fuse: Use -EDEADLK for DLM lock error instead of -EAGAIN

-EAGAIN is semantically overloaded for a DLM error and not
self describing, switch to -EDEADLK.
In order to allow a graceful daemon change, -EAGAIN is kept for now.

Signed-off-by: Bernd Schubert <bernd@bsbernd.com>

Author: bsbernd

Documentation/filesystems/fuse/fuse-AOP_TRUNCATED_PAGE-reason.txt

@@ -1,5 +1,5 @@
 =================================================================================
-WHY FUSE CONVERTS -EAGAIN TO AOP_TRUNCATED_PAGE IN fuse_read_folio()
+WHY FUSE CONVERTS -EDEADLK TO AOP_TRUNCATED_PAGE IN fuse_read_folio()
 =================================================================================
 
 TLDR: To prevent ABBA deadlock between page locks and DLM (cluster) locks.
@@ -292,21 +292,22 @@ Core-0: Application reading from FUSE filesystem
                          │  Sends FUSE_READ to userspace daemon
                          │  Daemon may need to acquire cluster lock
                          │
-                         └─ Might return -EAGAIN if DLM detects possible
+                         └─ Might return -EDEADLK if DLM detects possible
                             deadlock due to concurrant page invalidation
+                            - For now -EAGAIN handled the same
 
 
 ─────────────────────────────────────────────────────────────────────────────
 WHY FUSE NEEDS THIS CONVERSION
 ─────────────────────────────────────────────────────────────────────────────
 
-When fuse_simple_request() returns -EAGAIN:
+When fuse_simple_request() returns -EDEADLK:
   - Userspace FUSE daemon encountered transient failure
   - Could be: cluster lock contention, timeout, daemon busy
   - The page might be stale/modified during the wait
   - Page lock is STILL HELD at this point
 
-If -EAGAIN were returned directly:
+If -EDEADLK were returned directly:
   ❌ VFS would see error and fail the read
   ❌ Page would remain locked
   ❌ No retry mechanism triggered
@@ -342,11 +343,12 @@ Location: fs/fuse/file.c:947-964
    out:
       // LINE 962: ✓ CRITICAL - Always unlocks page before returning
       folio_unlock(folio);
-      return err;  // Returns AOP_TRUNCATED_PAGE if -EAGAIN occurred
+      return err;  // Returns AOP_TRUNCATED_PAGE if -EDEADLK occurred
   }
 
 This matches OCFS2's pattern:
   1. Detect lock contention (-EAGAIN from daemon)
+    - Note: confusing that OCFS2 uses -EAGAIN instead of -EDEADLK
   2. Unlock the page (line 962)
   3. Return AOP_TRUNCATED_PAGE
   4. VFS retries the operation
@@ -372,7 +374,7 @@ Callers in mm/filemap.c handle it consistently:
   - do_filemap_fault(): "if (error == AOP_TRUNCATED_PAGE) goto retry_find;"
 
 =================================================================================
-WHY NOT JUST RETURN -EAGAIN?
+WHY NOT JUST RETURN -EDEADLK (-EAGAIN)?
 =================================================================================
 
 From OCFS2 comments (fs/ocfs2/dlmglue.c:2555-2560):
@@ -434,7 +436,7 @@ T2     Call: ocfs2_read_folio()        DLM downconvert starts
                                         Lock: DLM inode lock
 
 T3     Try: DLM lock (NONBLOCK)        Want: page lock
-       Returns: -EAGAIN                ❌ BLOCKED (held by Core-0)
+       Returns: -EDEADLK               ❌ BLOCKED (held by Core-0)
 
 T4     ✓ Unlock: page lock             ✓ Acquires: page lock
        Return: AOP_TRUNCATED_PAGE      Continues: reclaim work
@@ -517,10 +519,10 @@ Caller 3: do_filemap_fault() - lines 3542-3543
 Effect: Re-handles the page fault from scratch
 
 =================================================================================
-SUMMARY: Why -EAGAIN → AOP_TRUNCATED_PAGE is Essential
+SUMMARY: Why -EDEADLK → AOP_TRUNCATED_PAGE is Essential
 =================================================================================
 
-The conversion -EAGAIN → AOP_TRUNCATED_PAGE in fuse_read_folio() is necessary:
+The conversion -EDEADLK → AOP_TRUNCATED_PAGE in fuse_read_folio() is necessary:
 
 1. **Prevent Deadlock**: Avoids page lock vs cluster lock ABBA deadlock
    - Core-0 holds page lock, wants DLM lock
@@ -535,7 +537,7 @@ The conversion -EAGAIN → AOP_TRUNCATED_PAGE in fuse_read_folio() is necessary:
 3. **Enable Retry**: Uses VFS's built-in retry mechanism properly
    - AOP_TRUNCATED_PAGE is understood by mm/filemap.c
    - Triggers automatic retry loops at multiple call sites
-   - -EAGAIN alone would just fail the operation
+   - -EDEADLK alone would just fail the operation
 
 4. **Ensure Fairness**: Allows fair lock acquisition through retry cycle
    - Other threads get chance to acquire locks
@@ -553,7 +555,7 @@ Core-0 (Read Path):
   - Entry: filemap_read()                        [mm/filemap.c:2675]
   - Page lock: folio_trylock()                   [mm/filemap.c:2466]
   - FUSE read: fuse_do_readfolio()               [fs/fuse/file.c:905]
-  - -EAGAIN conversion                           [fs/fuse/file.c:934-935]
+  - -EDEADLK conversion                           [fs/fuse/file.c:934-935]
   - Page unlock: folio_unlock()                  [fs/fuse/file.c:962]
 
 Core-1 (Reclaim Path):

fs/fuse/file.c

@@ -959,11 +959,14 @@ static int fuse_do_readfolio(struct file *file, struct folio *folio,
 	res = fuse_simple_request(fm, &ia.ap.args);
 	if (res < 0) {
 		/*
-		 * please refer to Documentation/filesystems/fuse/fuse-AOP_TRUNCATED_PAGE-reason.txt
-		 * why READ can return -EAGAIN from DLM subsystem.
-		 * XXX find a better DLM specific error code
+		 * Please refer to Documentation/filesystems/fuse/fuse-AOP_TRUNCATED_PAGE-reason.txt
+		 * why READ can return -EDEADLK from DLM subsystem.
+		 *
+		 * -EDEADLK: Preferred error code indicating DLM lock ordering violation
+		 *           (would cause deadlock with page lock)
+		 * -EAGAIN:  Legacy error code, maintained for backward compatibility
 		 */
-		if (res == -EAGAIN && fm->fc->dlm)
+		if ((res == -EDEADLK || res == -EAGAIN) && fm->fc->dlm)
 			res = AOP_TRUNCATED_PAGE;
 		return res;
 	}
@@ -1010,9 +1013,9 @@ static int fuse_iomap_read_folio_range(const struct iomap_iter *iter,
 	/*
 	 * TEMPORARY WORKAROUND for iomap write deadlock:
 	 *
-	 * When FUSE server returns -EAGAIN due to DLM,
-	 * fuse_do_readfolio() converts it to AOP_TRUNCATED_PAGE and
-	 * unlocks the folio (per AOP_TRUNCATED_PAGE contract).
+	 * When FUSE server returns -EDEADLK (or legacy -EAGAIN) due to DLM
+	 * lock contention, fuse_do_readfolio() converts it to AOP_TRUNCATED_PAGE
+	 * and unlocks the folio (per AOP_TRUNCATED_PAGE contract).
 	 *
 	 * However, iomap doesn't understand AOP_TRUNCATED_PAGE.
 	 * We need to: