mdbx: use SysV semaphores on systems without shared mutexes.

Change-Id: Ib2ad9ed137ab76999a2a8e832f9f77ff1a0788ca

src/elements/core.c

@@ -7494,7 +7494,7 @@ int __cold mdbx_env_create(MDBX_env **penv) {
     goto bailout;
   }
 
-#if MDBX_LOCKING > 0
+#if MDBX_LOCKING > MDBX_LOCKING_SYSV
   rc = mdbx_ipclock_stub(&env->me_lckless_stub.wlock);
 #endif /* MDBX_LOCKING */
   if (unlikely(rc != MDBX_SUCCESS)) {
@@ -8338,7 +8338,7 @@ static int __cold mdbx_setup_lck(MDBX_env *env, char *lck_pathname,
     env->me_maxreaders = UINT_MAX;
 #if MDBX_LOCKING > 0
     env->me_wlock = &env->me_lckless_stub.wlock;
-#endif /* MDBX_LOCKING */
+#endif /* MDBX_LOCKING > 0 */
     mdbx_debug("lck-setup:%s%s%s", " lck-less",
                (env->me_flags & MDBX_RDONLY) ? " readonly" : "",
                (rc == MDBX_RESULT_TRUE) ? " exclusive" : " cooperative");
@@ -8472,7 +8472,7 @@ static int __cold mdbx_setup_lck(MDBX_env *env, char *lck_pathname,
   env->me_meta_sync_txnid = &lck->mti_meta_sync_txnid;
 #if MDBX_LOCKING > 0
   env->me_wlock = &lck->mti_wlock;
-#endif /* MDBX_LOCKING */
+#endif /* MDBX_LOCKING > 0 */
   return lck_seize_rc;
 }
 
@@ -8704,6 +8704,14 @@ int __cold mdbx_env_open(MDBX_env *env, const char *path, unsigned flags,
   if (rc != MDBX_SUCCESS)
     goto bailout;
 
+#if MDBX_LOCKING == MDBX_LOCKING_SYSV
+  env->me_sysv_ipc.key = ftok(dxb_pathname, 42);
+  if (env->me_sysv_ipc.key == -1) {
+    rc = errno;
+    goto bailout;
+  }
+#endif /* MDBX_LOCKING */
+
   const int lck_rc = mdbx_setup_lck(env, lck_pathname, mode);
   if (MDBX_IS_ERROR(lck_rc)) {
     rc = lck_rc;
@@ -8930,7 +8938,7 @@ int __cold mdbx_env_close_ex(MDBX_env *env, int dont_sync) {
               mdbx_fastmutex_destroy(&env->me_remap_guard) == MDBX_SUCCESS);
 #endif /* Windows */
 
-#if MDBX_LOCKING > 0
+#if MDBX_LOCKING > MDBX_LOCKING_SYSV
   mdbx_ensure(env, mdbx_ipclock_destroy(&env->me_lckless_stub.wlock) == 0);
 #endif /* MDBX_LOCKING */
 

src/elements/internals.h

@@ -204,9 +204,11 @@
 #endif /* MDBX_64BIT_CAS */
 
 #define MDBX_LOCKING_WIN32FILES -1
+#define MDBX_LOCKING_SYSV 5         /* SystemV IPC semaphores */
 #define MDBX_LOCKING_POSIX1988 1988 /* POSIX-1 Shared anonymous semaphores */
 #define MDBX_LOCKING_POSIX2001 2001 /* POSIX-2001 Shared Mutexes */
 #define MDBX_LOCKING_POSIX2008 2008 /* POSIX-2008 Robust Mutexes */
+#define MDBX_LOCKING_BENAPHORE 1995 /* BeOS Benaphores, aka Futexes */
 
 #if defined(_WIN32) || defined(_WIN64)
 #define MDBX_LOCKING MDBX_LOCKING_WIN32FILES
@@ -227,8 +229,10 @@
 #else
 #define MDBX_LOCKING MDBX_LOCKING_POSIX2001
 #endif
-#else
+#elif defined(__sun) || defined(__SVR4) || defined(__svr4__)
 #define MDBX_LOCKING MDBX_LOCKING_POSIX1988
+#else
+#define MDBX_LOCKING MDBX_LOCKING_SYSV
 #endif
 #define MDBX_LOCKING_CONFIG "AUTO=" STRINGIFY(MDBX_LOCKING)
 #else
@@ -507,8 +511,18 @@ typedef struct MDBX_page {
 
 #pragma pack(pop)
 
-#if MDBX_LOCKING > 0
+#if MDBX_LOCKING == MDBX_LOCKING_WIN32FILES
-#if MDBX_LOCKING == MDBX_LOCKING_POSIX2001 ||                                  \
+#define MDBX_CLOCK_SIGN UINT32_C(0xF10C)
+typedef void mdbx_ipclock_t;
+#elif MDBX_LOCKING == MDBX_LOCKING_SYSV
+
+#define MDBX_CLOCK_SIGN UINT32_C(0xF18D)
+typedef mdbx_pid_t mdbx_ipclock_t;
+#ifndef EOWNERDEAD
+#define EOWNERDEAD MDBX_RESULT_TRUE
+#endif
+
+#elif MDBX_LOCKING == MDBX_LOCKING_POSIX2001 ||                                \
     MDBX_LOCKING == MDBX_LOCKING_POSIX2008
 #define MDBX_CLOCK_SIGN UINT32_C(0x8017)
 typedef pthread_mutex_t mdbx_ipclock_t;
@@ -517,13 +531,11 @@ typedef pthread_mutex_t mdbx_ipclock_t;
 typedef sem_t mdbx_ipclock_t;
 #else
 #error "FIXME"
-#endif
+#endif /* MDBX_LOCKING */
 
+#if MDBX_LOCKING > MDBX_LOCKING_SYSV
 MDBX_INTERNAL_FUNC int mdbx_ipclock_stub(mdbx_ipclock_t *ipc);
 MDBX_INTERNAL_FUNC int mdbx_ipclock_destroy(mdbx_ipclock_t *ipc);
-
-#else
-#define MDBX_CLOCK_SIGN UINT32_C(0xF10C)
 #endif /* MDBX_LOCKING */
 
 /* Reader Lock Table
@@ -642,10 +654,10 @@ typedef struct MDBX_lockinfo {
 
   alignas(MDBX_CACHELINE_SIZE) /* cacheline ---------------------------------*/
 
-  /* Write transation lok. */
+  /* Write transation lock. */
 #if MDBX_LOCKING > 0
       mdbx_ipclock_t mti_wlock;
-#endif /* MDBX_LOCKING */
+#endif /* MDBX_LOCKING > 0 */
 
   volatile txnid_t mti_oldest_reader;
 
@@ -668,7 +680,7 @@ typedef struct MDBX_lockinfo {
   /* Readeaders registration lock. */
 #if MDBX_LOCKING > 0
       mdbx_ipclock_t mti_rlock;
-#endif /* MDBX_LOCKING */
+#endif /* MDBX_LOCKING > 0 */
 
   /* The number of slots that have been used in the reader table.
    * This always records the maximum count, it is not decremented
@@ -1020,9 +1032,16 @@ struct MDBX_env {
   MDBX_txn *me_txn0;          /* prealloc'd write transaction */
 
   /* write-txn lock */
+#if MDBX_LOCKING == MDBX_LOCKING_SYSV
+  union {
+    key_t key;
+    int semid;
+  } me_sysv_ipc;
+#endif /* MDBX_LOCKING == MDBX_LOCKING_SYSV */
+
 #if MDBX_LOCKING > 0
   mdbx_ipclock_t *me_wlock;
-#endif /* MDBX_LOCKING */
+#endif /* MDBX_LOCKING > 0 */
 
   MDBX_dbx *me_dbxs;           /* array of static DB info */
   uint16_t *me_dbflags;        /* array of flags from MDBX_db.md_flags */
@@ -1050,7 +1069,7 @@ struct MDBX_env {
   struct {
 #if MDBX_LOCKING > 0
     mdbx_ipclock_t wlock;
-#endif /* MDBX_LOCKING */
+#endif /* MDBX_LOCKING > 0 */
     txnid_t oldest;
     uint64_t sync_timestamp;
     uint64_t autosync_period;

src/elements/lck-posix.c

@@ -13,6 +13,7 @@
  */
 
 #include "internals.h"
+#include <sys/sem.h>
 
 /*----------------------------------------------------------------------------*/
 /* global constructor/destructor */
@@ -195,7 +196,7 @@ MDBX_INTERNAL_FUNC int mdbx_rpid_check(MDBX_env *env, uint32_t pid) {
 
 /*---------------------------------------------------------------------------*/
 
-#if MDBX_LOCKING > 0
+#if MDBX_LOCKING > MDBX_LOCKING_SYSV
 MDBX_INTERNAL_FUNC int mdbx_ipclock_stub(mdbx_ipclock_t *ipc) {
 #if MDBX_LOCKING == MDBX_LOCKING_POSIX1988
   return sem_init(ipc, false, 1) ? errno : 0;
@@ -217,7 +218,7 @@ MDBX_INTERNAL_FUNC int mdbx_ipclock_destroy(mdbx_ipclock_t *ipc) {
 #error "FIXME"
 #endif
 }
-#endif /* MDBX_LOCKING */
+#endif /* MDBX_LOCKING > MDBX_LOCKING_SYSV */
 
 MDBX_INTERNAL_FUNC int __cold mdbx_lck_seize(MDBX_env *env) {
   assert(env->me_fd != INVALID_HANDLE_VALUE);
@@ -365,7 +366,10 @@ MDBX_INTERNAL_FUNC int __cold mdbx_lck_destroy(MDBX_env *env,
              OFF_T_MAX) == 0) {
 
     mdbx_verbose("%s: got exclusive, drown locks", __func__);
-#if MDBX_LOCKING > 0
+#if MDBX_LOCKING == MDBX_LOCKING_SYSV
+    if (env->me_sysv_ipc.semid != -1)
+      rc = semctl(env->me_sysv_ipc.semid, 2, IPC_RMID) ? errno : 0;
+#else
     rc = mdbx_ipclock_destroy(&env->me_lck->mti_rlock);
     if (rc == 0)
       rc = mdbx_ipclock_destroy(&env->me_lck->mti_wlock);
@@ -431,8 +435,58 @@ MDBX_INTERNAL_FUNC int __cold mdbx_lck_init(MDBX_env *env,
     return MDBX_SUCCESS /* currently don't need any initialization
       if LCK already opened/used inside current process */
         ;
+#if MDBX_LOCKING == MDBX_LOCKING_SYSV
+  int semid = -1;
+  if (global_uniqueness_flag) {
+    struct stat st;
+    if (fstat(env->me_fd, &st))
+      return errno;
+  sysv_retry_create:
+    semid = semget(env->me_sysv_ipc.key, 2,
+                   IPC_CREAT | IPC_EXCL |
+                       (st.st_mode & (S_IRWXU | S_IRWXG | S_IRWXO)));
+    if (unlikely(semid == -1)) {
+      int err = errno;
+      if (err != EEXIST)
+        return err;
 
-#if MDBX_LOCKING == MDBX_LOCKING_POSIX1988
+      /* remove and re-create semaphore set */
+      semid = semget(env->me_sysv_ipc.key, 2, 0);
+      if (semid == -1) {
+        err = errno;
+        if (err != ENOENT)
+          return err;
+        goto sysv_retry_create;
+      }
+      if (semctl(semid, 2, IPC_RMID)) {
+        err = errno;
+        if (err != EIDRM)
+          return err;
+      }
+      goto sysv_retry_create;
+    }
+
+    unsigned short val_array[2] = {1, 1};
+    if (semctl(semid, 2, SETALL, val_array))
+      return errno;
+  } else {
+    semid = semget(env->me_sysv_ipc.key, 2, 0);
+    if (semid == -1)
+      return errno;
+
+    /* check read & write access */
+    struct semid_ds data[2];
+    if (semctl(semid, 2, IPC_STAT, data) || semctl(semid, 2, IPC_SET, data))
+      return errno;
+  }
+
+  env->me_sysv_ipc.semid = semid;
+
+  return MDBX_SUCCESS;
+
+#elif MDBX_LOCKING == MDBX_LOCKING_FUTEX
+#warning "TODO"
+#elif MDBX_LOCKING == MDBX_LOCKING_POSIX1988
 
   /* don't initialize semaphores twice */
   if (global_uniqueness_flag == MDBX_RESULT_TRUE) {
@@ -519,7 +573,7 @@ bailout:
 static int __cold mdbx_ipclock_failed(MDBX_env *env, mdbx_ipclock_t *ipc,
                                       const int err) {
   int rc = err;
-#if MDBX_LOCKING == MDBX_LOCKING_POSIX2008
+#if MDBX_LOCKING == MDBX_LOCKING_POSIX2008 || MDBX_LOCKING == MDBX_LOCKING_SYSV
   if (err == EOWNERDEAD) {
     /* We own the mutex. Clean up after dead previous owner. */
 
@@ -533,12 +587,15 @@ static int __cold mdbx_ipclock_failed(MDBX_env *env, mdbx_ipclock_t *ipc,
         rc = MDBX_PANIC;
       }
     }
-    mdbx_notice("%cmutex owner died, %s", (rlocked ? 'r' : 'w'),
+    mdbx_notice("%clock owner died, %s", (rlocked ? 'r' : 'w'),
                 (rc ? "this process' env is hosed" : "recovering"));
 
     int check_rc = mdbx_reader_check0(env, rlocked, NULL);
     check_rc = (check_rc == MDBX_SUCCESS) ? MDBX_RESULT_TRUE : check_rc;
 
+#if MDBX_LOCKING == MDBX_LOCKING_SYSV
+    rc = (rc == MDBX_SUCCESS) ? check_rc : rc;
+#else
 #if defined(PTHREAD_MUTEX_ROBUST) || defined(pthread_mutex_consistent)
     int mreco_rc = pthread_mutex_consistent(ipc);
 #elif defined(PTHREAD_MUTEX_ROBUST_NP) || defined(pthread_mutex_consistent_np)
@@ -551,17 +608,21 @@ static int __cold mdbx_ipclock_failed(MDBX_env *env, mdbx_ipclock_t *ipc,
     check_rc = (mreco_rc == 0) ? check_rc : mreco_rc;
 
     if (unlikely(mreco_rc))
-      mdbx_error("mutex recovery failed, %s", mdbx_strerror(mreco_rc));
+      mdbx_error("lock recovery failed, %s", mdbx_strerror(mreco_rc));
 
     rc = (rc == MDBX_SUCCESS) ? check_rc : rc;
     if (MDBX_IS_ERROR(rc))
       pthread_mutex_unlock(ipc);
+#endif /* MDBX_LOCKING == MDBX_LOCKING_POSIX2008 */
     return rc;
   }
 #elif MDBX_LOCKING == MDBX_LOCKING_POSIX2001
   (void)ipc;
 #elif MDBX_LOCKING == MDBX_LOCKING_POSIX1988
   (void)ipc;
+#elif MDBX_LOCKING == MDBX_LOCKING_FUTEX
+#warning "TODO"
+  (void)ipc;
 #else
 #error "FIXME"
 #endif /* MDBX_LOCKING */
@@ -588,6 +649,19 @@ static int mdbx_ipclock_lock(MDBX_env *env, mdbx_ipclock_t *ipc,
     }
   } else if (sem_wait(ipc))
     rc = errno;
+#elif MDBX_LOCKING == MDBX_LOCKING_SYSV
+  struct sembuf op = {.sem_num = (ipc != env->me_wlock),
+                      .sem_op = -1,
+                      .sem_flg = dont_wait ? IPC_NOWAIT | SEM_UNDO : SEM_UNDO};
+  int rc;
+  if (semop(env->me_sysv_ipc.semid, &op, 1)) {
+    rc = errno;
+    if (dont_wait && rc == EAGAIN)
+      rc = MDBX_BUSY;
+  } else {
+    rc = *ipc ? EOWNERDEAD : MDBX_SUCCESS;
+    *ipc = env->me_pid;
+  }
 #else
 #error "FIXME"
 #endif /* MDBX_LOCKING */
@@ -597,12 +671,21 @@ static int mdbx_ipclock_lock(MDBX_env *env, mdbx_ipclock_t *ipc,
   return rc;
 }
 
-static int mdbx_ipclock_unlock(mdbx_ipclock_t *ipc) {
+static int mdbx_ipclock_unlock(MDBX_env *env, mdbx_ipclock_t *ipc) {
 #if MDBX_LOCKING == MDBX_LOCKING_POSIX2001 ||                                  \
     MDBX_LOCKING == MDBX_LOCKING_POSIX2008
   int rc = pthread_mutex_unlock(ipc);
+  (void)env;
 #elif MDBX_LOCKING == MDBX_LOCKING_POSIX1988
   int rc = sem_post(ipc) ? errno : MDBX_SUCCESS;
+  (void)env;
+#elif MDBX_LOCKING == MDBX_LOCKING_SYSV
+  if (unlikely(*ipc != (pid_t)env->me_pid))
+    return EPERM;
+  *ipc = 0;
+  struct sembuf op = {
+      .sem_num = (ipc != env->me_wlock), .sem_op = 1, .sem_flg = SEM_UNDO};
+  int rc = semop(env->me_sysv_ipc.semid, &op, 1) ? errno : MDBX_SUCCESS;
 #else
 #error "FIXME"
 #endif /* MDBX_LOCKING */
@@ -619,7 +702,7 @@ MDBX_INTERNAL_FUNC int mdbx_rdt_lock(MDBX_env *env) {
 
 MDBX_INTERNAL_FUNC void mdbx_rdt_unlock(MDBX_env *env) {
   mdbx_trace("%s", ">>");
-  int rc = mdbx_ipclock_unlock(&env->me_lck->mti_rlock);
+  int rc = mdbx_ipclock_unlock(env, &env->me_lck->mti_rlock);
   mdbx_trace("<< rc %d", rc);
   if (unlikely(rc != MDBX_SUCCESS))
     mdbx_panic("%s() failed: errcode %d\n", __func__, rc);
@@ -636,7 +719,7 @@ int mdbx_txn_lock(MDBX_env *env, bool dont_wait) {
 
 void mdbx_txn_unlock(MDBX_env *env) {
   mdbx_trace("%s", ">>");
-  int rc = mdbx_ipclock_unlock(env->me_wlock);
+  int rc = mdbx_ipclock_unlock(env, env->me_wlock);
   mdbx_trace("<< rc %d", rc);
   if (unlikely(rc != MDBX_SUCCESS))
     mdbx_panic("%s() failed: errcode %d\n", __func__, rc);

src/elements/osal.h

@@ -125,6 +125,9 @@
 
 #if defined(__sun) || defined(__SVR4) || defined(__svr4__)
 #include <kstat.h>
+/* On Solaris, it's easier to add a missing prototype rather than find a
+ * combination of #defines that break nothing. */
+__extern_C key_t ftok(const char *, int);
 #endif /* SunOS/Solaris */
 
 #if defined(_WIN32) || defined(_WIN64)
@@ -194,6 +197,7 @@ static inline void *mdbx_realloc(void *ptr, size_t bytes) {
 #include <semaphore.h>
 #include <signal.h>
 #include <sys/file.h>
+#include <sys/ipc.h>
 #include <sys/mman.h>
 #include <sys/param.h>
 #include <sys/stat.h>