mdbx: Merge branch 'tsan-cleanup' into devel.

Some work as a part of https://github.com/ReOpen/ReOpenLDAP/issues/62
This commit is contained in:
Leo Yuriev 2015-12-13 19:41:19 +03:00
commit 297d9e71ac
3 changed files with 125 additions and 50 deletions

View File

@ -30,7 +30,7 @@ IPROGS := mdbx_stat mdbx_copy mdbx_dump mdbx_load mdbx_chk
IDOCS := mdb_stat.1 mdb_copy.1 mdb_dump.1 mdb_load.1
PROGS := $(IPROGS) mtest0 mtest1 mtest2 mtest3 mtest4 mtest5 mtest6 wbench
SRC_LMDB := mdb.c midl.c lmdb.h midl.h
SRC_LMDB := mdb.c midl.c lmdb.h midl.h reopen.h barriers.h
SRC_MDBX := $(SRC_LMDB) mdbx.h
.PHONY: mdbx lmdb all install clean check tests coverage

132
mdb.c
View File

@ -1092,6 +1092,10 @@ static int mdb_reader_check0(MDB_env *env, int rlocked, int *dead);
static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int_ai, mdb_cmp_int_a2, mdb_cmp_int_ua;
/** @endcond */
#ifdef __SANITIZE_THREAD__
static pthread_mutex_t tsan_mutex = PTHREAD_MUTEX_INITIALIZER;
#endif
/** Return the library version info. */
char * __cold
mdb_version(int *major, int *minor, int *patch)
@ -1857,17 +1861,22 @@ static MDB_meta* mdb_meta_head_w(MDB_env *env) {
return a;
}
static MDB_meta* mdb_meta_head_r(MDB_env *env) {
static ATTRIBUTE_NO_SANITIZE_THREAD /* LY: avoid tsan-trap by meta->mm_txnid */
MDB_meta* mdb_meta_head_r(MDB_env *env) {
MDB_meta* a = METAPAGE_1(env);
MDB_meta* b = METAPAGE_2(env), *h;
txnid_t head_txnid;
int loop = 0, rc;
do {
while(1) {
#ifdef __SANITIZE_THREAD__
pthread_mutex_lock(&tsan_mutex);
pthread_mutex_unlock(&tsan_mutex);
#endif
head_txnid = env->me_txns->mti_txnid;
mdb_assert(env, a->mm_txnid != b->mm_txnid || head_txnid == 0);
if (a->mm_txnid == head_txnid)
if (likely(a->mm_txnid == head_txnid))
return a;
if (likely(b->mm_txnid == head_txnid))
return b;
@ -1877,9 +1886,13 @@ static MDB_meta* mdb_meta_head_r(MDB_env *env) {
__asm__ __volatile__("pause");
#endif
mdb_coherent_barrier();
if (loop > 2)
pthread_yield();
} while (++loop < 5);
loop += 1;
if(likely(loop < 3))
continue;
if(unlikely(loop > 5))
break;
pthread_yield();
}
rc = mdb_mutex_lock(env, MDB_MUTEX(env, w));
h = mdb_meta_head_w(env);
@ -1898,8 +1911,8 @@ static int mdb_meta_lt(MDB_meta* a, MDB_meta* b) {
}
/** Find oldest txnid still referenced. */
static txnid_t
mdb_find_oldest(MDB_env *env, int *laggard)
static ATTRIBUTE_NO_SANITIZE_THREAD /* LY: avoid tsan-trap by reader[].mr_txnid */
txnid_t mdb_find_oldest(MDB_env *env, int *laggard)
{
int i, reader;
MDB_reader *r = env->me_txns->mti_readers;
@ -1968,7 +1981,7 @@ mdb_oomkick(MDB_env *env, txnid_t oldest)
break;
if (rc) {
r->mr_txnid = (txnid_t)-1L;
r->mr_txnid = ~(txnid_t)0;
if (rc > 1) {
r->mr_tid = 0;
r->mr_pid = 0;
@ -2695,16 +2708,39 @@ mdb_reader_pid(MDB_env *env, int op, pid_t pid)
}
}
static ATTRIBUTE_NO_SANITIZE_THREAD
txnid_t lead_txnid__tsan_workaround(MDB_txn *txn, MDB_reader *r)
{
while(1) { /* LY: Retry on a race, ITS#7970. */
MDB_meta *meta = mdb_meta_head_r(txn->mt_env);
txnid_t lead = meta->mm_txnid;
r->mr_txnid = lead;
mdb_coherent_barrier();
/* Copy the DB info and flags */
memcpy(txn->mt_dbs, meta->mm_dbs, CORE_DBS * sizeof(MDB_db));
txn->mt_next_pgno = meta->mm_last_pg+1;
if (likely(lead == txn->mt_env->me_txns->mti_txnid)) {
#ifdef __SANITIZE_THREAD__
pthread_mutex_lock(&tsan_mutex);
pthread_mutex_unlock(&tsan_mutex);
#endif
return lead;
}
#if defined(__i386__) || defined(__x86_64__)
__asm__ __volatile__("pause");
#endif
}
}
/** Common code for #mdb_txn_begin() and #mdb_txn_renew().
* @param[in] txn the transaction handle to initialize
* @return 0 on success, non-zero on failure.
*/
static int
mdb_txn_renew0(MDB_txn *txn)
mdb_txn_renew0(MDB_txn *txn, unsigned flags)
{
MDB_env *env = txn->mt_env;
MDB_meta *meta;
unsigned i, nr, flags = txn->mt_flags;
unsigned i, nr;
uint16_t x;
int rc, new_notls = 0;
@ -2713,9 +2749,11 @@ mdb_txn_renew0(MDB_txn *txn)
return MDB_PANIC;
}
if ((flags &= MDB_TXN_RDONLY) != 0) {
if (flags & MDB_TXN_RDONLY) {
struct MDB_rthc *rthc = NULL;
MDB_reader *r = NULL;
txn->mt_flags = MDB_TXN_RDONLY;
if (likely(env->me_flags & MDB_ENV_TXKEY)) {
mdb_assert(env, !(env->me_flags & MDB_NOTLS));
rthc = pthread_getspecific(env->me_txkey);
@ -2741,23 +2779,26 @@ mdb_txn_renew0(MDB_txn *txn)
}
if (likely(r)) {
if (unlikely(r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1))
if (unlikely(r->mr_pid != env->me_pid || r->mr_txnid != ~(txnid_t)0))
return MDB_BAD_RSLOT;
} else {
pid_t pid = env->me_pid;
pthread_t tid = pthread_self();
pthread_mutex_t *rmutex = MDB_MUTEX(env, r);
if (unlikely(!env->me_live_reader)) {
rc = mdb_reader_pid(env, F_SETLK, pid);
if (unlikely(rc != MDB_SUCCESS))
return rc;
env->me_live_reader = 1;
}
rc = mdb_mutex_lock(env, rmutex);
if (unlikely(rc != MDB_SUCCESS))
return rc;
if (unlikely(!env->me_live_reader)) {
rc = mdb_reader_pid(env, F_SETLK, pid);
if (unlikely(rc != MDB_SUCCESS)) {
mdb_mutex_unlock(env, rmutex);
return rc;
}
env->me_live_reader = 1;
}
nr = env->me_txns->mti_numreaders;
for (i=0; i<nr; i++)
if (env->me_txns->mti_readers[i].mr_pid == 0)
@ -2774,7 +2815,7 @@ mdb_txn_renew0(MDB_txn *txn)
* When it will be closed, we can finally claim it.
*/
r->mr_pid = 0;
r->mr_txnid = (txnid_t)-1;
r->mr_txnid = ~(txnid_t)0;
r->mr_tid = tid;
mdb_coherent_barrier();
if (i == nr)
@ -2792,16 +2833,7 @@ mdb_txn_renew0(MDB_txn *txn)
}
}
do { /* LY: Retry on a race, ITS#7970. */
meta = mdb_meta_head_r(env);
r->mr_txnid = meta->mm_txnid;
mdb_coherent_barrier();
/* Copy the DB info and flags */
memcpy(txn->mt_dbs, meta->mm_dbs, CORE_DBS * sizeof(MDB_db));
txn->mt_next_pgno = meta->mm_last_pg+1;
} while(unlikely(r->mr_txnid != env->me_txns->mti_txnid));
txn->mt_txnid = r->mr_txnid;
txn->mt_txnid = lead_txnid__tsan_workaround(txn, r);
txn->mt_u.reader = r;
txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
} else {
@ -2810,10 +2842,14 @@ mdb_txn_renew0(MDB_txn *txn)
if (unlikely(rc))
return rc;
meta = mdb_meta_head_w(env);
txn->mt_txnid = meta->mm_txnid;
#ifdef __SANITIZE_THREAD__
pthread_mutex_lock(&tsan_mutex);
pthread_mutex_unlock(&tsan_mutex);
#endif
MDB_meta *meta = mdb_meta_head_w(env);
txn->mt_txnid = meta->mm_txnid + 1;
txn->mt_flags = flags;
txn->mt_txnid++;
#if MDB_DEBUG
if (unlikely(txn->mt_txnid == mdb_debug_edge)) {
if (! mdb_debug_logger)
@ -2842,8 +2878,6 @@ mdb_txn_renew0(MDB_txn *txn)
txn->mt_next_pgno = meta->mm_last_pg+1;
}
txn->mt_flags = flags;
/* Setup db info */
txn->mt_numdbs = env->me_numdbs;
for (i=CORE_DBS; i<txn->mt_numdbs; i++) {
@ -2880,7 +2914,7 @@ mdb_txn_renew(MDB_txn *txn)
if (unlikely(!F_ISSET(txn->mt_flags, MDB_TXN_RDONLY|MDB_TXN_FINISHED)))
return EINVAL;
rc = mdb_txn_renew0(txn);
rc = mdb_txn_renew0(txn, MDB_TXN_RDONLY);
if (rc == MDB_SUCCESS) {
mdb_debug("renew txn %zu%c %p on mdbenv %p, root page %zu",
txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
@ -2988,13 +3022,12 @@ mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned flags, MDB_txn **ret)
} else { /* MDB_RDONLY */
txn->mt_dbiseqs = env->me_dbiseqs;
renew:
rc = mdb_txn_renew0(txn);
rc = mdb_txn_renew0(txn, flags);
}
if (unlikely(rc)) {
if (txn != env->me_txn0)
free(txn);
} else {
txn->mt_flags |= flags; /* could not change txn=me_txn0 earlier */
txn->mt_signature = MDBX_MT_SIGNATURE;
*ret = txn;
mdb_debug("begin txn %zu%c %p on mdbenv %p, root page %zu",
@ -3050,8 +3083,8 @@ mdb_dbis_update(MDB_txn *txn, int keep)
env->me_numdbs = n;
}
int
mdbx_txn_straggler(MDB_txn *txn, int *percent)
ATTRIBUTE_NO_SANITIZE_THREAD /* LY: avoid tsan-trap by me_txn, mm_last_pg and mt_next_pgno */
int mdbx_txn_straggler(MDB_txn *txn, int *percent)
{
MDB_env *env;
MDB_meta *meta;
@ -3069,9 +3102,11 @@ mdbx_txn_straggler(MDB_txn *txn, int *percent)
env = txn->mt_env;
meta = mdb_meta_head_r(env);
if (percent) {
long cent = env->me_maxpg / 100;
long last = env->me_txn ? env->me_txn0->mt_next_pgno : meta->mm_last_pg;
*percent = (last + cent / 2) / (cent ? cent : 1);
size_t maxpg = env->me_maxpg;
size_t last = meta->mm_last_pg + 1;
if (env->me_txn)
last = env->me_txn0->mt_next_pgno;
*percent = (last + maxpg / 2) * 100u / maxpg;
}
lag = meta->mm_txnid - txn->mt_u.reader->mr_txnid;
return (0 > (long) lag) ? ~0u >> 1: lag;
@ -3103,7 +3138,7 @@ mdb_txn_end(MDB_txn *txn, unsigned mode)
if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
if (txn->mt_u.reader) {
txn->mt_u.reader->mr_txnid = (txnid_t)-1;
txn->mt_u.reader->mr_txnid = ~(txnid_t)0;
if (!(env->me_flags & MDB_NOTLS)) {
txn->mt_u.reader = NULL; /* txn does not own reader */
} else if (mode & MDB_END_SLOT) {
@ -4966,7 +5001,6 @@ mdb_env_close0(MDB_env *env)
mdb_ensure(env, rthc->rc_reader == reader);
rthc->rc_reader = NULL;
reader->mr_rthc = NULL;
free(rthc);
}
reader->mr_pid = 0;
}
@ -9307,7 +9341,7 @@ mdb_env_copyfd0(MDB_env *env, HANDLE fd)
if (unlikely(rc))
goto leave;
rc = mdb_txn_renew0(txn);
rc = mdb_txn_renew0(txn, MDB_RDONLY);
if (rc) {
mdb_mutex_unlock(env, wmutex);
goto leave;
@ -10061,7 +10095,7 @@ mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
for (i=0; i<rdrs; i++) {
if (mr[i].mr_pid) {
txnid_t txnid = mr[i].mr_txnid;
if (txnid == (txnid_t)-1l)
if (txnid == ~(txnid_t)0)
sprintf(buf, "%10d %zx -\n",
(int) mr[i].mr_pid, (size_t) mr[i].mr_tid);
else

View File

@ -67,6 +67,14 @@
# endif
#endif /* __forceinline */
#ifndef __noinline
# if defined(__GNUC__) || defined(__clang__)
# define __noinline __attribute__((noinline))
# elif defined(_MSC_VER)
# define __noinline __declspec(noinline)
# endif
#endif /* __noinline */
#ifndef __must_check_result
# if defined(__GNUC__) || defined(__clang__)
# define __must_check_result __attribute__((warn_unused_result))
@ -115,6 +123,8 @@
#ifndef __noreturn
# if defined(__GNUC__) || defined(__clang__)
# define __noreturn __attribute__((noreturn))
# elif defined(__MSC_VER)
# define __noreturn __declspec(noreturn)
# else
# define __noreturn
# endif
@ -123,6 +133,8 @@
#ifndef __nothrow
# if defined(__GNUC__) || defined(__clang__)
# define __nothrow __attribute__((nothrow))
# elif defined(__MSC_VER)
# define __nothrow __declspec(nothrow)
# else
# define __nothrow
# endif
@ -217,4 +229,33 @@ __extern_C void __assert_fail(
# define VALGRIND_CHECK_MEM_IS_DEFINED(a,s) (0)
#endif /* ! USE_VALGRIND */
#if defined(__has_feature)
# if __has_feature(thread_sanitizer)
# define __SANITIZE_THREAD__ 1
# endif
#endif
#ifdef __SANITIZE_THREAD__
# define ATTRIBUTE_NO_SANITIZE_THREAD __attribute__((no_sanitize_thread, noinline))
#else
# define ATTRIBUTE_NO_SANITIZE_THREAD
#endif
#if defined(__has_feature)
# if __has_feature(address_sanitizer)
# define __SANITIZE_ADDRESS__ 1
# endif
#endif
#ifdef __SANITIZE_ADDRESS__
# include <sanitizer/asan_interface.h>
# define ATTRIBUTE_NO_SANITIZE_ADDRESS __attribute__((no_sanitize_address, noinline))
#else
# define ASAN_POISON_MEMORY_REGION(addr, size) \
((void)(addr), (void)(size))
# define ASAN_UNPOISON_MEMORY_REGION(addr, size) \
((void)(addr), (void)(size))
# define ATTRIBUTE_NO_SANITIZE_ADDRESS
#endif /* __SANITIZE_ADDRESS__ */
#endif /* _REOPEN_H */