From 600e2b6ce0b96cd442e0d5a805aa69cf0179f728 Mon Sep 17 00:00:00 2001
From: Hallvard Furuseth <hallvard@openldap.org>
Date: Fri, 3 Jul 2015 23:07:30 +0200
Subject: [PATCH] Support robust mutexes/locks. Add mdb_mutex_t etc.
---
libraries/liblmdb/lmdb.h | 15 ++-
libraries/liblmdb/mdb.c | 224 ++++++++++++++++++++++++++++-----------
2 files changed, 174 insertions(+), 65 deletions(-)
diff --git a/libraries/liblmdb/lmdb.h b/libraries/liblmdb/lmdb.h
index ee302a2365..1e934f4767 100644
--- a/libraries/liblmdb/lmdb.h
+++ b/libraries/liblmdb/lmdb.h
@@ -49,9 +49,13 @@
* stale locks can block further operation.
*
* Fix: Check for stale readers periodically, using the
- * #mdb_reader_check function or the \ref mdb_stat_1 "mdb_stat" tool. Or just
- * make all programs using the database close it; the lockfile
- * is always reset on first open of the environment.
+ * #mdb_reader_check function or the \ref mdb_stat_1 "mdb_stat" tool.
+ * Stale writers will be cleared automatically on some systems:
+ * - Windows - automatic
+ * - Linux, systems using POSIX mutexes with Robust option - automatic
+ * - not on BSD, systems using POSIX semaphores.
+ * Otherwise just make all programs using the database close it;
+ * the lockfile is always reset on first open of the environment.
*
* - On BSD systems or others configured with MDB_USE_POSIX_SEM,
* startup can fail due to semaphores owned by another userid.
@@ -106,6 +110,9 @@
* for stale readers is performed or the lockfile is reset,
* since the process may not remove it from the lockfile.
*
+ * This does not apply to write transactions if the system clears
+ * stale writers, see above.
+ *
* - If you do that anyway, do a periodic check for stale readers. Or
* close the environment once in a while, so the lockfile can get reset.
*
@@ -391,7 +398,7 @@ typedef enum MDB_cursor_op {
#define MDB_PAGE_NOTFOUND (-30797)
/** Located page was wrong type */
#define MDB_CORRUPTED (-30796)
- /** Update of meta page failed, probably I/O error */
+ /** Update of meta page failed or environment had fatal error */
#define MDB_PANIC (-30795)
/** Environment version mismatch */
#define MDB_VERSION_MISMATCH (-30794)
diff --git a/libraries/liblmdb/mdb.c b/libraries/liblmdb/mdb.c
index 1cd915042d..1cb99e7a91 100644
--- a/libraries/liblmdb/mdb.c
+++ b/libraries/liblmdb/mdb.c
@@ -224,6 +224,18 @@ extern int cacheflush(char *addr, int nbytes, int cache);
# define mdb_func_ "<mdb_unknown>"
#endif
+/* Internal error codes, not exposed outside liblmdb */
+#define MDB_NO_ROOT (MDB_LAST_ERRCODE + 10)
+#ifdef _WIN32
+#define MDB_OWNERDEAD ((int) WAIT_ABANDONED)
+#elif defined(MDB_USE_POSIX_MUTEX) && defined(EOWNERDEAD)
+#define MDB_OWNERDEAD EOWNERDEAD /**< #LOCK_MUTEX0() result if dead owner */
+#endif
+
+#ifdef MDB_OWNERDEAD
+#define MDB_ROBUST_SUPPORTED 1
+#endif
+
#ifdef _WIN32
#define MDB_USE_HASH 1
#define MDB_PIDLOCK 0
@@ -231,6 +243,7 @@ extern int cacheflush(char *addr, int nbytes, int cache);
#define pthread_t HANDLE
#define pthread_mutex_t HANDLE
#define pthread_cond_t HANDLE
+typedef HANDLE mdb_mutex_t, mdb_mutexref_t;
#define pthread_key_t DWORD
#define pthread_self() GetCurrentThreadId()
#define pthread_key_create(x,y) \
@@ -244,10 +257,9 @@ extern int cacheflush(char *addr, int nbytes, int cache);
#define pthread_cond_wait(cond,mutex) do{SignalObjectAndWait(*mutex, *cond, INFINITE, FALSE); WaitForSingleObject(*mutex, INFINITE);}while(0)
#define THREAD_CREATE(thr,start,arg) thr=CreateThread(NULL,0,start,arg,0,NULL)
#define THREAD_FINISH(thr) WaitForSingleObject(thr, INFINITE)
-#define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_rmutex)
-#define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_rmutex)
-#define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_wmutex)
-#define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_wmutex)
+#define LOCK_MUTEX0(mutex) WaitForSingleObject(mutex, INFINITE)
+#define UNLOCK_MUTEX(mutex) ReleaseMutex(mutex)
+#define mdb_mutex_consistent(mutex) 0
#define getpid() GetCurrentProcessId()
#define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
#define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
@@ -272,10 +284,9 @@ extern int cacheflush(char *addr, int nbytes, int cache);
#ifdef MDB_USE_POSIX_SEM
-#define LOCK_MUTEX_R(env) mdb_sem_wait((env)->me_rmutex)
-#define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
-#define LOCK_MUTEX_W(env) mdb_sem_wait((env)->me_wmutex)
-#define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
+typedef sem_t *mdb_mutex_t, *mdb_mutexref_t;
+#define LOCK_MUTEX0(mutex) mdb_sem_wait(mutex)
+#define UNLOCK_MUTEX(mutex) sem_post(mutex)
static int
mdb_sem_wait(sem_t *sem)
@@ -286,21 +297,25 @@ mdb_sem_wait(sem_t *sem)
}
#else /* MDB_USE_POSIX_MUTEX: */
- /** Lock the reader mutex.
+ /** Shared mutex/semaphore as it is stored (mdb_mutex_t), and as
+ * local variables keep it (mdb_mutexref_t).
+ *
+ * When #mdb_mutexref_t is a pointer declaration and #mdb_mutex_t is
+ * not, then it is array[size 1] so it can be assigned to a pointer.
+ * @{
*/
-#define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
- /** Unlock the reader mutex.
+typedef pthread_mutex_t mdb_mutex_t[1], *mdb_mutexref_t;
+ /* @} */
+ /** Lock the reader or writer mutex.
+ * Returns 0 or a code to give #mdb_mutex_failed(), as in #LOCK_MUTEX().
*/
-#define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
-
- /** Lock the writer mutex.
- * Only a single write transaction is allowed at a time. Other writers
- * will block waiting for this mutex.
+#define LOCK_MUTEX0(mutex) pthread_mutex_lock(mutex)
+ /** Unlock the reader or writer mutex.
*/
-#define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
- /** Unlock the writer mutex.
+#define UNLOCK_MUTEX(mutex) pthread_mutex_unlock(mutex)
+ /** Mark mutex-protected data as repaired, after death of previous owner.
*/
-#define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
+#define mdb_mutex_consistent(mutex) pthread_mutex_consistent(mutex)
#endif /* MDB_USE_POSIX_SEM */
/** Get the error code for the last failed system function.
@@ -334,6 +349,19 @@ mdb_sem_wait(sem_t *sem)
/** @} */
+#ifdef MDB_ROBUST_SUPPORTED
+ /** Lock mutex, handle any error, set rc = result.
+ * Return 0 on success, nonzero (not rc) on error.
+ */
+#define LOCK_MUTEX(rc, env, mutex) \
+ (((rc) = LOCK_MUTEX0(mutex)) && \
+ ((rc) = mdb_mutex_failed(env, mutex, rc)))
+static int mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc);
+#else
+#define LOCK_MUTEX(rc, env, mutex) ((rc) = LOCK_MUTEX0(mutex))
+#define mdb_mutex_failed(env, mutex, rc) (rc)
+#endif
+
#ifndef _WIN32
/** A flag for opening a file and requesting synchronous data writes.
* This is only used when writing a meta page. It's not strictly needed;
@@ -650,9 +678,9 @@ typedef struct MDB_txbody {
char mtb_rmname[MNAME_LEN];
#else
/** Mutex protecting access to this table.
- * This is the reader lock that #LOCK_MUTEX_R acquires.
+ * This is the reader table lock used with LOCK_MUTEX().
*/
- pthread_mutex_t mtb_mutex;
+ mdb_mutex_t mtb_rmutex;
#endif
/** The ID of the last transaction committed to the database.
* This is recorded here only for convenience; the value can always
@@ -672,7 +700,7 @@ typedef struct MDB_txninfo {
MDB_txbody mtb;
#define mti_magic mt1.mtb.mtb_magic
#define mti_format mt1.mtb.mtb_format
-#define mti_mutex mt1.mtb.mtb_mutex
+#define mti_rmutex mt1.mtb.mtb_rmutex
#define mti_rmname mt1.mtb.mtb_rmname
#define mti_txnid mt1.mtb.mtb_txnid
#define mti_numreaders mt1.mtb.mtb_numreaders
@@ -683,7 +711,7 @@ typedef struct MDB_txninfo {
char mt2_wmname[MNAME_LEN];
#define mti_wmname mt2.mt2_wmname
#else
- pthread_mutex_t mt2_wmutex;
+ mdb_mutex_t mt2_wmutex;
#define mti_wmutex mt2.mt2_wmutex
#endif
char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
@@ -1170,11 +1198,13 @@ struct MDB_env {
int me_live_reader; /**< have liveness lock in reader table */
#ifdef _WIN32
int me_pidquery; /**< Used in OpenProcess */
- HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
- HANDLE me_wmutex;
-#elif defined(MDB_USE_POSIX_SEM)
- sem_t *me_rmutex; /* Shared mutexes are not supported */
- sem_t *me_wmutex;
+#endif
+#ifdef MDB_USE_POSIX_MUTEX /* Posix mutexes reside in shared mem */
+# define me_rmutex me_txns->mti_rmutex /**< Shared reader lock */
+# define me_wmutex me_txns->mti_wmutex /**< Shared writer lock */
+#else
+ mdb_mutex_t me_rmutex;
+ mdb_mutex_t me_wmutex;
#endif
void *me_userctx; /**< User-settable context */
MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
@@ -1264,6 +1294,7 @@ static void mdb_xcursor_init2(MDB_cursor *mc, MDB_xcursor *src_mx, int force);
static int mdb_drop0(MDB_cursor *mc, int subs);
static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
+static int mdb_reader_check0(MDB_env *env, int rlocked, int *dead);
/** @cond */
static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
@@ -1298,7 +1329,7 @@ static char *const mdb_errstr[] = {
"MDB_NOTFOUND: No matching key/data pair found",
"MDB_PAGE_NOTFOUND: Requested page not found",
"MDB_CORRUPTED: Located page was wrong type",
- "MDB_PANIC: Update of meta page failed",
+ "MDB_PANIC: Update of meta page failed or environment had fatal error",
"MDB_VERSION_MISMATCH: Database environment version mismatch",
"MDB_INVALID: File is not an LMDB file",
"MDB_MAP_FULL: Environment mapsize limit reached",
@@ -2540,6 +2571,7 @@ mdb_txn_renew0(MDB_txn *txn)
} else {
MDB_PID_T pid = env->me_pid;
MDB_THR_T tid = pthread_self();
+ mdb_mutexref_t rmutex = env->me_rmutex;
if (!env->me_live_reader) {
rc = mdb_reader_pid(env, Pidset, pid);
@@ -2548,13 +2580,14 @@ mdb_txn_renew0(MDB_txn *txn)
env->me_live_reader = 1;
}
- LOCK_MUTEX_R(env);
+ if (LOCK_MUTEX(rc, env, rmutex))
+ return rc;
nr = ti->mti_numreaders;
for (i=0; i<nr; i++)
if (ti->mti_readers[i].mr_pid == 0)
break;
if (i == env->me_maxreaders) {
- UNLOCK_MUTEX_R(env);
+ UNLOCK_MUTEX(rmutex);
return MDB_READERS_FULL;
}
r = &ti->mti_readers[i];
@@ -2571,7 +2604,7 @@ mdb_txn_renew0(MDB_txn *txn)
ti->mti_numreaders = ++nr;
env->me_close_readers = nr;
r->mr_pid = pid;
- UNLOCK_MUTEX_R(env);
+ UNLOCK_MUTEX(rmutex);
new_notls = (env->me_flags & MDB_NOTLS);
if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
@@ -2590,8 +2623,8 @@ mdb_txn_renew0(MDB_txn *txn)
} else {
/* Not yet touching txn == env->me_txn0, it may be active */
if (ti) {
- LOCK_MUTEX_W(env);
-
+ if (LOCK_MUTEX(rc, env, env->me_wmutex))
+ return rc;
txn->mt_txnid = ti->mti_txnid;
meta = env->me_metas[txn->mt_txnid & 1];
} else {
@@ -2868,7 +2901,7 @@ mdb_txn_reset0(MDB_txn *txn, const char *act)
env->me_txn = NULL;
/* The writer mutex was locked in mdb_txn_begin. */
if (env->me_txns)
- UNLOCK_MUTEX_W(env);
+ UNLOCK_MUTEX(env->me_wmutex);
} else {
txn->mt_parent->mt_child = NULL;
env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
@@ -3468,7 +3501,7 @@ done:
mdb_dbis_update(txn, 1);
if (env->me_txns)
- UNLOCK_MUTEX_W(env);
+ UNLOCK_MUTEX(env->me_wmutex);
if (txn != env->me_txn0)
free(txn);
@@ -4519,8 +4552,11 @@ mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
if ((rc = pthread_mutexattr_init(&mattr))
|| (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
- || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
- || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
+#ifdef MDB_ROBUST_SUPPORTED
+ || (rc = pthread_mutexattr_setrobust(&mattr, PTHREAD_MUTEX_ROBUST))
+#endif
+ || (rc = pthread_mutex_init(env->me_txns->mti_rmutex, &mattr))
+ || (rc = pthread_mutex_init(env->me_txns->mti_wmutex, &mattr)))
goto fail;
pthread_mutexattr_destroy(&mattr);
#endif /* _WIN32 || MDB_USE_POSIX_SEM */
@@ -4577,8 +4613,8 @@ fail:
* environment and re-opening it with the new flags.
*/
#define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
-#define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP| \
- MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
+#define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY| \
+ MDB_WRITEMAP|MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
#if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
# error "Persistent DB flags & env flags overlap, but both go in mm_flags"
@@ -6165,7 +6201,6 @@ int
mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
unsigned int flags)
{
- enum { MDB_NO_ROOT = MDB_LAST_ERRCODE+10 }; /* internal code */
MDB_env *env;
MDB_node *leaf = NULL;
MDB_page *fp, *mp, *sub_root = NULL;
@@ -8928,6 +8963,7 @@ static int ESECT
mdb_env_copyfd0(MDB_env *env, HANDLE fd)
{
MDB_txn *txn = NULL;
+ mdb_mutexref_t wmutex = NULL;
int rc;
size_t wsize;
char *ptr;
@@ -8952,11 +8988,13 @@ mdb_env_copyfd0(MDB_env *env, HANDLE fd)
mdb_txn_reset0(txn, "reset-stage1");
/* Temporarily block writers until we snapshot the meta pages */
- LOCK_MUTEX_W(env);
+ wmutex = env->me_wmutex;
+ if (LOCK_MUTEX(rc, env, wmutex))
+ goto leave;
rc = mdb_txn_renew0(txn);
if (rc) {
- UNLOCK_MUTEX_W(env);
+ UNLOCK_MUTEX(wmutex);
goto leave;
}
}
@@ -8980,8 +9018,8 @@ mdb_env_copyfd0(MDB_env *env, HANDLE fd)
break;
}
}
- if (env->me_txns)
- UNLOCK_MUTEX_W(env);
+ if (wmutex)
+ UNLOCK_MUTEX(wmutex);
if (rc)
goto leave;
@@ -9654,17 +9692,22 @@ mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
int ESECT
mdb_reader_check(MDB_env *env, int *dead)
{
- unsigned int i, j, rdrs;
- MDB_reader *mr;
- MDB_PID_T *pids, pid;
- int count = 0;
-
if (!env)
return EINVAL;
if (dead)
*dead = 0;
- if (!env->me_txns)
- return MDB_SUCCESS;
+ return env->me_txns ? mdb_reader_check0(env, 0, dead) : MDB_SUCCESS;
+}
+
+/** As #mdb_reader_check(). rlocked = <caller locked the reader mutex>. */
+static int mdb_reader_check0(MDB_env *env, int rlocked, int *dead)
+{
+ mdb_mutexref_t rmutex = rlocked ? NULL : env->me_rmutex;
+ unsigned int i, j, rdrs;
+ MDB_reader *mr;
+ MDB_PID_T *pids, pid;
+ int rc = MDB_SUCCESS, count = 0;
+
rdrs = env->me_txns->mti_numreaders;
pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
if (!pids)
@@ -9672,22 +9715,32 @@ mdb_reader_check(MDB_env *env, int *dead)
pids[0] = 0;
mr = env->me_txns->mti_readers;
for (i=0; i<rdrs; i++) {
- if (mr[i].mr_pid && mr[i].mr_pid != env->me_pid) {
- pid = mr[i].mr_pid;
+ pid = mr[i].mr_pid;
+ if (pid && pid != env->me_pid) {
if (mdb_pid_insert(pids, pid) == 0) {
if (!mdb_reader_pid(env, Pidcheck, pid)) {
- LOCK_MUTEX_R(env);
- /* Recheck, a new process may have reused pid */
- if (!mdb_reader_pid(env, Pidcheck, pid)) {
- for (j=i; j<rdrs; j++)
+ /* Stale reader found */
+ j = i;
+ if (rmutex) {
+ if ((rc = LOCK_MUTEX0(rmutex)) != 0) {
+ if ((rc = mdb_mutex_failed(env, rmutex, rc)))
+ break;
+ rdrs = 0; /* the above checked all readers */
+ } else {
+ /* Recheck, a new process may have reused pid */
+ if (mdb_reader_pid(env, Pidcheck, pid))
+ j = rdrs;
+ }
+ }
+ for (; j<rdrs; j++)
if (mr[j].mr_pid == pid) {
DPRINTF(("clear stale reader pid %u txn %"Z"d",
(unsigned) pid, mr[j].mr_txnid));
mr[j].mr_pid = 0;
count++;
}
- }
- UNLOCK_MUTEX_R(env);
+ if (rmutex)
+ UNLOCK_MUTEX(rmutex);
}
}
}
@@ -9695,6 +9748,55 @@ mdb_reader_check(MDB_env *env, int *dead)
free(pids);
if (dead)
*dead = count;
- return MDB_SUCCESS;
+ return rc;
+}
+
+#ifdef MDB_ROBUST_SUPPORTED
+/** Handle #LOCK_MUTEX0() failure.
+ * Try to repair the lock file if the mutex owner died.
+ * @param[in] env the environment handle
+ * @param[in] mutex LOCK_MUTEX0() mutex
+ * @param[in] rc LOCK_MUTEX0() error (nonzero)
+ * @return 0 on success with the mutex locked, or an error code on failure.
+ */
+static int mdb_mutex_failed(MDB_env *env, mdb_mutexref_t mutex, int rc)
+{
+ int toggle, rlocked, rc2;
+
+ if (rc == MDB_OWNERDEAD) {
+ /* We own the mutex. Clean up after dead previous owner. */
+ rc = MDB_SUCCESS;
+ rlocked = (mutex == env->me_rmutex);
+ if (!rlocked) {
+ /* Keep mti_txnid updated, otherwise next writer can
+ * overwrite data which latest meta page refers to.
+ */
+ toggle = mdb_env_pick_meta(env);
+ env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
+ /* env is hosed if the dead thread was ours */
+ if (env->me_txn) {
+ env->me_flags |= MDB_FATAL_ERROR;
+ env->me_txn = NULL;
+ rc = MDB_PANIC;
+ }
+ }
+ DPRINTF(("%cmutex owner died, %s", (rlocked ? 'r' : 'w'),
+ (rc ? "this process' env is hosed" : "recovering")));
+ rc2 = mdb_reader_check0(env, rlocked, NULL);
+ if (rc2 == 0)
+ rc2 = mdb_mutex_consistent(mutex);
+ if (rc || (rc = rc2)) {
+ DPRINTF(("LOCK_MUTEX recovery failed, %s", mdb_strerror(rc)));
+ UNLOCK_MUTEX(mutex);
+ }
+ } else {
+#ifdef _WIN32
+ rc = ErrCode();
+#endif
+ DPRINTF(("LOCK_MUTEX failed, %s", mdb_strerror(rc)));
+ }
+
+ return rc;
}
+#endif /* MDB_ROBUST_SUPPORTED */
/** @} */
--
GitLab