chunk upload is still possible, only creating the upload
marker file if not already present. There is still a race between the
concurrent uploads, however since they encode for the same data
(except compression level) both are valid, even when the winner of
the backup upload is not the winner of the chunk insert. However, if
one of the concurrent uploads fails, the other one must fail chunk
insertion and the backup job fail as well, as then there is no
guarantee that garbage collection could not have cleaned up the chunk
in the mean time, even if such a constellation is very unlikely.

Avoid this overhead for regular datastores by passing and checking
the backend type for the chunk store.

Signed-off-by: Christian Ebner <c.ebner@proxmox.com>
---
 pbs-datastore/src/chunk_store.rs              | 136 +++++++++++++++++-
 pbs-datastore/src/datastore.rs                |  32 +++++
 .../src/local_datastore_lru_cache.rs          |   3 +-
 src/api2/backup/upload_chunk.rs               |  19 ++-
 src/api2/config/datastore.rs                  |   2 +
 5 files changed, 179 insertions(+), 13 deletions(-)

diff --git a/pbs-datastore/src/chunk_store.rs b/pbs-datastore/src/chunk_store.rs
index 65d74d68e..5b1f397bd 100644
--- a/pbs-datastore/src/chunk_store.rs
+++ b/pbs-datastore/src/chunk_store.rs
@@ -7,7 +7,7 @@ use std::time::Duration;
 use anyhow::{bail, format_err, Context, Error};
 use tracing::{info, warn};
 
-use pbs_api_types::{DatastoreFSyncLevel, GarbageCollectionStatus};
+use pbs_api_types::{DatastoreBackendType, DatastoreFSyncLevel, GarbageCollectionStatus};
 use proxmox_io::ReadExt;
 use proxmox_s3_client::S3Client;
 use proxmox_sys::fs::{create_dir, create_path, file_type_from_file_stat, CreateOptions};
@@ -22,6 +22,10 @@ use crate::file_formats::{
 };
 use crate::DataBlob;
 
+/// Filename extension for local datastore cache marker files indicating
+/// the chunk being uploaded to the backend
+const BACKEND_UPLOAD_MARKER_EXT: &str = "backend-upload";
+
 /// File system based chunk store
 pub struct ChunkStore {
     name: String, // used for error reporting
@@ -30,6 +34,7 @@ pub struct ChunkStore {
     mutex: Mutex<()>,
     locker: Option<Arc<Mutex<ProcessLocker>>>,
     sync_level: DatastoreFSyncLevel,
+    datastore_backend_type: DatastoreBackendType,
 }
 
 // TODO: what about sysctl setting vm.vfs_cache_pressure (0 - 100) ?
@@ -77,6 +82,7 @@ impl ChunkStore {
             mutex: Mutex::new(()),
             locker: None,
             sync_level: Default::default(),
+            datastore_backend_type: DatastoreBackendType::default(),
         }
     }
 
@@ -97,6 +103,7 @@ impl ChunkStore {
         uid: nix::unistd::Uid,
         gid: nix::unistd::Gid,
         sync_level: DatastoreFSyncLevel,
+        datastore_backend_type: DatastoreBackendType,
     ) -> Result<Self, Error>
     where
         P: Into<PathBuf>,
@@ -151,7 +158,7 @@ impl ChunkStore {
             }
         }
 
-        Self::open(name, base, sync_level)
+        Self::open(name, base, sync_level, datastore_backend_type)
     }
 
     fn lockfile_path<P: Into<PathBuf>>(base: P) -> PathBuf {
@@ -185,6 +192,7 @@ impl ChunkStore {
         name: &str,
         base: P,
         sync_level: DatastoreFSyncLevel,
+        datastore_backend_type: DatastoreBackendType,
     ) -> Result<Self, Error> {
         let base: PathBuf = base.into();
 
@@ -201,6 +209,7 @@ impl ChunkStore {
             locker: Some(locker),
             mutex: Mutex::new(()),
             sync_level,
+            datastore_backend_type,
         })
     }
 
@@ -557,10 +566,96 @@ impl ChunkStore {
         Ok(())
     }
 
+    // Try to insert a new backend upload marker, signaling to garbage collection that there is an
+    // in-progress upload for this chunk.
+    //
+    // Returns true if the marker was created or pre-existed (concurrent upload), returns false
+    // if the chunk has been inserted since, the marker file not being created and the upload
+    // must be avoided.
+    pub(crate) fn insert_backend_upload_marker(&self, digest: &[u8; 32]) -> Result<bool, Error> {
+        if self.datastore_backend_type == DatastoreBackendType::Filesystem {
+            bail!("cannot create backend upload marker, not a cache store");
+        }
+        let (marker_path, digest_str) = self.chunk_backed_upload_marker_path(digest);
+
+        let _lock = self.mutex.lock();
+
+        if self.cond_touch_chunk(digest, false)? {
+            return Ok(false);
+        }
+
+        if let Err(err) = std::fs::File::options()
+            .write(true)
+            .create_new(true)
+            .open(&marker_path)
+        {
+            if err.kind() != std::io::ErrorKind::AlreadyExists {
+                return Err(err).with_context(|| {
+                    format!("failed to create backend upload marker for chunk {digest_str}")
+                });
+            }
+        }
+        Ok(true)
+    }
+
+    pub(crate) fn cleanup_backend_upload_marker(&self, digest: &[u8; 32]) -> Result<(), Error> {
+        if self.datastore_backend_type == DatastoreBackendType::Filesystem {
+            bail!("cannot cleanup backend upload marker, not a cache store");
+        }
+        let (marker_path, _digest_str) = self.chunk_backed_upload_marker_path(digest);
+
+        let _lock = self.mutex.lock();
+
+        if let Err(err) = std::fs::remove_file(marker_path) {
+            if err.kind() != std::io::ErrorKind::NotFound {
+                bail!("failed to cleanup backend upload marker: {err}");
+            }
+        }
+        Ok(())
+    }
+
+    pub(crate) fn backend_upload_marker_exists(&self, digest: &[u8; 32]) -> Result<bool, Error> {
+        let (marker_path, _digest_str) = self.chunk_backed_upload_marker_path(digest);
+        if let Err(err) = std::fs::metadata(marker_path) {
+            if err.kind() != std::io::ErrorKind::NotFound {
+                bail!("failed to check backend upload marker: {err}");
+            }
+            return Ok(false);
+        }
+        Ok(true)
+    }
+
     pub fn insert_chunk(&self, chunk: &DataBlob, digest: &[u8; 32]) -> Result<(bool, u64), Error> {
         self.insert_chunk_impl(chunk, digest, |_, _| Ok(()))
     }
 
+    pub fn insert_chunk_and_remove_upload_marker(
+        &self,
+        chunk: &DataBlob,
+        digest: &[u8; 32],
+    ) -> Result<(bool, u64), Error> {
+        self.insert_chunk_impl(chunk, digest, |digest, pre_existing| {
+            if self.datastore_backend_type != DatastoreBackendType::Filesystem {
+                // Must fail as if the marker is no longer present but the chunk file is not
+                // present, as this indicates that a concurrent upload failed, no guarantee that
+                // garbage collection did not cleanup chunks in the mean time.
+                let (chunk_path, digest_str) = self.chunk_path(digest);
+                if let Err(err) =
+                    std::fs::remove_file(chunk_path.with_extension(BACKEND_UPLOAD_MARKER_EXT))
+                {
+                    if !(pre_existing && err.kind() == std::io::ErrorKind::NotFound) {
+                        bail!(
+                            "concurrent upload failed on store '{}' for {digest_str} - {err}",
+                            self.name,
+                        )
+                    }
+                }
+            }
+
+            Ok(())
+        })
+    }
+
     fn insert_chunk_impl(
         &self,
         chunk: &DataBlob,
@@ -692,6 +787,15 @@ impl ChunkStore {
         Ok(())
     }
 
+    /// Generate file path for backend upload marker of given chunk digest.
+    fn chunk_backed_upload_marker_path(&self, digest: &[u8; 32]) -> (PathBuf, String) {
+        let (chunk_path, digest_str) = self.chunk_path(digest);
+        (
+            chunk_path.with_extension(BACKEND_UPLOAD_MARKER_EXT),
+            digest_str,
+        )
+    }
+
     pub fn relative_path(&self, path: &Path) -> PathBuf {
         // unwrap: only `None` in unit tests
         assert!(self.locker.is_some());
@@ -777,14 +881,26 @@ fn test_chunk_store1() {
 
     if let Err(_e) = std::fs::remove_dir_all(".testdir") { /* ignore */ }
 
-    let chunk_store = ChunkStore::open("test", &path, DatastoreFSyncLevel::None);
+    let chunk_store = ChunkStore::open(
+        "test",
+        &path,
+        DatastoreFSyncLevel::None,
+        DatastoreBackendType::Filesystem,
+    );
     assert!(chunk_store.is_err());
 
     let user = nix::unistd::User::from_uid(nix::unistd::Uid::current())
         .unwrap()
         .unwrap();
-    let chunk_store =
-        ChunkStore::create("test", &path, user.uid, user.gid, DatastoreFSyncLevel::None).unwrap();
+    let chunk_store = ChunkStore::create(
+        "test",
+        &path,
+        user.uid,
+        user.gid,
+        DatastoreFSyncLevel::None,
+        DatastoreBackendType::Filesystem,
+    )
+    .unwrap();
 
     let (chunk, digest) = crate::data_blob::DataChunkBuilder::new(&[0u8, 1u8])
         .build()
@@ -796,8 +912,14 @@ fn test_chunk_store1() {
     let (exists, _) = chunk_store.insert_chunk(&chunk, &digest).unwrap();
     assert!(exists);
 
-    let chunk_store =
-        ChunkStore::create("test", &path, user.uid, user.gid, DatastoreFSyncLevel::None);
+    let chunk_store = ChunkStore::create(
+        "test",
+        &path,
+        user.uid,
+        user.gid,
+        DatastoreFSyncLevel::None,
+        DatastoreBackendType::Filesystem,
+    );
     assert!(chunk_store.is_err());
 
     if let Err(_e) = std::fs::remove_dir_all(".testdir") { /* ignore */ }
diff --git a/pbs-datastore/src/datastore.rs b/pbs-datastore/src/datastore.rs
index a6b17e3c3..e40b6883b 100644
--- a/pbs-datastore/src/datastore.rs
+++ b/pbs-datastore/src/datastore.rs
@@ -336,10 +336,14 @@ impl DataStore {
                 DatastoreTuning::API_SCHEMA
                     .parse_property_string(config.tuning.as_deref().unwrap_or(""))?,
             )?;
+            let backend_config: DatastoreBackendConfig =
+                config.backend.as_deref().unwrap_or("").parse()?;
+            let backend_type = backend_config.ty.unwrap_or_default();
             Arc::new(ChunkStore::open(
                 name,
                 config.absolute_path(),
                 tuning.sync_level.unwrap_or_default(),
+                backend_type,
             )?)
         };
 
@@ -424,10 +428,16 @@ impl DataStore {
             DatastoreTuning::API_SCHEMA
                 .parse_property_string(config.tuning.as_deref().unwrap_or(""))?,
         )?;
+        let backend_config: DatastoreBackendConfig = serde_json::from_value(
+            DatastoreBackendConfig::API_SCHEMA
+                .parse_property_string(config.backend.as_deref().unwrap_or(""))?,
+        )?;
+        let backend_type = backend_config.ty.unwrap_or_default();
         let chunk_store = ChunkStore::open(
             &name,
             config.absolute_path(),
             tuning.sync_level.unwrap_or_default(),
+            backend_type,
         )?;
         let inner = Arc::new(Self::with_store_and_config(
             Arc::new(chunk_store),
@@ -1651,6 +1661,13 @@ impl DataStore {
                     let atime = match std::fs::metadata(&chunk_path) {
                         Ok(stat) => stat.accessed()?,
                         Err(err) if err.kind() == std::io::ErrorKind::NotFound => {
+                            if self
+                                .inner
+                                .chunk_store
+                                .backend_upload_marker_exists(&digest)?
+                            {
+                                continue;
+                            }
                             // File not found, delete by setting atime to unix epoch
                             info!("Not found, mark for deletion: {}", content.key);
                             SystemTime::UNIX_EPOCH
@@ -1857,6 +1874,21 @@ impl DataStore {
         self.inner.chunk_store.insert_chunk(chunk, digest)
     }
 
+    // Try to insert a new backend upload marker, signaling to garbage collection that there is an
+    // in-progress upload for this chunk.
+    //
+    // Returns true if the marker was created or pre-existed (concurrent upload), returns false
+    // if the chunk has been inserted since, the marker file not being created and the upload
+    // must be avoided.
+    pub fn insert_backend_upload_marker(&self, digest: &[u8; 32]) -> Result<bool, Error> {
+        self.inner.chunk_store.insert_backend_upload_marker(digest)
+    }
+
+    /// Remove the marker file signaling an in-progress upload to the backend
+    pub fn cleanup_backend_upload_marker(&self, digest: &[u8; 32]) -> Result<(), Error> {
+        self.inner.chunk_store.cleanup_backend_upload_marker(digest)
+    }
+
     pub fn stat_chunk(&self, digest: &[u8; 32]) -> Result<std::fs::Metadata, Error> {
         let (chunk_path, _digest_str) = self.inner.chunk_store.chunk_path(digest);
         std::fs::metadata(chunk_path).map_err(Error::from)
diff --git a/pbs-datastore/src/local_datastore_lru_cache.rs b/pbs-datastore/src/local_datastore_lru_cache.rs
index fe3b51a55..cdad77031 100644
--- a/pbs-datastore/src/local_datastore_lru_cache.rs
+++ b/pbs-datastore/src/local_datastore_lru_cache.rs
@@ -34,7 +34,8 @@ impl LocalDatastoreLruCache {
     ///
     /// Fails if the chunk cannot be inserted successfully.
     pub fn insert(&self, digest: &[u8; 32], chunk: &DataBlob) -> Result<(), Error> {
-        self.store.insert_chunk(chunk, digest)?;
+        self.store
+            .insert_chunk_and_remove_upload_marker(chunk, digest)?;
         self.cache
             .insert(*digest, (), |digest| self.store.clear_chunk(&digest))
     }
diff --git a/src/api2/backup/upload_chunk.rs b/src/api2/backup/upload_chunk.rs
index 8dd7e4d52..7d1f863ed 100644
--- a/src/api2/backup/upload_chunk.rs
+++ b/src/api2/backup/upload_chunk.rs
@@ -259,6 +259,7 @@ async fn upload_to_backend(
                     data.len()
                 );
             }
+            let datastore = env.datastore.clone();
 
             if env.no_cache {
                 let object_key = pbs_datastore::s3::object_key_from_digest(&digest)?;
@@ -272,11 +273,11 @@ async fn upload_to_backend(
             // Avoid re-upload to S3 if the chunk is either present in the LRU cache or the chunk
             // file exists on filesystem. The latter means that the chunk has been present in the
             // past an was not cleaned up by garbage collection, so contained in the S3 object store.
-            if env.datastore.cache_contains(&digest) {
+            if datastore.cache_contains(&digest) {
                 tracing::info!("Skip upload of cached chunk {}", hex::encode(digest));
                 return Ok((digest, size, encoded_size, true));
             }
-            if let Ok(true) = env.datastore.cond_touch_chunk(&digest, false) {
+            if let Ok(true) = datastore.cond_touch_chunk(&digest, false) {
                 tracing::info!(
                     "Skip upload of already encountered chunk {}",
                     hex::encode(digest)
@@ -286,16 +287,24 @@ async fn upload_to_backend(
 
             tracing::info!("Upload of new chunk {}", hex::encode(digest));
             let object_key = pbs_datastore::s3::object_key_from_digest(&digest)?;
-            let is_duplicate = s3_client
+            if !datastore.insert_backend_upload_marker(&digest)? {
+                return Ok((digest, size, encoded_size, true));
+            }
+            let is_duplicate = match s3_client
                 .upload_no_replace_with_retry(object_key, data.clone())
                 .await
-                .map_err(|err| format_err!("failed to upload chunk to s3 backend - {err:#}"))?;
+            {
+                Ok(is_duplicate) => is_duplicate,
+                Err(err) => {
+                    datastore.cleanup_backend_upload_marker(&digest)?;
+                    bail!("failed to upload chunk to s3 backend - {err:#}");
+                }
+            };
 
             // Only insert the chunk into the cache after it has been successufuly uploaded.
             // Although less performant than doing this in parallel, it is required for consisency
             // since chunks are considered as present on the backend if the file exists in the local
             // cache store.
-            let datastore = env.datastore.clone();
             tracing::info!("Caching of chunk {}", hex::encode(digest));
             let _ = tokio::task::spawn_blocking(move || {
                 let chunk = DataBlob::from_raw(data.to_vec())?;
diff --git a/src/api2/config/datastore.rs b/src/api2/config/datastore.rs
index 3b03c0466..541bd0a04 100644
--- a/src/api2/config/datastore.rs
+++ b/src/api2/config/datastore.rs
@@ -173,6 +173,7 @@ pub(crate) fn do_create_datastore(
                 &datastore.name,
                 &path,
                 tuning.sync_level.unwrap_or_default(),
+                backend_type,
             )
         })?
     } else {
@@ -204,6 +205,7 @@ pub(crate) fn do_create_datastore(
             backup_user.uid,
             backup_user.gid,
             tuning.sync_level.unwrap_or_default(),
+            backend_type,
         )?
     };
 
-- 
2.47.3



_______________________________________________
pbs-devel mailing list
pbs-devel@lists.proxmox.com
https://lists.proxmox.com/cgi-bin/mailman/listinfo/pbs-devel


  parent reply	other threads:[~2025-10-08 15:21 UTC|newest]

Thread overview: 13+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2025-10-08 15:21 [pbs-devel] [PATCH proxmox-backup v2 00/12] s3 store: fix issues with chunk s3 backend upload and cache eviction Christian Ebner
2025-10-08 15:21 ` [pbs-devel] [PATCH proxmox-backup v2 01/12] datastore: gc: inline single callsite method Christian Ebner
2025-10-08 15:21 ` [pbs-devel] [PATCH proxmox-backup v2 02/12] gc: chunk store: rework atime check and gc status into common helper Christian Ebner
2025-10-08 15:21 ` [pbs-devel] [PATCH proxmox-backup v2 03/12] chunk store: add unsafe signature to cache remove method Christian Ebner
2025-10-08 15:21 ` [pbs-devel] [PATCH proxmox-backup v2 04/12] local store cache: replace evicted cache chunks instead of truncate Christian Ebner
2025-10-08 15:21 ` [pbs-devel] [PATCH proxmox-backup v2 05/12] local store cache: serve response fetched from s3 backend Christian Ebner
2025-10-08 15:21 ` [pbs-devel] [PATCH proxmox-backup v2 06/12] local store cache: refactor fetch and insert of chunks for " Christian Ebner
2025-10-08 15:21 ` [pbs-devel] [PATCH proxmox-backup v2 07/12] local store cache: rework access cache fetching and insert logic Christian Ebner
2025-10-08 15:21 ` [pbs-devel] [PATCH proxmox-backup v2 08/12] local store cache: drop obsolete cacher implementation Christian Ebner
2025-10-08 15:21 ` [pbs-devel] [PATCH proxmox-backup v2 09/12] chunk store: refactor method for chunk insertion Christian Ebner
2025-10-08 15:21 ` Christian Ebner [this message]
2025-10-08 15:21 ` [pbs-devel] [PATCH proxmox-backup v2 11/12] api: chunk upload: fix race with garbage collection for no-cache on s3 Christian Ebner
2025-10-08 15:21 ` [pbs-devel] [PATCH proxmox-backup v2 12/12] pull: guard chunk upload and only insert into cache after upload Christian Ebner

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=20251008152125.849216-11-c.ebner@proxmox.com \
    --to=c.ebner@proxmox.com \
    --cc=pbs-devel@lists.proxmox.com \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox
Service provided by Proxmox Server Solutions GmbH | Privacy | Legal