Merge pull request #80 from Max042004/fix-fcntl-flock-ltype

Translate fcntl flock l_type between Linux and macOS

Author: jserv

src/syscall/fs.c

@@ -768,11 +768,33 @@ int64_t sys_fcntl(guest_t *g, int fd, int cmd, uint64_t arg)
         memcpy(&l_start, lflock + 8, 8); /* offset 8 due to padding */
         memcpy(&l_len, lflock + 16, 8);
 
+        /* l_type constants differ between Linux and macOS/BSD:
+         *   Linux: F_RDLCK=0, F_WRLCK=1, F_UNLCK=2
+         *   macOS: F_RDLCK=1, F_UNLCK=2, F_WRLCK=3
+         * Passing the Linux value straight through makes a Linux F_RDLCK (0)
+         * an invalid type on macOS, which fcntl() rejects with EINVAL. This is
+         * the lock POSIX databases (e.g. SQLite) take first, so it must map. */
+        short mac_type;
+        switch (l_type) {
+        case 0: /* LINUX_F_RDLCK */
+            mac_type = F_RDLCK;
+            break;
+        case 1: /* LINUX_F_WRLCK */
+            mac_type = F_WRLCK;
+            break;
+        case 2: /* LINUX_F_UNLCK */
+            mac_type = F_UNLCK;
+            break;
+        default:
+            host_fd_ref_close(&host_ref);
+            return -LINUX_EINVAL;
+        }
+
         struct flock mac_fl = {
             .l_start = l_start,
             .l_len = l_len,
             .l_pid = 0,
-            .l_type = l_type, /* F_RDLCK=0, F_WRLCK=1, F_UNLCK=2 same on both */
+            .l_type = mac_type,
             .l_whence = l_whence, /* SEEK_SET=0, SEEK_CUR=1, SEEK_END=2 same */
         };
 
@@ -784,7 +806,20 @@ int64_t sys_fcntl(guest_t *g, int fd, int cmd, uint64_t arg)
 
         /* For F_GETLK, write back the result */
         if (cmd == 5) {
-            int16_t rt = mac_fl.l_type, rw = mac_fl.l_whence;
+            /* Map macOS l_type back to Linux constants (see above). */
+            int16_t rt;
+            switch (mac_fl.l_type) {
+            case F_RDLCK:
+                rt = 0; /* LINUX_F_RDLCK */
+                break;
+            case F_WRLCK:
+                rt = 1; /* LINUX_F_WRLCK */
+                break;
+            default:
+                rt = 2; /* LINUX_F_UNLCK */
+                break;
+            }
+            int16_t rw = mac_fl.l_whence;
             int64_t rs = mac_fl.l_start, rl = mac_fl.l_len;
             int32_t rp = mac_fl.l_pid;
             memset(lflock, 0, sizeof(lflock));

tests/manifest.txt

@@ -42,6 +42,7 @@ test-socket
 
 [section] Syscall coverage tests
 test-file-ops
+test-flock
 test-sysinfo
 test-io-opt
 test-syscall-smoke

tests/test-flock.c

@@ -0,0 +1,99 @@
+/* Test POSIX advisory record locking via fcntl(F_SETLK/F_GETLK/F_SETLKW)
+ *
+ * Copyright 2026 elfuse contributors
+ * Copyright 2025 Moritz Angermann, zw3rk pte. ltd.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Regression coverage for the Linux<->macOS struct flock translation. The
+ * l_type constants differ between the two ABIs (Linux F_RDLCK=0/F_WRLCK=1,
+ * macOS F_RDLCK=1/F_WRLCK=3), so passing the guest value straight through to
+ * the host made the very first lock SQLite takes (a shared F_RDLCK) fail with
+ * EINVAL and surface as "disk I/O error". The byte offsets below mirror the
+ * ones SQLite locks around its 1GiB "pending byte".
+ */
+
+#include <errno.h>
+#include <fcntl.h>
+#include <stdlib.h>
+#include <string.h>
+#include <unistd.h>
+
+#include "test-harness.h"
+
+#define PENDING_BYTE 0x40000000L
+#define RESERVED_BYTE (PENDING_BYTE + 1)
+#define SHARED_FIRST (PENDING_BYTE + 2)
+#define SHARED_SIZE 510
+
+static int set_lock(int fd, short type, off_t start, off_t len)
+{
+    struct flock fl = {
+        .l_type = type,
+        .l_whence = SEEK_SET,
+        .l_start = start,
+        .l_len = len,
+    };
+    return fcntl(fd, F_SETLK, &fl);
+}
+
+int main(void)
+{
+    int passes = 0, fails = 0;
+    const char *path = "/tmp/elfuse-test-flock.db";
+
+    int fd = open(path, O_RDWR | O_CREAT | O_TRUNC, 0644);
+    if (fd < 0) {
+        perror("open");
+        return 1;
+    }
+
+    /* Shared read lock -- this is the call that regressed to EINVAL. */
+    TEST("F_SETLK F_RDLCK (shared)");
+    EXPECT_EQ(set_lock(fd, F_RDLCK, SHARED_FIRST, SHARED_SIZE), 0,
+              "shared read lock rejected");
+
+    /* Promote the pending byte to a write lock, then drop it. */
+    TEST("F_SETLK F_WRLCK (pending)");
+    EXPECT_EQ(set_lock(fd, F_WRLCK, PENDING_BYTE, 1), 0,
+              "pending write lock rejected");
+
+    TEST("F_SETLK F_WRLCK (reserved)");
+    EXPECT_EQ(set_lock(fd, F_WRLCK, RESERVED_BYTE, 1), 0,
+              "reserved write lock rejected");
+
+    TEST("F_SETLK F_UNLCK (release shared)");
+    EXPECT_EQ(set_lock(fd, F_UNLCK, SHARED_FIRST, SHARED_SIZE), 0,
+              "unlock rejected");
+
+    /* Blocking variant must take the same translation path. */
+    TEST("F_SETLKW F_WRLCK");
+    struct flock wfl = {
+        .l_type = F_WRLCK,
+        .l_whence = SEEK_SET,
+        .l_start = 0,
+        .l_len = 16,
+    };
+    EXPECT_EQ(fcntl(fd, F_SETLKW, &wfl), 0, "F_SETLKW rejected");
+
+    /* F_GETLK on a region this process already write-locks must report back a
+     * Linux l_type. Linux reports F_UNLCK for locks held by the *same* owner,
+     * so the only thing we can assert portably is that the type round-trips to
+     * a valid Linux constant and the call succeeds. */
+    TEST("F_GETLK round-trips l_type");
+    struct flock gfl = {
+        .l_type = F_WRLCK,
+        .l_whence = SEEK_SET,
+        .l_start = 0,
+        .l_len = 16,
+    };
+    int gr = fcntl(fd, F_GETLK, &gfl);
+    EXPECT_TRUE(gr == 0 && (gfl.l_type == F_UNLCK || gfl.l_type == F_RDLCK ||
+                            gfl.l_type == F_WRLCK),
+                "F_GETLK returned an invalid l_type");
+
+    close(fd);
+    unlink(path);
+
+    SUMMARY("test-flock");
+    return fails == 0 ? 0 : 1;
+}