aboutsummaryrefslogtreecommitdiffstats
path: root/erts/emulator/sys/common
diff options
context:
space:
mode:
authorLukas Larsson <lukas@erlang.org>2017-05-30 16:35:18 +0200
committerLukas Larsson <lukas@erlang.org>2017-10-02 10:35:52 +0200
commit988f5f5e8061ce2e135a314ca782788eda478a06 (patch)
tree1609c42abb9ed03a176865ee32ad3de803c392d3 /erts/emulator/sys/common
parent1f9003a3dd9bbf9e3dcd84b8fdecef04a3b5e9c7 (diff)
downloadotp-988f5f5e8061ce2e135a314ca782788eda478a06.tar.gz
otp-988f5f5e8061ce2e135a314ca782788eda478a06.tar.bz2
otp-988f5f5e8061ce2e135a314ca782788eda478a06.zip
erts: Move all I/O polling to a seperate thread
Diffstat (limited to 'erts/emulator/sys/common')
-rw-r--r--erts/emulator/sys/common/erl_check_io.c2304
-rw-r--r--erts/emulator/sys/common/erl_check_io.h92
-rw-r--r--erts/emulator/sys/common/erl_poll.c2640
-rw-r--r--erts/emulator/sys/common/erl_poll.h281
-rw-r--r--erts/emulator/sys/common/erl_poll_api.h122
5 files changed, 2330 insertions, 3109 deletions
diff --git a/erts/emulator/sys/common/erl_check_io.c b/erts/emulator/sys/common/erl_check_io.c
index ff0f3ea121..3131b96536 100644
--- a/erts/emulator/sys/common/erl_check_io.c
+++ b/erts/emulator/sys/common/erl_check_io.c
@@ -29,7 +29,6 @@
#endif
#define ERL_CHECK_IO_C__
-#define ERTS_WANT_BREAK_HANDLING
#ifndef WANT_NONBLOCKING
# define WANT_NONBLOCKING
#endif
@@ -44,151 +43,108 @@
#define ERTS_WANT_TIMER_WHEEL_API
#include "erl_time.h"
+#if 0
+#define DEBUG_PRINT(FMT, ...) erts_printf(FMT "\r\n", ##__VA_ARGS__)
+#define DEBUG_PRINT_FD(FMT, STATE, ...) \
+ DEBUG_PRINT("%d: " FMT " (ev=%s, ac=%s, flg=%d)", \
+ (STATE) ? (STATE)->fd : (ErtsSysFdType)-1, ##__VA_ARGS__, \
+ ev2str((STATE) ? (STATE)->events : ERTS_POLL_EV_NONE), \
+ ev2str((STATE) ? (STATE)->active_events : ERTS_POLL_EV_NONE), \
+ (STATE) ? (STATE)->flags : ERTS_EV_FLAG_CLEAR)
+#define DEBUG_PRINT_MODE
+#else
+#define DEBUG_PRINT(...)
+#endif
+
+#ifndef DEBUG_PRINT_FD
+#define DEBUG_PRINT_FD(...)
+#endif
+
#ifndef ERTS_SYS_CONTINOUS_FD_NUMBERS
# include "safe_hash.h"
# define DRV_EV_STATE_HTAB_SIZE 1024
#endif
-typedef char EventStateType;
-#define ERTS_EV_TYPE_NONE ((EventStateType) 0)
-#define ERTS_EV_TYPE_DRV_SEL ((EventStateType) 1) /* driver_select */
-#define ERTS_EV_TYPE_STOP_USE ((EventStateType) 2) /* pending stop_select */
-#define ERTS_EV_TYPE_NIF ((EventStateType) 3) /* enif_select */
-#define ERTS_EV_TYPE_STOP_NIF ((EventStateType) 4) /* pending nif stop */
-
-typedef char EventStateFlags;
-#define ERTS_EV_FLAG_USED ((EventStateFlags) 1) /* ERL_DRV_USE has been turned on */
-#define ERTS_EV_FLAG_DEFER_IN_EV ((EventStateFlags) 2)
-#define ERTS_EV_FLAG_DEFER_OUT_EV ((EventStateFlags) 4)
+typedef enum {
+ ERTS_EV_TYPE_NONE = 0,
+ ERTS_EV_TYPE_DRV_SEL = 1, /* driver_select */
+ ERTS_EV_TYPE_STOP_USE = 2, /* pending stop_select */
+ ERTS_EV_TYPE_NIF = 3, /* enif_select */
+ ERTS_EV_TYPE_STOP_NIF = 4 /* pending nif stop */
+} EventStateType;
-#ifdef DEBUG
-# define ERTS_ACTIVE_FD_INC 2
+typedef enum {
+ ERTS_EV_FLAG_CLEAR = 0,
+ ERTS_EV_FLAG_USED = 1, /* ERL_DRV_USE has been turned on */
+#ifdef ERTS_ENABLE_KERNEL_POLL
+ ERTS_EV_FLAG_FALLBACK = 2, /* Set when kernel poll rejected fd
+ and it was put in the nkp version */
#else
-# define ERTS_ACTIVE_FD_INC 128
+ ERTS_EV_FLAG_FALLBACK = ERTS_EV_FLAG_CLEAR,
#endif
-#define ERTS_CHECK_IO_POLL_RES_LEN 512
-
-#if defined(ERTS_KERNEL_POLL_VERSION)
-# define ERTS_CIO_EXPORT(FUNC) FUNC ## _kp
-#elif defined(ERTS_NO_KERNEL_POLL_VERSION)
-# define ERTS_CIO_EXPORT(FUNC) FUNC ## _nkp
-#else
-# define ERTS_CIO_EXPORT(FUNC) FUNC
-#endif
+ /* Combinations */
+ ERTS_EV_FLAG_USED_FALLBACK = ERTS_EV_FLAG_USED | ERTS_EV_FLAG_FALLBACK
+} EventStateFlags;
-#define ERTS_CIO_POLL_CTL ERTS_POLL_EXPORT(erts_poll_control)
-#define ERTS_CIO_POLL_CTLV ERTS_POLL_EXPORT(erts_poll_controlv)
-#define ERTS_CIO_POLL_WAIT ERTS_POLL_EXPORT(erts_poll_wait)
-#define ERTS_CIO_POLL_INTR ERTS_POLL_EXPORT(erts_poll_interrupt)
-#define ERTS_CIO_POLL_INTR_TMD ERTS_POLL_EXPORT(erts_poll_interrupt_timed)
-#define ERTS_CIO_NEW_POLLSET ERTS_POLL_EXPORT(erts_poll_create_pollset)
-#define ERTS_CIO_FREE_POLLSET ERTS_POLL_EXPORT(erts_poll_destroy_pollset)
-#define ERTS_CIO_POLL_MAX_FDS ERTS_POLL_EXPORT(erts_poll_max_fds)
-#define ERTS_CIO_POLL_INIT ERTS_POLL_EXPORT(erts_poll_init)
-#define ERTS_CIO_POLL_INFO ERTS_POLL_EXPORT(erts_poll_info)
+#define flag2str(flags) \
+ ((flags) == ERTS_EV_FLAG_CLEAR ? "CLEAR" : \
+ ((flags) == ERTS_EV_FLAG_USED ? "USED" : \
+ ((flags) == ERTS_EV_FLAG_FALLBACK ? "FLBK" : \
+ ((flags) == ERTS_EV_FLAG_USED_FALLBACK ? "USED|FLBK" : "ERROR"))))
-#define GET_FD(fd) fd
+/* How many events that can be handled at once by one erts_poll_wait call */
+#define ERTS_CHECK_IO_POLL_RES_LEN 512
-static struct pollset_info
+/* Each I/O Poll Thread has one ErtsPollThread each. The ps field
+ can point to either a private ErtsPollSet or a shared one.
+ At the moment only kqueue and epoll pollsets can be
+ shared across threads.
+*/
+typedef struct erts_poll_thread
{
- ErtsPollSet ps;
- erts_atomic_t in_poll_wait; /* set while doing poll */
- erts_atomic_t check_io_time;
- struct {
- int six; /* start index */
- int eix; /* end index */
- erts_atomic32_t no;
- int size;
- ErtsSysFdType *array;
- } active_fd;
- erts_atomic_t removed_list; /* struct removed_fd* */
-}*pollsetv;
-
-#define NUM_OF_POLLSETS 1
+ ErtsPollSet *ps;
+ ErtsPollResFd *pollres;
+ int pollres_len;
+} ErtsPollThread;
-#ifdef ERTS_ENABLE_KERNEL_POLL
-void erts_init_check_io_kp(void);
-void erts_init_check_io_nkp(void);
-int ERTS_CIO_EXPORT(driver_select)(ErlDrvPort, ErlDrvEvent, int, int);
-int ERTS_CIO_EXPORT(enif_select)(ErlNifEnv*, ErlNifEvent, enum ErlNifSelectFlags, void*, const ErlNifPid*, Eterm);
-Uint ERTS_CIO_EXPORT(erts_check_io_size)(void);
-Eterm ERTS_CIO_EXPORT(erts_check_io_info)(void *);
-int ERTS_CIO_EXPORT(erts_check_io_max_files)(void);
-void ERTS_CIO_EXPORT(erts_check_io_interrupt)(struct pollset_info*, int);
-void ERTS_CIO_EXPORT(erts_check_io_interrupt_timed)(struct pollset_info*, int, ErtsMonotonicTime);
-void ERTS_CIO_EXPORT(erts_check_io)(int);
-struct pollset_info* ERTS_CIO_EXPORT(erts_get_pollset)(int);
-int ERTS_CIO_EXPORT(erts_check_io_debug)(ErtsCheckIoDebugInfo *);
-void ERTS_CIO_EXPORT(erts_io_notify_port_task_executed)(ErtsPortTaskHandle *pthp);
-#ifdef ERTS_ENABLE_LOCK_COUNT
-void ERTS_CIO_EXPORT(erts_lcnt_update_cio_locks)(int enable);
-#endif
+/* pollsetv contains pointers to the ErtsPollSets that are in use.
+ * Which pollset to use is determined by hashing the fd.
+ */
+static ErtsPollSet **pollsetv;
+#if ERTS_POLL_USE_FALLBACK
+static ErtsPollSet *flbk_pollset;
#endif
-
-/* ToDo: Was inline in erl_check_io.h but now need struct pollset_info */
-void
-ERTS_CIO_EXPORT(erts_io_notify_port_task_executed)(ErtsPortTaskHandle *pthp)
-{
- ErtsIoTask *itp = ErtsContainerStruct(pthp, ErtsIoTask, task);
- erts_aint_t ci_time = erts_atomic_read_acqb(&itp->pollset->check_io_time);
- erts_atomic_set_relb(&itp->executed_time, ci_time);
-}
-
-
-struct pollset_info* ERTS_CIO_EXPORT(erts_get_pollset)(int sched_nr)
-{
- ASSERT(sched_nr > 0 && sched_nr <= erts_no_schedulers);
- return &pollsetv[sched_nr - 1];
-}
+static ErtsPollThread *psiv;
typedef struct {
#ifndef ERTS_SYS_CONTINOUS_FD_NUMBERS
SafeHashBucket hb;
#endif
ErtsSysFdType fd;
- struct pollset_info *pollset;
struct {
ErtsDrvSelectDataState *select; /* ERTS_EV_TYPE_DRV_SEL */
ErtsNifSelectDataState *nif; /* ERTS_EV_TYPE_NIF */
union {
erts_driver_t* drv_ptr; /* ERTS_EV_TYPE_STOP_USE */
ErtsResource* resource; /* ERTS_EV_TYPE_STOP_NIF */
- }stop;
+ } stop;
} driver;
- ErtsPollEvents events;
- unsigned short remove_cnt; /* number of removed_fd's referring to this fd */
+ ErtsPollEvents events; /* The events that have been selected upon */
+ ErtsPollEvents active_events; /* The events currently active in the pollset */
EventStateType type;
EventStateFlags flags;
} ErtsDrvEventState;
-struct removed_fd {
- struct removed_fd *next;
-#ifdef ERTS_SYS_CONTINOUS_FD_NUMBERS
- ErtsSysFdType fd;
-#else
- ErtsDrvEventState* state;
- #ifdef DEBUG
- ErtsSysFdType fd;
- #endif
-#endif
-
-};
-
struct drv_ev_state_shared {
- /* The layout of this struct must be independent of kp/nkp compilation */
-#ifdef ERTS_SYS_CONTINOUS_FD_NUMBERS
- int max_fds;
-#endif
-
-#define DRV_EV_STATE_LOCK_CNT 128
union {
erts_mtx_t lck;
byte _cache_line_alignment[ERTS_ALC_CACHE_LINE_ALIGN_SIZE(sizeof(erts_mtx_t))];
- }locks[DRV_EV_STATE_LOCK_CNT];
+ } locks[ERTS_CHECK_IO_DRV_EV_STATE_LOCK_CNT];
#ifdef ERTS_SYS_CONTINOUS_FD_NUMBERS
+ int max_fds;
erts_atomic_t len;
ErtsDrvEventState *v;
erts_mtx_t grow_lock; /* prevent lock-hogging of racing growers */
@@ -200,51 +156,121 @@ struct drv_ev_state_shared {
#endif
};
-#ifndef ERTS_KERNEL_POLL_VERSION
+int ERTS_WRITE_UNLIKELY(erts_no_pollsets) = 1;
+int ERTS_WRITE_UNLIKELY(erts_no_poll_threads) = 1;
struct drv_ev_state_shared drv_ev_state;
-#else
-extern struct drv_ev_state_shared drv_ev_state;
-#endif
-static ERTS_INLINE erts_mtx_t* fd_mtx(ErtsSysFdType fd)
-{
+static ERTS_INLINE int fd_hash(ErtsSysFdType fd) {
int hash = (int)fd;
# ifndef ERTS_SYS_CONTINOUS_FD_NUMBERS
hash ^= (hash >> 9);
# endif
- return &drv_ev_state.locks[hash % DRV_EV_STATE_LOCK_CNT].lck;
+ return hash;
}
-#ifndef ERTS_SYS_CONTINOUS_FD_NUMBERS
+static ERTS_INLINE erts_mtx_t* fd_mtx(ErtsSysFdType fd)
+{
+ return &drv_ev_state.locks[fd_hash(fd) % ERTS_CHECK_IO_DRV_EV_STATE_LOCK_CNT].lck;
+}
+
+#ifdef ERTS_SYS_CONTINOUS_FD_NUMBERS
+
+static ERTS_INLINE ErtsDrvEventState *get_drv_ev_state(ErtsSysFdType fd)
+{
+ return &drv_ev_state.v[(int) fd];
+}
+
+#define new_drv_ev_state(State, fd) (State)
+#define erase_drv_ev_state(State)
+
+static ERTS_INLINE int grow_drv_ev_state(ErtsSysFdType fd) {
+ int i;
+ int old_len;
+ int new_len;
+
+ if ((unsigned)fd >= (unsigned)erts_atomic_read_nob(&drv_ev_state.len)) {
+
+ if (fd < 0 || fd >= drv_ev_state.max_fds)
+ return 0;
-static ERTS_INLINE ErtsDrvEventState *hash_get_drv_ev_state(ErtsSysFdType fd)
+ erts_mtx_lock(&drv_ev_state.grow_lock);
+ old_len = erts_atomic_read_nob(&drv_ev_state.len);
+ if (fd >= old_len) {
+ new_len = erts_poll_new_table_len(old_len, fd + 1);
+ if (new_len > drv_ev_state.max_fds)
+ new_len = drv_ev_state.max_fds;
+
+ for (i=0; i<ERTS_CHECK_IO_DRV_EV_STATE_LOCK_CNT; i++) { /* lock all fd's */
+ erts_mtx_lock(&drv_ev_state.locks[i].lck);
+ }
+ drv_ev_state.v = (drv_ev_state.v
+ ? erts_realloc(ERTS_ALC_T_DRV_EV_STATE,
+ drv_ev_state.v,
+ sizeof(ErtsDrvEventState)*new_len)
+ : erts_alloc(ERTS_ALC_T_DRV_EV_STATE,
+ sizeof(ErtsDrvEventState)*new_len));
+ ERTS_CT_ASSERT(ERTS_EV_TYPE_NONE == 0);
+ sys_memzero(drv_ev_state.v+old_len,
+ sizeof(ErtsDrvEventState) * (new_len - old_len));
+ for (i = old_len; i < new_len; i++) {
+ drv_ev_state.v[i].fd = (ErtsSysFdType) i;
+ }
+ erts_atomic_set_nob(&drv_ev_state.len, new_len);
+ for (i=0; i<ERTS_CHECK_IO_DRV_EV_STATE_LOCK_CNT; i++) {
+ erts_mtx_unlock(&drv_ev_state.locks[i].lck);
+ }
+ }
+ /*else already grown by racing thread */
+
+ erts_mtx_unlock(&drv_ev_state.grow_lock);
+ }
+ return 1;
+}
+
+static int drv_ev_state_len(void)
+{
+ return erts_atomic_read_nob(&drv_ev_state.len);
+}
+
+#else /* !ERTS_SYS_CONTINOUS_FD_NUMBERS */
+
+static ERTS_INLINE ErtsDrvEventState *get_drv_ev_state(ErtsSysFdType fd)
{
ErtsDrvEventState tmpl;
tmpl.fd = fd;
return (ErtsDrvEventState *) safe_hash_get(&drv_ev_state.tab, (void *) &tmpl);
}
-static ERTS_INLINE ErtsDrvEventState* hash_new_drv_ev_state(ErtsSysFdType fd)
+static ERTS_INLINE ErtsDrvEventState* new_drv_ev_state(ErtsDrvEventState *state,
+ ErtsSysFdType fd)
{
ErtsDrvEventState tmpl;
+
+ if (state)
+ return state;
+
tmpl.fd = fd;
- tmpl.pollset = NULL;
tmpl.driver.select = NULL;
tmpl.driver.nif = NULL;
tmpl.driver.stop.drv_ptr = NULL;
tmpl.events = 0;
- tmpl.remove_cnt = 0;
+ tmpl.active_events = 0;
tmpl.type = ERTS_EV_TYPE_NONE;
tmpl.flags = 0;
+
return (ErtsDrvEventState *) safe_hash_put(&drv_ev_state.tab, (void *) &tmpl);
}
-static ERTS_INLINE void hash_erase_drv_ev_state(ErtsDrvEventState *state)
+static ERTS_INLINE void erase_drv_ev_state(ErtsDrvEventState *state)
{
- ASSERT(state->remove_cnt == 0);
safe_hash_erase(&drv_ev_state.tab, (void *) state);
}
+static int drv_ev_state_len(void)
+{
+ return erts_atomic_read_nob(&drv_ev_state.tab.nitems);
+}
+
#endif /* !ERTS_SYS_CONTINOUS_FD_NUMBERS */
static void stale_drv_select(Eterm id, ErtsDrvEventState *state, int mode);
@@ -268,36 +294,39 @@ steal_pending_stop_use(erts_dsprintf_buf_t*, ErlDrvPort, ErtsDrvEventState*,
static void
steal_pending_stop_nif(erts_dsprintf_buf_t *dsbufp, ErtsResource*,
ErtsDrvEventState *state, int mode, int on);
-
-ERTS_SCHED_PREF_QUICK_ALLOC_IMPL(removed_fd, struct removed_fd, 64, ERTS_ALC_T_FD_LIST)
+static ERTS_INLINE void
+check_fd_cleanup(ErtsDrvEventState *state,
+ ErtsDrvSelectDataState **free_select,
+ ErtsNifSelectDataState **free_nif);
+#ifdef DEBUG_PRINT_MODE
+static char *drvmode2str(int mode);
+static char *nifmode2str(enum ErlNifSelectFlags mode);
+#endif
static ERTS_INLINE void
-init_iotask(ErtsIoTask *io_task, struct pollset_info* psi)
+init_iotask(ErtsIoTask *io_task, ErtsSysFdType fd)
{
erts_port_task_handle_init(&io_task->task);
- erts_atomic_init_nob(&io_task->executed_time, ~((erts_aint_t) 0));
- io_task->pollset = psi;
+ io_task->fd = fd;
}
static ERTS_INLINE int
-is_iotask_active(ErtsIoTask *io_task, erts_aint_t current_cio_time)
-{
+is_iotask_active(ErtsIoTask *io_task)
+{
if (erts_port_task_is_scheduled(&io_task->task))
return 1;
- if (erts_atomic_read_nob(&io_task->executed_time) == current_cio_time)
- return 1;
return 0;
}
static ERTS_INLINE ErtsDrvSelectDataState *
-alloc_drv_select_data(struct pollset_info* psi)
+alloc_drv_select_data(ErtsSysFdType fd)
{
ErtsDrvSelectDataState *dsp = erts_alloc(ERTS_ALC_T_DRV_SEL_D_STATE,
sizeof(ErtsDrvSelectDataState));
dsp->inport = NIL;
dsp->outport = NIL;
- init_iotask(&dsp->iniotask, psi);
- init_iotask(&dsp->outiotask, psi);
+ init_iotask(&dsp->iniotask, fd);
+ init_iotask(&dsp->outiotask, fd);
return dsp;
}
@@ -308,8 +337,6 @@ alloc_nif_select_data(void)
sizeof(ErtsNifSelectDataState));
dsp->in.pid = NIL;
dsp->out.pid = NIL;
- dsp->in.ddeselect_cnt = 0;
- dsp->out.ddeselect_cnt = 0;
return dsp;
}
@@ -327,176 +354,126 @@ free_nif_select_data(ErtsNifSelectDataState *dsp)
erts_free(ERTS_ALC_T_NIF_SEL_D_STATE, dsp);
}
-static ERTS_INLINE void
-remember_removed(ErtsDrvEventState *state)
+static ERTS_INLINE int
+get_pollset_id(ErtsSysFdType fd)
{
- struct removed_fd *fdlp;
- ERTS_LC_ASSERT(erts_lc_mtx_is_locked(fd_mtx(state->fd)));
- if (erts_atomic_read_nob(&state->pollset->in_poll_wait)) {
- erts_aint_t was_next, exp_next;
- state->remove_cnt++;
- ASSERT(state->remove_cnt > 0);
- fdlp = removed_fd_alloc();
- #if defined(ERTS_SYS_CONTINOUS_FD_NUMBERS) || defined(DEBUG)
- fdlp->fd = state->fd;
- #endif
- #ifndef ERTS_SYS_CONTINOUS_FD_NUMBERS
- fdlp->state = state;
- #endif
-
- /* Lockless atomic insertion in removed_list */
- was_next = erts_atomic_read_acqb(&state->pollset->removed_list);
- do {
- exp_next = was_next;
- fdlp->next = (struct removed_fd*) exp_next;
- was_next = erts_atomic_cmpxchg_mb(&state->pollset->removed_list,
- (erts_aint_t) fdlp,
- exp_next);
- }while (was_next != exp_next);
- }
+ return fd_hash(fd) % erts_no_pollsets;
}
+static ERTS_INLINE ErtsPollSet *
+get_pollset(ErtsSysFdType fd)
+{
+ return pollsetv[get_pollset_id(fd)];
+}
-static ERTS_INLINE int
-is_removed(ErtsDrvEventState *state)
+#if ERTS_POLL_USE_FALLBACK
+static ERTS_INLINE ErtsPollSet *
+get_fallback(void)
{
- /* Note that there is a possible race here, where an fd is removed
- (increasing remove_cnt) and then added again just before erts_poll_wait
- is called by erts_check_io. Any polled event on the re-added fd will then
- be falsely ignored. But that does not matter, as the event will trigger
- again next time erl_check_io is called. */
- return state->remove_cnt > 0;
+ return flbk_pollset;
}
+#endif
-static void
-forget_removed(struct pollset_info* psi)
+/*
+ * Place a fd within a pollset. This will automatically use
+ * the fallback ps if needed.
+ */
+static ERTS_INLINE ErtsPollEvents
+erts_io_control_wakeup(ErtsDrvEventState *state, ErtsPollOp op,
+ ErtsPollEvents pe, int *wake_poller)
{
- struct removed_fd* fdlp;
- struct removed_fd* tofree;
+ ErtsSysFdType fd = state->fd;
+ ErtsPollEvents res = 0;
+ EventStateFlags flags = state->flags;
- fdlp = (struct removed_fd*) erts_atomic_xchg_mb(&psi->removed_list,
- (erts_aint_t) NULL);
+ ERTS_LC_ASSERT(erts_lc_mtx_is_locked(fd_mtx(state->fd)));
- while (fdlp) {
- ErtsResource* resource = NULL;
- erts_driver_t* drv_ptr = NULL;
- erts_mtx_t* mtx;
- ErtsSysFdType fd;
- ErtsDrvEventState *state;
+ if (!(flags & ERTS_EV_FLAG_FALLBACK)) {
+ res = erts_poll_control(get_pollset(fd), fd, op, pe, wake_poller);
-#ifdef ERTS_SYS_CONTINOUS_FD_NUMBERS
- fd = fdlp->fd;
- mtx = fd_mtx(fd);
- erts_mtx_lock(mtx);
- state = &drv_ev_state.v[(int) fd];
-#else
- state = fdlp->state;
- fd = state->fd;
- ASSERT(fd == fdlp->fd);
- mtx = fd_mtx(fd);
- erts_mtx_lock(mtx);
-#endif
- ASSERT(state->remove_cnt > 0);
- if (--state->remove_cnt == 0) {
- switch (state->type) {
- case ERTS_EV_TYPE_STOP_NIF:
- /* Now we can call stop */
- resource = state->driver.stop.resource;
- state->driver.stop.resource = NULL;
- ASSERT(resource);
- state->type = ERTS_EV_TYPE_NONE;
- state->flags = 0;
- goto case_ERTS_EV_TYPE_NONE;
-
- case ERTS_EV_TYPE_STOP_USE:
- /* Now we can call stop_select */
- drv_ptr = state->driver.stop.drv_ptr;
- ASSERT(drv_ptr);
- state->type = ERTS_EV_TYPE_NONE;
- state->flags = 0;
- state->driver.stop.drv_ptr = NULL;
- /* Fall through */
- case ERTS_EV_TYPE_NONE:
- case_ERTS_EV_TYPE_NONE:
- state->pollset = NULL;
-#ifndef ERTS_SYS_CONTINOUS_FD_NUMBERS
- hash_erase_drv_ev_state(state);
-#endif
- break;
- case ERTS_EV_TYPE_DRV_SEL:
- break;
- default:
- ASSERT(0);
- }
- }
- erts_mtx_unlock(mtx);
- if (drv_ptr) {
- int was_unmasked = erts_block_fpe();
- DTRACE1(driver_stop_select, drv_ptr->name);
- LTTNG1(driver_stop_select, drv_ptr->name);
- (*drv_ptr->stop_select) ((ErlDrvEvent) fd, NULL);
- erts_unblock_fpe(was_unmasked);
- if (drv_ptr->handle) {
- erts_ddll_dereference_driver(drv_ptr->handle);
- }
- }
- if (resource) {
- erts_resource_stop(resource, (ErlNifEvent)fd, 0);
- enif_release_resource(resource->data);
+#if ERTS_POLL_USE_FALLBACK
+ if (op == ERTS_POLL_OP_ADD && res == ERTS_POLL_EV_NVAL) {
+ /* When an add fails with NVAL, the poll/kevent operation could not
+ put that fd in the pollset, so we instead put it into a fallback pollset */
+ state->flags |= ERTS_EV_FLAG_FALLBACK;
+ res = erts_poll_control_flbk(get_fallback(), fd, op, pe, wake_poller);
}
-
- tofree = fdlp;
- fdlp = fdlp->next;
- removed_fd_free(tofree);
+ } else {
+ ASSERT(op != ERTS_POLL_OP_ADD);
+ res = erts_poll_control_flbk(get_fallback(), fd, op, pe, wake_poller);
+#endif
}
+
+ return res;
}
-#ifdef ERTS_SYS_CONTINOUS_FD_NUMBERS
-static void
-grow_drv_ev_state(int min_ix)
+static ERTS_INLINE ErtsPollEvents
+erts_io_control(ErtsDrvEventState *state, ErtsPollOp op, ErtsPollEvents pe)
{
- int i;
- int old_len;
- int new_len;
+ int wake_poller = 0;
+ return erts_io_control_wakeup(state, op, pe, &wake_poller);
+}
- erts_mtx_lock(&drv_ev_state.grow_lock);
- old_len = erts_atomic_read_nob(&drv_ev_state.len);
- if (min_ix >= old_len) {
- new_len = erts_poll_new_table_len(old_len, min_ix + 1);
- if (new_len > drv_ev_state.max_fds)
- new_len = drv_ev_state.max_fds;
-
- for (i=0; i<DRV_EV_STATE_LOCK_CNT; i++) { /* lock all fd's */
- erts_mtx_lock(&drv_ev_state.locks[i].lck);
- }
- drv_ev_state.v = (drv_ev_state.v
- ? erts_realloc(ERTS_ALC_T_DRV_EV_STATE,
- drv_ev_state.v,
- sizeof(ErtsDrvEventState)*new_len)
- : erts_alloc(ERTS_ALC_T_DRV_EV_STATE,
- sizeof(ErtsDrvEventState)*new_len));
- for (i = old_len; i < new_len; i++) {
- drv_ev_state.v[i].fd = (ErtsSysFdType) i;
- drv_ev_state.v[i].pollset = NULL;
- drv_ev_state.v[i].driver.select = NULL;
- drv_ev_state.v[i].driver.stop.drv_ptr = NULL;
- drv_ev_state.v[i].driver.nif = NULL;
- drv_ev_state.v[i].events = 0;
- drv_ev_state.v[i].remove_cnt = 0;
- drv_ev_state.v[i].type = ERTS_EV_TYPE_NONE;
- drv_ev_state.v[i].flags = 0;
- }
- erts_atomic_set_nob(&drv_ev_state.len, new_len);
- for (i=0; i<DRV_EV_STATE_LOCK_CNT; i++) {
- erts_mtx_unlock(&drv_ev_state.locks[i].lck);
- }
+/* ToDo: Was inline in erl_check_io.h but now need struct erts_poll_thread */
+void
+erts_io_notify_port_task_executed(ErtsPortTaskType type,
+ ErtsPortTaskHandle *pthp,
+ void (*reset_handle)(ErtsPortTaskHandle *))
+{
+ ErtsIoTask *itp = ErtsContainerStruct(pthp, ErtsIoTask, task);
+ ErtsSysFdType fd = itp->fd;
+ erts_mtx_t *mtx = fd_mtx(fd);
+ int active_events;
+ ErtsDrvEventState *state;
+ ErtsDrvSelectDataState *free_select = NULL;
+ ErtsNifSelectDataState *free_nif = NULL;
+
+ erts_mtx_lock(mtx);
+ state = get_drv_ev_state(fd);
+
+ active_events = state->active_events;
+
+ switch (type) {
+ case ERTS_PORT_TASK_INPUT:
+
+ DEBUG_PRINT_FD("executed ready_input", state);
+
+ ASSERT(!(state->active_events & ERTS_POLL_EV_IN));
+ if (state->events & ERTS_POLL_EV_IN)
+ active_events |= ERTS_POLL_EV_IN;
+ break;
+ case ERTS_PORT_TASK_OUTPUT:
+
+ DEBUG_PRINT_FD("executed ready_output", state);
+
+ ASSERT(!(state->active_events & ERTS_POLL_EV_OUT));
+ if (state->events & ERTS_POLL_EV_OUT)
+ active_events |= ERTS_POLL_EV_OUT;
+ break;
+ default:
+ erts_exit(ERTS_ABORT_EXIT, "Invalid IO port task type");
+ break;
}
- /*else already grown by racing thread */
- erts_mtx_unlock(&drv_ev_state.grow_lock);
-}
-#endif /* ERTS_SYS_CONTINOUS_FD_NUMBERS */
+ reset_handle(pthp);
+
+ if (active_events) {
+ /* This is not needed if active_events has not changed */
+ if (state->active_events != active_events) {
+ state->active_events = active_events;
+ erts_io_control(state, ERTS_POLL_OP_MOD, active_events);
+ }
+ } else {
+ check_fd_cleanup(state, &free_select, &free_nif);
+ }
+ erts_mtx_unlock(mtx);
+
+ if (free_select)
+ free_drv_select_data(free_select);
+ if (free_nif)
+ free_nif_select_data(free_nif);
+}
static ERTS_INLINE void
abort_task(Eterm id, ErtsPortTaskHandle *pthp, EventStateType type)
@@ -542,16 +519,14 @@ abort_tasks(ErtsDrvEventState *state, int mode)
static void
deselect(ErtsDrvEventState *state, int mode)
{
- int do_wake = 0;
ErtsPollEvents rm_events;
ERTS_LC_ASSERT(erts_lc_mtx_is_locked(fd_mtx(state->fd)));
- ASSERT(state->events);
abort_tasks(state, mode);
- if (!mode)
+ if (!mode) {
rm_events = state->events;
- else {
+ } else {
rm_events = 0;
ASSERT(state->type == ERTS_EV_TYPE_DRV_SEL);
if (mode & ERL_DRV_READ) {
@@ -564,18 +539,16 @@ deselect(ErtsDrvEventState *state, int mode)
}
}
- ERTS_CIO_POLL_CTL(state->pollset->ps, state->fd, rm_events,
- 0, &do_wake);
state->events &= ~rm_events;
+ state->active_events &= ~rm_events;
if (!(state->events)) {
+ erts_io_control(state, ERTS_POLL_OP_DEL, 0);
switch (state->type) {
case ERTS_EV_TYPE_NIF:
state->driver.nif->in.pid = NIL;
state->driver.nif->out.pid = NIL;
- state->driver.nif->in.ddeselect_cnt = 0;
- state->driver.nif->out.ddeselect_cnt = 0;
- enif_release_resource(state->driver.stop.resource);
+ enif_release_resource(state->driver.stop.resource->data);
state->driver.stop.resource = NULL;
break;
case ERTS_EV_TYPE_DRV_SEL:
@@ -588,15 +561,15 @@ deselect(ErtsDrvEventState *state, int mode)
ASSERT(0);
break;
}
-
state->type = ERTS_EV_TYPE_NONE;
state->flags = 0;
- remember_removed(state);
+ } else {
+ erts_io_control(state, ERTS_POLL_OP_MOD, state->active_events);
}
}
#ifdef ERTS_SYS_CONTINOUS_FD_NUMBERS
-# define IS_FD_UNKNOWN(state) ((state)->type == ERTS_EV_TYPE_NONE && (state)->remove_cnt == 0)
+# define IS_FD_UNKNOWN(state) ((state)->type == ERTS_EV_TYPE_NONE)
#else
# define IS_FD_UNKNOWN(state) ((state) == NULL)
#endif
@@ -606,17 +579,13 @@ check_fd_cleanup(ErtsDrvEventState *state,
ErtsDrvSelectDataState **free_select,
ErtsNifSelectDataState **free_nif)
{
- erts_aint_t current_cio_time;
-
ERTS_LC_ASSERT(erts_lc_mtx_is_locked(fd_mtx(state->fd)));
-
- current_cio_time = erts_atomic_read_acqb(&state->pollset->check_io_time);
*free_select = NULL;
if (state->driver.select
&& (state->type != ERTS_EV_TYPE_DRV_SEL)
- && !is_iotask_active(&state->driver.select->iniotask, current_cio_time)
- && !is_iotask_active(&state->driver.select->outiotask, current_cio_time)) {
-
+ && !is_iotask_active(&state->driver.select->iniotask)
+ && !is_iotask_active(&state->driver.select->outiotask)) {
+
*free_select = state->driver.select;
state->driver.select = NULL;
}
@@ -628,14 +597,10 @@ check_fd_cleanup(ErtsDrvEventState *state,
}
if (((state->type != ERTS_EV_TYPE_NONE)
- | state->remove_cnt
| (state->driver.nif != NULL)
| (state->driver.select != NULL)) == 0) {
- state->pollset = NULL;
-#ifndef ERTS_SYS_CONTINOUS_FD_NUMBERS
- hash_erase_drv_ev_state(state);
-#endif
+ erase_drv_ev_state(state);
}
}
@@ -645,253 +610,8 @@ check_fd_cleanup(ErtsDrvEventState *state,
# define MUST_DEFER(MAY_SLEEP) (MAY_SLEEP)
#endif
-static ERTS_INLINE int
-check_cleanup_active_fd(struct pollset_info* psi,
- ErtsSysFdType fd,
-#if ERTS_CIO_DEFER_ACTIVE_EVENTS
- ErtsPollControlEntry *pce,
- int *pce_ix,
-#endif
- erts_aint_t current_cio_time,
- int may_sleep)
-{
- ErtsDrvEventState *state;
- int active = 0;
- erts_mtx_t *mtx = fd_mtx(fd);
- void *free_select = NULL;
- void *free_nif = NULL;
-#if ERTS_CIO_DEFER_ACTIVE_EVENTS
- ErtsPollEvents evon = 0, evoff = 0;
-#endif
-
- erts_mtx_lock(mtx);
-
-#ifdef ERTS_SYS_CONTINOUS_FD_NUMBERS
- state = &drv_ev_state.v[(int) fd];
-#else
- state = hash_get_drv_ev_state(fd); /* may be NULL! */
-#endif
- if (state && state->pollset == psi)
- {
- if (state->driver.select) {
-#if ERTS_CIO_DEFER_ACTIVE_EVENTS
- if (is_iotask_active(&state->driver.select->iniotask, current_cio_time)) {
- active = 1;
- if (MUST_DEFER(may_sleep)
- && (state->events & ERTS_POLL_EV_IN)
- && !(state->flags & ERTS_EV_FLAG_DEFER_IN_EV)) {
- evoff |= ERTS_POLL_EV_IN;
- state->flags |= ERTS_EV_FLAG_DEFER_IN_EV;
- }
- }
- else if (state->flags & ERTS_EV_FLAG_DEFER_IN_EV) {
- if (state->events & ERTS_POLL_EV_IN)
- evon |= ERTS_POLL_EV_IN;
- state->flags &= ~ERTS_EV_FLAG_DEFER_IN_EV;
- }
- if (is_iotask_active(&state->driver.select->outiotask, current_cio_time)) {
- active = 1;
- if (MUST_DEFER(may_sleep)
- && (state->events & ERTS_POLL_EV_OUT)
- && !(state->flags & ERTS_EV_FLAG_DEFER_OUT_EV)) {
- evoff |= ERTS_POLL_EV_OUT;
- state->flags |= ERTS_EV_FLAG_DEFER_OUT_EV;
- }
- }
- else if (state->flags & ERTS_EV_FLAG_DEFER_OUT_EV) {
- if (state->events & ERTS_POLL_EV_OUT) {
- evon |= ERTS_POLL_EV_OUT;
- }
- state->flags &= ~ERTS_EV_FLAG_DEFER_OUT_EV;
- }
- if (active)
- (void) 0;
- else
-#else
- if (is_iotask_active(&state->driver.select->iniotask, current_cio_time)
- || is_iotask_active(&state->driver.select->outiotask, current_cio_time))
- active = 1;
- else
-#endif
- if (state->type != ERTS_EV_TYPE_DRV_SEL) {
- free_select = state->driver.select;
- state->driver.select = NULL;
- }
-#if ERTS_CIO_DEFER_ACTIVE_EVENTS
- if (evon) {
- int do_wake = 0;
- ERTS_CIO_POLL_CTL(psi->ps, state->fd, evon, 1, &do_wake);
- }
-#endif
- }
-
- if (state->driver.nif) {
- ErtsPollEvents rm_events = 0;
- if (state->driver.nif->in.ddeselect_cnt) {
- ASSERT(state->type == ERTS_EV_TYPE_NIF);
- ASSERT(state->events & ERTS_POLL_EV_IN);
- ASSERT(is_nil(state->driver.nif->in.pid));
- if (may_sleep || state->driver.nif->in.ddeselect_cnt == 1) {
- rm_events = ERTS_POLL_EV_IN;
- state->driver.nif->in.ddeselect_cnt = 0;
- }
- }
- if (state->driver.nif->out.ddeselect_cnt) {
- ASSERT(state->type == ERTS_EV_TYPE_NIF);
- ASSERT(state->events & ERTS_POLL_EV_OUT);
- ASSERT(is_nil(state->driver.nif->out.pid));
- if (may_sleep || state->driver.nif->out.ddeselect_cnt == 1) {
- rm_events |= ERTS_POLL_EV_OUT;
- state->driver.nif->out.ddeselect_cnt = 0;
- }
- }
- if (rm_events) {
- int do_wake = 0;
- state->events = ERTS_CIO_POLL_CTL(state->pollset->ps, state->fd,
- rm_events, 0, &do_wake);
- }
- if (state->events)
- active = 1;
- else if (state->type != ERTS_EV_TYPE_NIF) {
- free_nif = state->driver.nif;
- state->driver.nif = NULL;
- }
- }
-
-#ifndef ERTS_SYS_CONTINOUS_FD_NUMBERS
- if (((state->type != ERTS_EV_TYPE_NONE) | state->remove_cnt | active) == 0)
- hash_erase_drv_ev_state(state);
-#endif
- }
-
- erts_mtx_unlock(mtx);
-
- if (free_select)
- free_drv_select_data(free_select);
- if (free_nif)
- free_nif_select_data(free_nif);
-
-#if ERTS_CIO_DEFER_ACTIVE_EVENTS
- if (evoff) {
- ErtsPollControlEntry *pcep = &pce[(*pce_ix)++];
- pcep->fd = fd;
- pcep->events = evoff;
- pcep->on = 0;
- }
-#endif
-
- return active;
-}
-
-static void
-check_cleanup_active_fds(struct pollset_info* psi,
- erts_aint_t current_cio_time, int may_sleep)
-{
- int six = psi->active_fd.six;
- int eix = psi->active_fd.eix;
- erts_aint32_t no = erts_atomic32_read_dirty(&psi->active_fd.no);
- const int size = psi->active_fd.size;
- int ix = six;
-#if ERTS_CIO_DEFER_ACTIVE_EVENTS
- /* every fd might add one entry */
- Uint pce_sz = sizeof(ErtsPollControlEntry)*no;
- ErtsPollControlEntry *pctrl_entries = (pce_sz
- ? erts_alloc(ERTS_ALC_T_TMP, pce_sz)
- : NULL);
- int pctrl_ix = 0;
-#endif
-
- while (ix != eix) {
- ErtsSysFdType fd = psi->active_fd.array[ix];
- int nix = ix + 1;
- if (nix >= size)
- nix = 0;
- ASSERT(fd != ERTS_SYS_FD_INVALID);
- if (!check_cleanup_active_fd(psi, fd,
-#if ERTS_CIO_DEFER_ACTIVE_EVENTS
- pctrl_entries,
- &pctrl_ix,
-#endif
- current_cio_time,
- may_sleep)) {
- no--;
- if (ix == six) {
-#ifdef DEBUG
- psi->active_fd.array[ix] = ERTS_SYS_FD_INVALID;
-#endif
- six = nix;
- }
- else {
- psi->active_fd.array[ix] = psi->active_fd.array[six];
-#ifdef DEBUG
- psi->active_fd.array[six] = ERTS_SYS_FD_INVALID;
-#endif
- six++;
- if (six >= size)
- six = 0;
- }
- }
- ix = nix;
- }
-
-#if ERTS_CIO_DEFER_ACTIVE_EVENTS
- ASSERT(pctrl_ix <= pce_sz/sizeof(ErtsPollControlEntry));
- if (pctrl_ix)
- ERTS_CIO_POLL_CTLV(psi->ps, pctrl_entries, pctrl_ix);
- if (pctrl_entries)
- erts_free(ERTS_ALC_T_TMP, pctrl_entries);
-#endif
-
- psi->active_fd.six = six;
- psi->active_fd.eix = eix;
- erts_atomic32_set_relb(&psi->active_fd.no, no);
-}
-
-static void grow_active_fds(struct pollset_info *psi)
-{
- ASSERT(psi->active_fd.six == psi->active_fd.eix);
- psi->active_fd.six = 0;
- psi->active_fd.eix = psi->active_fd.size;
- psi->active_fd.size += ERTS_ACTIVE_FD_INC;
- psi->active_fd.array = erts_realloc(ERTS_ALC_T_ACTIVE_FD_ARR,
- psi->active_fd.array,
- psi->active_fd.size*sizeof(ErtsSysFdType));
-#ifdef DEBUG
- {
- int i;
- for (i = psi->active_fd.eix + 1; i < psi->active_fd.size; i++)
- psi->active_fd.array[i] = ERTS_SYS_FD_INVALID;
- }
-#endif
-}
-
-static ERTS_INLINE void
-add_active_fd(struct pollset_info *psi, ErtsSysFdType fd)
-{
- int eix = psi->active_fd.eix;
- const int size = psi->active_fd.size;
-
- psi->active_fd.array[eix] = fd;
-
- erts_atomic32_set_relb(&psi->active_fd.no,
- (erts_atomic32_read_dirty(&psi->active_fd.no)
- + 1));
-
- eix++;
- if (eix >= size)
- eix = 0;
- psi->active_fd.eix = eix;
-
- if (psi->active_fd.six == eix) {
- grow_active_fds(psi);
- }
-}
-
int
-ERTS_CIO_EXPORT(driver_select)(ErlDrvPort ix,
- ErlDrvEvent e,
- int mode,
- int on)
+driver_select(ErlDrvPort ix, ErlDrvEvent e, int mode, int on)
{
void (*stop_select_fn)(ErlDrvEvent, void*) = NULL;
Port *prt = erts_drvport2port(ix);
@@ -899,6 +619,7 @@ ERTS_CIO_EXPORT(driver_select)(ErlDrvPort ix,
ErtsSysFdType fd = (ErtsSysFdType) e;
ErtsPollEvents ctl_events = (ErtsPollEvents) 0;
ErtsPollEvents old_events;
+ ErtsPollOp ctl_op = ERTS_POLL_OP_MOD;
ErtsDrvEventState *state;
int wake_poller = 0;
int ret;
@@ -907,32 +628,29 @@ ERTS_CIO_EXPORT(driver_select)(ErlDrvPort ix,
#ifdef USE_VM_PROBES
DTRACE_CHARBUF(name, 64);
#endif
+ ERTS_MSACC_PUSH_AND_SET_STATE(ERTS_MSACC_STATE_CHECK_IO);
- if (prt == ERTS_INVALID_ERL_DRV_PORT)
+ if (prt == ERTS_INVALID_ERL_DRV_PORT) {
+ ERTS_MSACC_POP_STATE();
return -1;
+ }
ERTS_LC_ASSERT(erts_lc_is_port_locked(prt));
#ifdef ERTS_SYS_CONTINOUS_FD_NUMBERS
- if ((unsigned)fd >= (unsigned)erts_atomic_read_nob(&drv_ev_state.len)) {
- if (fd < 0) {
- return -1;
- }
- if (fd >= drv_ev_state.max_fds) {
- drv_select_large_fd_error(ix, fd, mode, on);
- return -1;
- }
- grow_drv_ev_state(fd);
+ if (!grow_drv_ev_state(fd)) {
+ if (fd > 0) drv_select_large_fd_error(ix, fd, mode, on);
+ ERTS_MSACC_POP_STATE();
+ return -1;
}
#endif
erts_mtx_lock(fd_mtx(fd));
-#ifdef ERTS_SYS_CONTINOUS_FD_NUMBERS
- state = &drv_ev_state.v[(int) fd];
-#else
- state = hash_get_drv_ev_state(fd); /* may be NULL! */
-#endif
+ state = get_drv_ev_state(fd); /* may be NULL! */
+
+ DEBUG_PRINT_FD("driver_select(%T, %p, %s, %d)",
+ state, id, fd, drvmode2str(mode), on);
if (!on) {
if (IS_FD_UNKNOWN(state)) {
@@ -950,15 +668,10 @@ ERTS_CIO_EXPORT(driver_select)(ErlDrvPort ix,
}
else if ((mode&ERL_DRV_USE_NO_CALLBACK) == ERL_DRV_USE) {
mode |= (ERL_DRV_READ | ERL_DRV_WRITE);
- wake_poller = 1; /* to eject fd from pollset (if needed) */
}
}
-#ifndef ERTS_SYS_CONTINOUS_FD_NUMBERS
- if (state == NULL) {
- state = hash_new_drv_ev_state(fd);
- }
-#endif
+ state = new_drv_ev_state(state, fd);
switch (state->type) {
case ERTS_EV_TYPE_NIF:
@@ -982,7 +695,9 @@ ERTS_CIO_EXPORT(driver_select)(ErlDrvPort ix,
ASSERT(state->type == ERTS_EV_TYPE_NONE);
break;
- }}
+ }
+ default: break;
+ }
if (mode & ERL_DRV_READ) {
if (state->type == ERTS_EV_TYPE_DRV_SEL) {
@@ -999,7 +714,7 @@ ERTS_CIO_EXPORT(driver_select)(ErlDrvPort ix,
drv_select_steal(ix, state, mode, on);
}
ctl_events |= ERTS_POLL_EV_OUT;
- }
+ }
ASSERT((state->type == ERTS_EV_TYPE_DRV_SEL) ||
@@ -1010,32 +725,31 @@ ERTS_CIO_EXPORT(driver_select)(ErlDrvPort ix,
if (on) {
ctl_events &= ~old_events;
state->events |= ctl_events;
+ if (ctl_events & ERTS_POLL_EV_IN && (!state->driver.select || !is_iotask_active(&state->driver.select->iniotask)))
+ state->active_events |= ERTS_POLL_EV_IN;
+ if (ctl_events & ERTS_POLL_EV_OUT && (!state->driver.select || !is_iotask_active(&state->driver.select->outiotask)))
+ state->active_events |= ERTS_POLL_EV_OUT;
+ if (old_events == 0 && !(state->flags & ERTS_EV_FLAG_USED)) {
+ ctl_op = ERTS_POLL_OP_ADD;
+ }
}
else {
ctl_events &= old_events;
state->events &= ~ctl_events;
+ state->active_events &= ~ctl_events;
- if (!(state->flags & ERTS_EV_FLAG_USED)
- && old_events && !state->events) {
- /*
- * Old driver removing all events. At least wake poller.
- * It will not make close() 100% safe but it will prevent
- * actions delayed by poll timeout.
- */
- wake_poller = 1;
+ if (!state->events) {
+ if (!(state->flags & ERTS_EV_FLAG_USED) || mode & ERL_DRV_USE)
+ ctl_op = ERTS_POLL_OP_DEL;
}
}
- if (ctl_events) {
+ if (ctl_events || ctl_op == ERTS_POLL_OP_DEL) {
ErtsPollEvents new_events;
- if (!state->pollset) {
- ErtsSchedulerData* esdp = erts_get_scheduler_data();
- ASSERT(esdp);
- state->pollset = esdp->pollset;
- }
-
- new_events = ERTS_CIO_POLL_CTL(state->pollset->ps, state->fd, ctl_events, on, &wake_poller);
+ new_events = erts_io_control_wakeup(state, ctl_op,
+ state->active_events,
+ &wake_poller);
if (new_events & (ERTS_POLL_EV_ERR|ERTS_POLL_EV_NVAL)) {
if (state->type == ERTS_EV_TYPE_DRV_SEL && !old_events) {
@@ -1048,14 +762,13 @@ ERTS_CIO_EXPORT(driver_select)(ErlDrvPort ix,
goto done;
}
- ASSERT(state->type == ERTS_EV_TYPE_DRV_SEL
- || state->type == ERTS_EV_TYPE_NONE);
+ ASSERT(state->type == ERTS_EV_TYPE_DRV_SEL || state->type == ERTS_EV_TYPE_NONE);
}
if (on) {
if (ctl_events) {
if (!state->driver.select)
- state->driver.select = alloc_drv_select_data(state->pollset);
+ state->driver.select = alloc_drv_select_data(state->fd);
if (state->type == ERTS_EV_TYPE_NONE)
state->type = ERTS_EV_TYPE_DRV_SEL;
ASSERT(state->type == ERTS_EV_TYPE_DRV_SEL);
@@ -1069,47 +782,44 @@ ERTS_CIO_EXPORT(driver_select)(ErlDrvPort ix,
}
}
else { /* off */
- if (state->type == ERTS_EV_TYPE_DRV_SEL) {
- if (ctl_events & ERTS_POLL_EV_IN) {
- abort_tasks(state, ERL_DRV_READ);
- state->driver.select->inport = NIL;
- }
- if (ctl_events & ERTS_POLL_EV_OUT) {
- abort_tasks(state, ERL_DRV_WRITE);
- state->driver.select->outport = NIL;
- }
- if (state->events == 0) {
- if (old_events != 0) {
- remember_removed(state);
- }
- if ((mode & ERL_DRV_USE) || !(state->flags & ERTS_EV_FLAG_USED)) {
- state->type = ERTS_EV_TYPE_NONE;
- state->flags = 0;
- }
- /*else keep it, as fd will probably be selected upon again */
- }
- }
- if ((mode & ERL_DRV_USE_NO_CALLBACK) == ERL_DRV_USE) {
- erts_driver_t* drv_ptr = prt->drv_ptr;
- ASSERT(state->events==0);
- if (state->remove_cnt == 0 || !wake_poller) {
- /* Safe to close fd now as it is not in pollset
- or there was no need to eject fd (kernel poll) */
- stop_select_fn = drv_ptr->stop_select;
+ if (state->type == ERTS_EV_TYPE_DRV_SEL) {
+ if (ctl_events & ERTS_POLL_EV_IN) {
+ abort_tasks(state, ERL_DRV_READ);
+ state->driver.select->inport = NIL;
+ }
+ if (ctl_events & ERTS_POLL_EV_OUT) {
+ abort_tasks(state, ERL_DRV_WRITE);
+ state->driver.select->outport = NIL;
+ }
+ if (state->events == 0) {
+ if ((mode & ERL_DRV_USE) || !(state->flags & ERTS_EV_FLAG_USED)) {
+ state->type = ERTS_EV_TYPE_NONE;
+ state->flags = 0;
+ }
+ /*else keep it, as fd will probably be selected upon again */
+ }
+ }
+ if ((mode & ERL_DRV_USE_NO_CALLBACK) == ERL_DRV_USE) {
+ erts_driver_t* drv_ptr = prt->drv_ptr;
+ ASSERT(state->events==0);
+ if (!wake_poller) {
+ /* Safe to close fd now as it is not in pollset
+ or there was no need to eject fd (kernel poll) */
+ stop_select_fn = drv_ptr->stop_select;
#ifdef USE_VM_PROBES
- strncpy(name, prt->drv_ptr->name, sizeof(name)-1);
- name[sizeof(name)-1] = '\0';
+ strncpy(name, prt->drv_ptr->name, sizeof(name)-1);
+ name[sizeof(name)-1] = '\0';
#endif
- }
- else {
- /* Not safe to close fd, postpone stop_select callback. */
- state->type = ERTS_EV_TYPE_STOP_USE;
- state->driver.stop.drv_ptr = drv_ptr;
- if (drv_ptr->handle) {
- erts_ddll_reference_referenced_driver(drv_ptr->handle);
- }
- }
- }
+ }
+ else {
+ /* Not safe to close fd, postpone stop_select callback. */
+ state->type = ERTS_EV_TYPE_STOP_USE;
+ state->driver.stop.drv_ptr = drv_ptr;
+ if (drv_ptr->handle) {
+ erts_ddll_reference_referenced_driver(drv_ptr->handle);
+ }
+ }
+ }
}
ret = 0;
@@ -1134,54 +844,47 @@ done_unknown:
if (free_nif)
free_nif_select_data(free_nif);
+ ERTS_MSACC_POP_STATE();
+
return ret;
}
int
-ERTS_CIO_EXPORT(enif_select)(ErlNifEnv* env,
- ErlNifEvent e,
- enum ErlNifSelectFlags mode,
- void* obj,
- const ErlNifPid* pid,
- Eterm ref)
+enif_select(ErlNifEnv* env,
+ ErlNifEvent e,
+ enum ErlNifSelectFlags mode,
+ void* obj,
+ const ErlNifPid* pid,
+ Eterm ref)
{
int on;
ErtsResource* resource = DATA_TO_RESOURCE(obj);
ErtsSysFdType fd = (ErtsSysFdType) e;
ErtsPollEvents ctl_events = (ErtsPollEvents) 0;
ErtsPollEvents old_events;
+ ErtsPollOp ctl_op = ERTS_POLL_OP_MOD;
ErtsDrvEventState *state;
- int wake_poller;
- int ret;
+ int ret, wake_poller = 0;
enum { NO_STOP=0, CALL_STOP, CALL_STOP_AND_RELEASE } call_stop = NO_STOP;
ErtsDrvSelectDataState *free_select = NULL;
ErtsNifSelectDataState *free_nif = NULL;
-#ifdef USE_VM_PROBES
- DTRACE_CHARBUF(name, 64);
-#endif
ASSERT(!(resource->monitors && resource->monitors->is_dying));
#ifdef ERTS_SYS_CONTINOUS_FD_NUMBERS
- if ((unsigned)fd >= (unsigned)erts_atomic_read_nob(&drv_ev_state.len)) {
- if (fd < 0) {
- return INT_MIN | ERL_NIF_SELECT_INVALID_EVENT;
- }
- if (fd >= drv_ev_state.max_fds) {
- nif_select_large_fd_error(fd, mode, resource, ref);
- return INT_MIN | ERL_NIF_SELECT_INVALID_EVENT;
- }
- grow_drv_ev_state(fd);
+ if (!grow_drv_ev_state(fd)) {
+ if (fd > 0) nif_select_large_fd_error(fd, mode, resource, ref);
+ return INT_MIN | ERL_NIF_SELECT_INVALID_EVENT;
}
#endif
erts_mtx_lock(fd_mtx(fd));
-#ifdef ERTS_SYS_CONTINOUS_FD_NUMBERS
- state = &drv_ev_state.v[(int) fd];
-#else
- state = hash_get_drv_ev_state(fd); /* may be NULL! */
-#endif
+ state = get_drv_ev_state(fd); /* may be NULL! */
+
+ DEBUG_PRINT_FD("enif_select(%T, %d, %s, %p, %T, %T)",
+ state, env->proc->common.id, fd, nifmode2str(mode), resource,
+ pid ? pid->pid : THE_NON_VALUE, ref);
if (mode & ERL_NIF_SELECT_STOP) {
ASSERT(resource->type->stop);
@@ -1193,13 +896,12 @@ ERTS_CIO_EXPORT(enif_select)(ErlNifEnv* env,
}
on = 0;
mode = ERL_DRV_READ | ERL_DRV_WRITE | ERL_DRV_USE;
- wake_poller = 1; /* to eject fd from pollset (if needed) */
ctl_events = ERTS_POLL_EV_IN | ERTS_POLL_EV_OUT;
+ ctl_op = ERTS_POLL_OP_DEL;
}
else {
on = 1;
ASSERT(mode);
- wake_poller = 0;
if (mode & ERL_DRV_READ) {
ctl_events |= ERTS_POLL_EV_IN;
}
@@ -1208,11 +910,7 @@ ERTS_CIO_EXPORT(enif_select)(ErlNifEnv* env,
}
}
-#ifndef ERTS_SYS_CONTINOUS_FD_NUMBERS
- if (state == NULL) {
- state = hash_new_drv_ev_state(fd);
- }
-#endif
+ state = new_drv_ev_state(state,fd);
switch (state->type) {
case ERTS_EV_TYPE_NIF:
@@ -1243,7 +941,9 @@ ERTS_CIO_EXPORT(enif_select)(ErlNifEnv* env,
}
ASSERT(state->type == ERTS_EV_TYPE_NONE);
break;
- }}
+ }
+ default: break;
+ }
ASSERT((state->type == ERTS_EV_TYPE_NIF) ||
(state->type == ERTS_EV_TYPE_NONE && !state->events));
@@ -1253,20 +953,22 @@ ERTS_CIO_EXPORT(enif_select)(ErlNifEnv* env,
if (on) {
ctl_events &= ~old_events;
state->events |= ctl_events;
+ state->active_events |= ctl_events;
+ if (state->type == ERTS_EV_TYPE_NONE)
+ ctl_op = ERTS_POLL_OP_ADD;
}
else {
ctl_events &= old_events;
state->events &= ~ctl_events;
+ state->active_events &= ~ctl_events;
}
- if (ctl_events) {
+ if (ctl_events || ctl_op == ERTS_POLL_OP_DEL) {
ErtsPollEvents new_events;
- if (!state->pollset) {
- state->pollset = erts_get_scheduler_data()->pollset;
- }
-
- new_events = ERTS_CIO_POLL_CTL(state->pollset->ps, state->fd, ctl_events, on, &wake_poller);
+ new_events = erts_io_control_wakeup(state, ctl_op,
+ state->active_events,
+ &wake_poller);
if (new_events & (ERTS_POLL_EV_ERR|ERTS_POLL_EV_NVAL)) {
if (state->type == ERTS_EV_TYPE_NIF && !old_events) {
@@ -1274,8 +976,6 @@ ERTS_CIO_EXPORT(enif_select)(ErlNifEnv* env,
state->flags = 0;
state->driver.nif->in.pid = NIL;
state->driver.nif->out.pid = NIL;
- state->driver.nif->in.ddeselect_cnt = 0;
- state->driver.nif->out.ddeselect_cnt = 0;
state->driver.stop.resource = NULL;
}
ret = INT_MIN | ERL_NIF_SELECT_FAILED;
@@ -1307,11 +1007,9 @@ ERTS_CIO_EXPORT(enif_select)(ErlNifEnv* env,
ASSERT(is_internal_ref(ref));
refn = internal_ref_numbers(ref);
state->driver.nif->in.immed = THE_NON_VALUE;
- state->driver.nif->in.refn[0] = refn[0];
- state->driver.nif->in.refn[1] = refn[1];
- state->driver.nif->in.refn[2] = refn[2];
+ sys_memcpy(state->driver.nif->in.refn, refn,
+ sizeof(state->driver.nif->in.refn));
}
- state->driver.nif->in.ddeselect_cnt = 0;
}
if (mode & ERL_DRV_WRITE) {
state->driver.nif->out.pid = recipient;
@@ -1321,11 +1019,9 @@ ERTS_CIO_EXPORT(enif_select)(ErlNifEnv* env,
ASSERT(is_internal_ref(ref));
refn = internal_ref_numbers(ref);
state->driver.nif->out.immed = THE_NON_VALUE;
- state->driver.nif->out.refn[0] = refn[0];
- state->driver.nif->out.refn[1] = refn[1];
- state->driver.nif->out.refn[2] = refn[2];
+ sys_memcpy(state->driver.nif->out.refn, refn,
+ sizeof(state->driver.nif->out.refn));
}
- state->driver.nif->out.ddeselect_cnt = 0;
}
ret = 0;
}
@@ -1333,14 +1029,9 @@ ERTS_CIO_EXPORT(enif_select)(ErlNifEnv* env,
if (state->type == ERTS_EV_TYPE_NIF) {
state->driver.nif->in.pid = NIL;
state->driver.nif->out.pid = NIL;
- state->driver.nif->in.ddeselect_cnt = 0;
- state->driver.nif->out.ddeselect_cnt = 0;
- if (old_events != 0) {
- remember_removed(state);
- }
}
ASSERT(state->events==0);
- if (state->remove_cnt == 0 || !wake_poller) {
+ if (!wake_poller) {
/*
* Safe to close fd now as it is not in pollset
* or there was no need to eject fd (kernel poll)
@@ -1455,7 +1146,7 @@ print_driver_name(erts_dsprintf_buf_t *dsbufp, Eterm id)
static void
steal(erts_dsprintf_buf_t *dsbufp, ErtsDrvEventState *state, int mode)
{
- erts_dsprintf(dsbufp, "stealing control of fd=%d from ", (int) GET_FD(state->fd));
+ erts_dsprintf(dsbufp, "stealing control of fd=%d from ", (int) state->fd);
switch (state->type) {
case ERTS_EV_TYPE_DRV_SEL: {
int deselect_mode = 0;
@@ -1479,7 +1170,7 @@ steal(erts_dsprintf_buf_t *dsbufp, ErtsDrvEventState *state, int mode)
if (deselect_mode)
deselect(state, deselect_mode);
else {
- erts_dsprintf(dsbufp, "no one", (int) GET_FD(state->fd));
+ erts_dsprintf(dsbufp, "no one", (int) state->fd);
ASSERT(0);
}
erts_dsprintf(dsbufp, "\n");
@@ -1510,7 +1201,7 @@ steal(erts_dsprintf_buf_t *dsbufp, ErtsDrvEventState *state, int mode)
break;
}
default:
- erts_dsprintf(dsbufp, "no one\n", (int) GET_FD(state->fd));
+ erts_dsprintf(dsbufp, "no one\n", (int) state->fd);
ASSERT(0);
}
}
@@ -1524,7 +1215,7 @@ print_drv_select_op(erts_dsprintf_buf_t *dsbufp,
"driver_select(%p, %d,%s%s%s%s, %d) "
"by ",
ix,
- (int) GET_FD(fd),
+ (int) fd,
mode & ERL_DRV_READ ? " ERL_DRV_READ" : "",
mode & ERL_DRV_WRITE ? " ERL_DRV_WRITE" : "",
mode & ERL_DRV_USE ? " ERL_DRV_USE" : "",
@@ -1541,7 +1232,7 @@ print_nif_select_op(erts_dsprintf_buf_t *dsbufp,
{
erts_dsprintf(dsbufp,
"enif_select(_, %d,%s%s%s, %T:%T, %T) ",
- (int) GET_FD(fd),
+ (int) fd,
mode & ERL_NIF_SELECT_READ ? " READ" : "",
mode & ERL_NIF_SELECT_WRITE ? " WRITE" : "",
mode & ERL_NIF_SELECT_STOP ? " STOP" : "",
@@ -1673,7 +1364,7 @@ steal_pending_stop_nif(erts_dsprintf_buf_t *dsbufp, ErtsResource* resource,
erts_dsprintf(dsbufp, "called before stop was called for NIF resource %T:%T\n",
rt->module, rt->name);
- enif_release_resource(state->driver.stop.resource);
+ enif_release_resource(state->driver.stop.resource->data);
state->type = ERTS_EV_TYPE_NONE;
state->flags = 0;
state->driver.stop.resource = NULL;
@@ -1687,8 +1378,7 @@ steal_pending_stop_nif(erts_dsprintf_buf_t *dsbufp, ErtsResource* resource,
static ERTS_INLINE int
io_task_schedule_allowed(ErtsDrvEventState *state,
- ErtsPortTaskType type,
- erts_aint_t current_cio_time)
+ ErtsPortTaskType type)
{
ErtsIoTask *io_task;
@@ -1708,42 +1398,41 @@ io_task_schedule_allowed(ErtsDrvEventState *state,
return 0;
}
- return !is_iotask_active(io_task, current_cio_time);
+ return !is_iotask_active(io_task);
}
static ERTS_INLINE void
-iready(Eterm id, ErtsDrvEventState *state, erts_aint_t current_cio_time)
+iready(Eterm id, ErtsDrvEventState *state)
{
if (io_task_schedule_allowed(state,
- ERTS_PORT_TASK_INPUT,
- current_cio_time)) {
+ ERTS_PORT_TASK_INPUT)) {
ErtsIoTask *iotask = &state->driver.select->iniotask;
- erts_atomic_set_nob(&iotask->executed_time, current_cio_time);
if (erts_port_task_schedule(id,
&iotask->task,
ERTS_PORT_TASK_INPUT,
(ErlDrvEvent) state->fd) != 0) {
stale_drv_select(id, state, ERL_DRV_READ);
- }
- add_active_fd(state->pollset, state->fd);
+ } else {
+ DEBUG_PRINT_FD("schedule ready_input(%T, %d)",
+ state, id, state->fd);
+ }
}
}
static ERTS_INLINE void
-oready(Eterm id, ErtsDrvEventState *state, erts_aint_t current_cio_time)
+oready(Eterm id, ErtsDrvEventState *state)
{
if (io_task_schedule_allowed(state,
- ERTS_PORT_TASK_OUTPUT,
- current_cio_time)) {
+ ERTS_PORT_TASK_OUTPUT)) {
ErtsIoTask *iotask = &state->driver.select->outiotask;
- erts_atomic_set_nob(&iotask->executed_time, current_cio_time);
if (erts_port_task_schedule(id,
&iotask->task,
ERTS_PORT_TASK_OUTPUT,
(ErlDrvEvent) state->fd) != 0) {
stale_drv_select(id, state, ERL_DRV_WRITE);
- }
- add_active_fd(state->pollset, state->fd);
+ } else {
+ DEBUG_PRINT_FD("schedule ready_output(%T, %d)", state, id, state->fd);
+ }
}
}
@@ -1798,91 +1487,56 @@ send_event_tuple(struct erts_nif_select_event* e, ErtsResource* resource,
static void bad_fd_in_pollset(ErtsDrvEventState *, Eterm inport, Eterm outport);
void
-ERTS_CIO_EXPORT(erts_check_io_interrupt)(struct pollset_info *psi, int set)
+erts_check_io_interrupt(ErtsPollThread *psi, int set)
{
- ERTS_CIO_POLL_INTR(psi->ps, set);
+ if (psi) {
+#if ERTS_POLL_USE_FALLBACK
+ if (psi->ps == get_fallback()) {
+ erts_poll_interrupt_flbk(psi->ps, set);
+ return;
+ }
+#endif
+ erts_poll_interrupt(psi->ps, set);
+ }
}
-void
-ERTS_CIO_EXPORT(erts_check_io_interrupt_timed)(struct pollset_info *psi,
- int set,
- ErtsMonotonicTime timeout_time)
-{
- ERTS_CIO_POLL_INTR_TMD(psi->ps, set, timeout_time);
+ErtsPollThread *
+erts_create_pollset_thread(int id) {
+ return psiv+id;
}
-#ifndef __WIN32__
-/*
- * Number of ignored events, for a lingering fd added by enif_select(),
- * until we deselect fd-event from pollset.
- */
-# define ERTS_NIF_DELAYED_DESELECT 20
-#else
-/* Disable delayed deselect as pollset cannot handle active events */
-# define ERTS_NIF_DELAYED_DESELECT 1
-#endif
-
void
-ERTS_CIO_EXPORT(erts_check_io)(int do_wait)
+erts_check_io(ErtsPollThread *psi)
{
- ErtsPollResFd *pollres;
int pollres_len;
- ErtsMonotonicTime timeout_time;
int poll_ret, i;
- erts_aint_t current_cio_time;
- ErtsSchedulerData *esdp = erts_get_scheduler_data();
- struct pollset_info *psi = esdp->pollset;
+ ERTS_MSACC_PUSH_AND_SET_STATE(ERTS_MSACC_STATE_CHECK_IO);
restart:
-#ifdef ERTS_BREAK_REQUESTED
- if (ERTS_BREAK_REQUESTED)
- erts_do_break_handling();
-#endif
-
- /* Figure out timeout value */
- timeout_time = (do_wait
- ? erts_check_next_timeout_time(esdp)
- : ERTS_POLL_NO_TIMEOUT /* poll only */);
-
- /*
- * No need for an atomic inc op when incrementing
- * erts_check_io_time, since only one thread can
- * check io at a time.
- */
- current_cio_time = erts_atomic_read_dirty(&psi->check_io_time);
- current_cio_time++;
- erts_atomic_set_relb(&psi->check_io_time, current_cio_time);
-
- check_cleanup_active_fds(psi,
- current_cio_time,
- timeout_time != ERTS_POLL_NO_TIMEOUT);
-
#ifdef ERTS_ENABLE_LOCK_CHECK
erts_lc_check_exact(NULL, 0); /* No locks should be locked */
#endif
- pollres_len = erts_atomic32_read_dirty(&psi->active_fd.no) + ERTS_CHECK_IO_POLL_RES_LEN;
+ pollres_len = psi->pollres_len;
- pollres = erts_alloc(ERTS_ALC_T_TMP, sizeof(ErtsPollResFd)*pollres_len);
+#if ERTS_POLL_USE_FALLBACK
+ if (psi->ps == get_fallback()) {
- erts_atomic_set_nob(&psi->in_poll_wait, 1);
+ poll_ret = erts_poll_wait_flbk(psi->ps, psi->pollres, &pollres_len);
- poll_ret = ERTS_CIO_POLL_WAIT(psi->ps, pollres, &pollres_len, timeout_time);
+ } else
+#endif
+ {
+ poll_ret = erts_poll_wait(psi->ps, psi->pollres, &pollres_len);
+ }
#ifdef ERTS_ENABLE_LOCK_CHECK
erts_lc_check_exact(NULL, 0); /* No locks should be locked */
#endif
-#ifdef ERTS_BREAK_REQUESTED
- if (ERTS_BREAK_REQUESTED)
- erts_do_break_handling();
-#endif
-
if (poll_ret != 0) {
- erts_atomic_set_nob(&psi->in_poll_wait, 0);
- forget_removed(psi);
- erts_free(ERTS_ALC_T_TMP, pollres);
+
if (poll_ret == EAGAIN) {
goto restart;
}
@@ -1898,64 +1552,78 @@ ERTS_CIO_EXPORT(erts_check_io)(int do_wait)
erl_errno_id(poll_ret), poll_ret);
erts_send_error_to_logger_nogl(dsbufp);
}
+ ERTS_MSACC_POP_STATE();
return;
}
for (i = 0; i < pollres_len; i++) {
- ErtsSysFdType fd = (ErtsSysFdType) pollres[i].fd;
+ erts_driver_t* drv_ptr = NULL;
+ ErtsResource* resource = NULL;
+ ErtsDrvSelectDataState *free_select = NULL;
+ ErtsNifSelectDataState *free_nif = NULL;
+ ErtsSysFdType fd = (ErtsSysFdType) ERTS_POLL_RES_GET_FD(&psi->pollres[i]);
ErtsDrvEventState *state;
+ ErtsPollEvents revents;
erts_mtx_lock(fd_mtx(fd));
+ state = get_drv_ev_state(fd);
-#ifdef ERTS_SYS_CONTINOUS_FD_NUMBERS
- state = &drv_ev_state.v[ (int) fd];
-#else
- state = hash_get_drv_ev_state(fd);
if (!state) {
- goto next_pollres;
+ erts_mtx_unlock(fd_mtx(fd));
+ continue;
}
-#endif
- /* Skip this fd if it was removed from pollset */
- if (is_removed(state) || state->pollset != psi) {
- goto next_pollres;
- }
+ revents = ERTS_POLL_RES_GET_EVTS(&psi->pollres[i]);
+
+ DEBUG_PRINT_FD("triggered %s", state, ev2str(revents));
+
+ if (revents & ERTS_POLL_EV_ERR) {
+ /*
+ * Handle error events by triggering all in/out events
+ * that has been selected on.
+ * We *do not* want to call a callback that corresponds
+ * to an event not selected.
+ */
+ revents = state->active_events;
+ state->active_events = 0;
+ } else {
+
+ /* Disregard any events that are not active at the moment,
+ for instance this could happen if the driver/nif does
+ select/deselect in rapid succession. */
+ revents &= state->active_events | ERTS_POLL_EV_NVAL;
+ state->active_events &= ~revents;
+
+ /* Reactivate the poll op if there are still active events */
+ if (state->active_events) {
+ DEBUG_PRINT_FD("re-enable %s", state, ev2str(state->active_events));
+ erts_io_control(state, ERTS_POLL_OP_MOD, state->active_events);
+ }
+ }
switch (state->type) {
case ERTS_EV_TYPE_DRV_SEL: { /* Requested via driver_select()... */
- ErtsPollEvents revents = pollres[i].events;
-
- if (revents & ERTS_POLL_EV_ERR) {
- /*
- * Handle error events by triggering all in/out events
- * that the driver has selected.
- * We *do not* want to call a callback that corresponds
- * to an event not selected.
- */
- revents = state->events;
- }
- else {
- revents &= (state->events | ERTS_POLL_EV_NVAL);
- }
if (revents & (ERTS_POLL_EV_IN|ERTS_POLL_EV_OUT)) {
if (revents & ERTS_POLL_EV_OUT) {
- oready(state->driver.select->outport, state, current_cio_time);
+ oready(state->driver.select->outport, state);
}
/* Someone might have deselected input since revents
- was read therefore, update revents... */
- revents &= state->events;
+ was read (true also on the non-smp emulator since
+ oready() may have been called); therefore, update
+ revents... */
+ revents &= state->events;
if (revents & ERTS_POLL_EV_IN) {
- iready(state->driver.select->inport, state, current_cio_time);
+ iready(state->driver.select->inport, state);
}
}
else if (revents & ERTS_POLL_EV_NVAL) {
bad_fd_in_pollset(state,
state->driver.select->inport,
state->driver.select->outport);
- add_active_fd(psi, state->fd);
+ check_fd_cleanup(state, &free_select, &free_nif);
}
break;
}
@@ -1964,89 +1632,116 @@ ERTS_CIO_EXPORT(erts_check_io)(int do_wait)
struct erts_nif_select_event in = {NIL};
struct erts_nif_select_event out = {NIL};
ErtsResource* resource = NULL;
- ErtsPollEvents revents = pollres[i].events;
-
- if (revents & ERTS_POLL_EV_ERR) {
- /*
- * Handle error events by triggering all in/out events
- * that the NIF has selected.
- * We *do not* want to send a message that corresponds
- * to an event not selected.
- */
- revents = state->events;
- }
- else {
- revents &= (state->events | ERTS_POLL_EV_NVAL);
- }
if (revents & (ERTS_POLL_EV_IN|ERTS_POLL_EV_OUT)) {
if (revents & ERTS_POLL_EV_OUT) {
if (is_not_nil(state->driver.nif->out.pid)) {
out = state->driver.nif->out;
resource = state->driver.stop.resource;
- state->driver.nif->out.ddeselect_cnt = ERTS_NIF_DELAYED_DESELECT;
state->driver.nif->out.pid = NIL;
- add_active_fd(psi, state->fd);
- }
- else {
- ASSERT(state->driver.nif->out.ddeselect_cnt >= 2);
- state->driver.nif->out.ddeselect_cnt--;
}
}
if (revents & ERTS_POLL_EV_IN) {
if (is_not_nil(state->driver.nif->in.pid)) {
in = state->driver.nif->in;
resource = state->driver.stop.resource;
- state->driver.nif->in.ddeselect_cnt = ERTS_NIF_DELAYED_DESELECT;
state->driver.nif->in.pid = NIL;
- add_active_fd(psi, state->fd);
- }
- else {
- ASSERT(state->driver.nif->in.ddeselect_cnt >= 2);
- state->driver.nif->in.ddeselect_cnt--;
}
}
+ state->events &= ~revents;
}
else if (revents & ERTS_POLL_EV_NVAL) {
bad_fd_in_pollset(state, NIL, NIL);
- add_active_fd(psi, state->fd);
+ check_fd_cleanup(state, &free_select, &free_nif);
}
erts_mtx_unlock(fd_mtx(fd));
+
if (is_not_nil(in.pid)) {
send_event_tuple(&in, resource, am_ready_input);
}
if (is_not_nil(out.pid)) {
send_event_tuple(&out, resource, am_ready_output);
}
- goto next_pollres_unlocked;
+ continue;
}
+ case ERTS_EV_TYPE_STOP_NIF: {
+ resource = state->driver.stop.resource;
+ state->type = ERTS_EV_TYPE_NONE;
+ goto case_ERTS_EV_TYPE_NONE;
+ }
+
+ case ERTS_EV_TYPE_STOP_USE: {
+#if ERTS_POLL_USE_FALLBACK
+ ASSERT(psi->ps == get_fallback());
+#endif
+ drv_ptr = state->driver.stop.drv_ptr;
+ state->type = ERTS_EV_TYPE_NONE;
+ /* fallthrough */
case ERTS_EV_TYPE_NONE: /* Deselected ... */
+ case_ERTS_EV_TYPE_NONE:
+ ASSERT(!state->events && !state->active_events && !state->flags);
+ check_fd_cleanup(state, &free_select, &free_nif);
break;
+ }
default: { /* Error */
erts_dsprintf_buf_t *dsbufp;
dsbufp = erts_create_logger_dsbuf();
erts_dsprintf(dsbufp,
"Invalid event request type for fd in erts_poll()! "
- "fd=%d, event request type=%sd\n", (int) state->fd,
+ "fd=%d, event request type=%d\n", (int) state->fd,
(int) state->type);
ASSERT(0);
deselect(state, 0);
- add_active_fd(psi, state->fd);
break;
}
}
- next_pollres:;
erts_mtx_unlock(fd_mtx(fd));
- next_pollres_unlocked:;
- }
- erts_atomic_set_nob(&psi->in_poll_wait, 0);
- erts_free(ERTS_ALC_T_TMP, pollres);
- forget_removed(psi);
+ if (drv_ptr) {
+ int was_unmasked = erts_block_fpe();
+ DTRACE1(driver_stop_select, drv_ptr->name);
+ LTTNG1(driver_stop_select, drv_ptr->name);
+ (*drv_ptr->stop_select)((ErlDrvEvent) fd, NULL);
+ erts_unblock_fpe(was_unmasked);
+ if (drv_ptr->handle) {
+ erts_ddll_dereference_driver(drv_ptr->handle);
+ }
+ }
+ if (resource) {
+ erts_resource_stop(resource, (ErlNifEvent)fd, 1);
+ enif_release_resource(resource->data);
+ }
+ if (free_select)
+ free_drv_select_data(free_select);
+ if (free_nif)
+ free_nif_select_data(free_nif);
+ }
+
+ /* The entire pollres array was filled with events,
+ * grow it for the next call. We do this for two reasons:
+ * 1. Pulling out more events in on go will increase throughput
+ * 2. If the polling implementation is not fair, this will make
+ * sure that we get all fds that we can. i.e. if 12 fds are
+ * constantly active, but we only have a pollres_len of 10,
+ * two of the fds may never be triggered depending on what the
+ * kernel decides to do.
+ **/
+ if (pollres_len == psi->pollres_len) {
+ int ev_state_len = drv_ev_state_len();
+ erts_free(ERTS_ALC_T_POLLSET, psi->pollres);
+ psi->pollres_len *= 2;
+ /* Never grow it larger than the current drv_ev_state.len size */
+ if (psi->pollres_len > ev_state_len)
+ psi->pollres_len = ev_state_len;
+ psi->pollres = erts_alloc(ERTS_ALC_T_POLLSET,
+ sizeof(ErtsPollResFd) * psi->pollres_len);
+ }
+
+ ERTS_MSACC_POP_STATE();
}
static void
@@ -2167,98 +1862,189 @@ static void drv_ev_state_free(void *des)
}
#endif
-#ifdef ERTS_ENABLE_KERNEL_POLL
+#define ERTS_MAX_NO_OF_POLL_THREADS ERTS_MAX_NO_OF_SCHEDULERS
-struct io_functions {
- int (*select)(ErlDrvPort, ErlDrvEvent, int, int);
- int (*enif_select)(ErlNifEnv*, ErlNifEvent, enum ErlNifSelectFlags, void*, const ErlNifPid*, Eterm);
- void (*check_io_as_interrupt)(struct pollset_info*);
- void (*check_io_interrupt)(struct pollset_info*, int);
- void (*check_io_interrupt_tmd)(struct pollset_info*, int, ErtsMonotonicTime);
- void (*check_io)(int);
- struct pollset_info* (*get_pollset)(int sched_nr);
- void (*notify_port_task_executed)(ErtsPortTaskHandle *pthp);
- int (*max_files)(void);
- Uint (*size)(void);
- Eterm (*info)(void *);
- int (*check_io_debug)(ErtsCheckIoDebugInfo *);
-#ifdef ERTS_ENABLE_LOCK_COUNT
- void (*lcnt_update_cio_locks)(int enable);
-#endif
-};
+static char *
+get_arg(char* rest, char** argv, int* ip)
+{
+ int i = *ip;
+ if (*rest == '\0') {
+ if (argv[i+1] == NULL) {
+ erts_fprintf(stderr, "too few arguments\n");
+ erts_usage();
+ }
+ argv[i++] = NULL;
+ rest = argv[i];
+ }
+ argv[i] = NULL;
+ *ip = i;
+ return rest;
+}
-# ifdef ERTS_KERNEL_POLL_VERSION
-struct io_functions erts_io_funcs = {0};
-# else
-extern struct io_functions erts_io_funcs;
-# endif
+static void
+parse_args(int *argc, char **argv, int concurrent_waiters)
+{
+ int i = 0, j;
+ int no_pollsets = 0, no_poll_threads = 0,
+ no_pollsets_percentage = 0,
+ no_poll_threads_percentage = 0;
+ ASSERT(argc && argv);
+ while (i < *argc) {
+ if(argv[i][0] == '-') {
+ switch (argv[i][1]) {
+ case 'I': {
+ if (strncmp(argv[i]+2, "Ot", 2) == 0) {
+ char *arg = get_arg(argv[i]+4, argv, &i);
+ if (sscanf(arg, "%d", &no_poll_threads) != 1 ||
+ no_poll_threads < 1 ||
+ ERTS_MAX_NO_OF_POLL_THREADS < no_poll_threads) {
+ erts_fprintf(stderr,"bad I/O poll threads number: %s\n", arg);
+ erts_usage();
+ }
+ } else if (strncmp(argv[i]+2, "Op", 3) == 0) {
+ char *arg = get_arg(argv[i]+4, argv, &i);
+ if (sscanf(arg, "%d", &no_pollsets) != 1 ||
+ no_pollsets < 1) {
+ erts_fprintf(stderr,"bad I/O pollset number: %s\n", arg);
+ erts_usage();
+ }
+ } else if (strncmp(argv[i]+2, "OPt", 4) == 0) {
+ char *arg = get_arg(argv[i]+5, argv, &i);
+ if (sscanf(arg, "%d", &no_poll_threads_percentage) != 1 ||
+ no_poll_threads_percentage < 0 ||
+ no_poll_threads_percentage > 100) {
+ erts_fprintf(stderr,"bad I/O poll thread percentage number: %s\n", arg);
+ erts_usage();
+ }
+ } else if (strncmp(argv[i]+2, "OPp", 4) == 0) {
+ char *arg = get_arg(argv[i]+5, argv, &i);
+ if (sscanf(arg, "%d", &no_pollsets_percentage) != 1 ||
+ no_pollsets_percentage < 0 ||
+ no_pollsets_percentage > 100) {
+ erts_fprintf(stderr,"bad I/O pollset percentage number: %s\n", arg);
+ erts_usage();
+ }
+ } else {
+ break;
+ }
+ break;
+ }
+ case 'K':
+ (void)get_arg(argv[i]+2, argv, &i);
+ break;
+ case '-':
+ goto args_parsed;
+ default:
+ break;
+ }
+ }
+ i++;
+ }
+
+args_parsed:
-#endif /* ERTS_ENABLE_KERNEL_POLL */
+ if (!concurrent_waiters) {
+ no_pollsets = no_poll_threads;
+ no_pollsets_percentage = 100;
+ }
+
+ if (no_poll_threads == 0) {
+ if (no_poll_threads_percentage == 0)
+ no_poll_threads = 1; /* This is the default */
+ else {
+ no_poll_threads = erts_no_schedulers * no_poll_threads_percentage / 100;
+ if (no_poll_threads < 1)
+ no_poll_threads = 1;
+ }
+ }
+
+ if (no_pollsets == 0) {
+ if (no_pollsets_percentage == 0)
+ no_pollsets = 1; /* This is the default */
+ else {
+ no_pollsets = no_poll_threads * no_pollsets_percentage / 100;
+ if (no_pollsets < 1)
+ no_pollsets = 1;
+ }
+ }
+
+ if (no_poll_threads < no_pollsets) {
+ erts_fprintf(stderr,
+ "number of IO poll threads has to be greater or equal to "
+ "the number of \nIO pollsets. Current values are set to: \n"
+ " -IOt %d -IOp %d\n",
+ no_poll_threads, no_pollsets);
+ erts_usage();
+ }
+
+ /* Handled arguments have been marked with NULL. Slide arguments
+ not handled towards the beginning of argv. */
+ for (i = 0, j = 0; i < *argc; i++) {
+ if (argv[i])
+ argv[j++] = argv[i];
+ }
+ *argc = j;
+
+ erts_no_pollsets = no_pollsets;
+ erts_no_poll_threads = no_poll_threads;
+}
void
-ERTS_CIO_EXPORT(erts_init_check_io)(void)
+erts_init_check_io(int *argc, char **argv)
{
- int j;
+ int j, concurrent_waiters, no_poll_threads;
ERTS_CT_ASSERT((INT_MIN & (ERL_NIF_SELECT_STOP_CALLED |
ERL_NIF_SELECT_STOP_SCHEDULED |
ERL_NIF_SELECT_INVALID_EVENT |
ERL_NIF_SELECT_FAILED)) == 0);
-#ifdef ERTS_ENABLE_KERNEL_POLL
- ASSERT(erts_io_funcs.select == NULL);
- erts_io_funcs.select = ERTS_CIO_EXPORT(driver_select);
- erts_io_funcs.enif_select = ERTS_CIO_EXPORT(enif_select);
- erts_io_funcs.check_io_interrupt = ERTS_CIO_EXPORT(erts_check_io_interrupt);
- erts_io_funcs.check_io_interrupt_tmd= ERTS_CIO_EXPORT(erts_check_io_interrupt_timed);
- erts_io_funcs.check_io = ERTS_CIO_EXPORT(erts_check_io);
- erts_io_funcs.get_pollset = ERTS_CIO_EXPORT(erts_get_pollset);
- erts_io_funcs.notify_port_task_executed = ERTS_CIO_EXPORT(erts_io_notify_port_task_executed);
- erts_io_funcs.max_files = ERTS_CIO_EXPORT(erts_check_io_max_files);
- erts_io_funcs.size = ERTS_CIO_EXPORT(erts_check_io_size);
- erts_io_funcs.info = ERTS_CIO_EXPORT(erts_check_io_info);
- erts_io_funcs.check_io_debug = ERTS_CIO_EXPORT(erts_check_io_debug);
-#ifdef ERTS_ENABLE_LOCK_COUNT
- erts_io_funcs.lcnt_update_cio_locks = ERTS_CIO_EXPORT(erts_lcnt_update_cio_locks);
-#endif
-#endif
-
- init_removed_fd_alloc();
-
- ERTS_CIO_POLL_INIT();
- pollsetv = erts_alloc(ERTS_ALC_T_POLLSET,
- sizeof(struct pollset_info)*erts_no_schedulers);
- for (j=0; j < erts_no_schedulers; j++) {
- struct pollset_info* psi = &pollsetv[j];
-
- erts_atomic_init_nob(&psi->check_io_time, 0);
- erts_atomic_init_nob(&psi->in_poll_wait, 0);
- psi->ps = ERTS_CIO_NEW_POLLSET();
- psi->active_fd.six = 0;
- psi->active_fd.eix = 0;
- erts_atomic32_init_nob(&psi->active_fd.no, 0);
- psi->active_fd.size = ERTS_ACTIVE_FD_INC;
- psi->active_fd.array = erts_alloc(ERTS_ALC_T_ACTIVE_FD_ARR,
- sizeof(ErtsSysFdType)*ERTS_ACTIVE_FD_INC);
-#ifdef DEBUG
- {
- int i;
- for (i = 0; i < ERTS_ACTIVE_FD_INC; i++)
- psi->active_fd.array[i] = ERTS_SYS_FD_INVALID;
- }
+
+ erts_poll_init(&concurrent_waiters);
+#if ERTS_POLL_USE_FALLBACK
+ erts_poll_init_flbk(NULL);
#endif
- erts_atomic_init_nob(&psi->removed_list, (erts_aint_t)NULL);
+ parse_args(argc, argv, concurrent_waiters);
+
+ /* Create the actual pollsets */
+ pollsetv = erts_alloc(ERTS_ALC_T_POLLSET,sizeof(ErtsPollSet *) * erts_no_pollsets);
+
+ for (j=0; j < erts_no_pollsets; j++)
+ pollsetv[j] = erts_poll_create_pollset(j);
+
+#if ERTS_POLL_USE_FALLBACK
+ flbk_pollset = erts_poll_create_pollset_flbk(-1);
+#endif
+
+ no_poll_threads = erts_no_poll_threads;
+#if ERTS_POLL_USE_FALLBACK
+ no_poll_threads++;
+#endif
+
+ psiv = erts_alloc(ERTS_ALC_T_POLLSET, sizeof(ErtsPollThread) * no_poll_threads);
+
+#if ERTS_POLL_USE_FALLBACK
+ psiv[0].pollres_len = ERTS_CHECK_IO_POLL_RES_LEN;
+ psiv[0].pollres = erts_alloc(ERTS_ALC_T_POLLSET,
+ sizeof(ErtsPollResFd) * ERTS_CHECK_IO_POLL_RES_LEN);
+ psiv[0].ps = get_fallback();
+ psiv++;
+#endif
+
+ for (j = 0; j < erts_no_poll_threads; j++) {
+ psiv[j].pollres_len = ERTS_CHECK_IO_POLL_RES_LEN;
+ psiv[j].pollres = erts_alloc(ERTS_ALC_T_POLLSET,
+ sizeof(ErtsPollResFd) * ERTS_CHECK_IO_POLL_RES_LEN);
+ psiv[j].ps = pollsetv[j % erts_no_pollsets];
}
- {
- int i;
- for (i=0; i<DRV_EV_STATE_LOCK_CNT; i++) {
- erts_mtx_init(&drv_ev_state.locks[i].lck, "drv_ev_state", make_small(i),
- ERTS_LOCK_FLAGS_PROPERTY_STATIC | ERTS_LOCK_FLAGS_CATEGORY_IO);
- }
+ for (j=0; j < ERTS_CHECK_IO_DRV_EV_STATE_LOCK_CNT; j++) {
+ erts_mtx_init(&drv_ev_state.locks[j].lck, "drv_ev_state", make_small(j),
+ ERTS_LOCK_FLAGS_PROPERTY_STATIC | ERTS_LOCK_FLAGS_CATEGORY_IO);
}
+
#ifdef ERTS_SYS_CONTINOUS_FD_NUMBERS
- drv_ev_state.max_fds = ERTS_CIO_POLL_MAX_FDS();
+ drv_ev_state.max_fds = erts_poll_max_fds();
erts_atomic_init_nob(&drv_ev_state.len, 0);
drv_ev_state.v = NULL;
erts_mtx_init(&drv_ev_state.grow_lock, "drv_ev_state_grow", NIL,
@@ -2281,24 +2067,29 @@ ERTS_CIO_EXPORT(erts_init_check_io)(void)
}
int
-ERTS_CIO_EXPORT(erts_check_io_max_files)(void)
+erts_check_io_max_files(void)
{
#ifdef ERTS_SYS_CONTINOUS_FD_NUMBERS
return drv_ev_state.max_fds;
#else
- return ERTS_POLL_EXPORT(erts_poll_max_fds)();
+ return erts_poll_max_fds();
#endif
}
Uint
-ERTS_CIO_EXPORT(erts_check_io_size)(void)
+erts_check_io_size(void)
{
Uint res = 0;
ErtsPollInfo pi;
int i;
- for (i = 0; i < erts_no_schedulers; i++) {
- ERTS_CIO_POLL_INFO(pollsetv[i].ps, &pi);
+#if ERTS_POLL_USE_FALLBACK
+ erts_poll_info(get_fallback(), &pi);
+ res += pi.memory_size;
+#endif
+
+ for (i = 0; i < erts_no_pollsets; i++) {
+ erts_poll_info(pollsetv[i], &pi);
res += pi.memory_size;
}
#ifdef ERTS_SYS_CONTINOUS_FD_NUMBERS
@@ -2318,58 +2109,77 @@ ERTS_CIO_EXPORT(erts_check_io_size)(void)
}
Eterm
-ERTS_CIO_EXPORT(erts_check_io_info)(void *proc)
+erts_check_io_info(void *proc)
{
Process *p = (Process *) proc;
Eterm tags[16], values[16], res, list = NIL;
Uint sz, *szp, *hp, **hpp;
ErtsPollInfo *piv;
- Sint i, j;
-
- piv = erts_alloc(ERTS_ALC_T_TMP,
- sizeof(ErtsPollInfo) * erts_no_schedulers);
+ Sint i, j = 0, len;
+ int no_pollsets = erts_no_pollsets + ERTS_POLL_USE_FALLBACK;
+ ERTS_CT_ASSERT(ERTS_POLL_USE_FALLBACK == 0 || ERTS_POLL_USE_FALLBACK == 1);
- for (j = 0; j < erts_no_schedulers; j++) {
- struct pollset_info *psi = &pollsetv[j];
- erts_aint_t cio_time = erts_atomic_read_acqb(&psi->check_io_time);
+ piv = erts_alloc(ERTS_ALC_T_TMP, sizeof(ErtsPollInfo) * no_pollsets);
- piv[j].active_fds = (int) erts_atomic32_read_acqb(&psi->active_fd.no);
- while (1) {
- erts_aint_t post_cio_time;
- int post_active_fds;
+#if ERTS_POLL_USE_FALLBACK
+ erts_poll_info_flbk(get_fallback(), &piv[0]);
+ piv[0].poll_threads = 1;
+ piv[0].active_fds = 0;
+ piv++;
+#endif
- ERTS_CIO_POLL_INFO(psi->ps, &piv[j]);
+ for (j = 0; j < erts_no_pollsets; j++) {
+ erts_poll_info(pollsetv[j], &piv[j]);
+ piv[j].active_fds = 0;
+ piv[j].poll_threads = erts_no_poll_threads / erts_no_pollsets;
+ if (erts_no_poll_threads % erts_no_pollsets > j)
+ piv[j].poll_threads++;
+ }
- post_cio_time = erts_atomic_read_mb(&psi->check_io_time);
- post_active_fds = (int) erts_atomic32_read_acqb(&psi->active_fd.no);
- if (cio_time == post_cio_time && piv[j].active_fds == post_active_fds)
- break;
- cio_time = post_cio_time;
- piv[j].active_fds = post_active_fds;
+#ifdef ERTS_SYS_CONTINOUS_FD_NUMBERS
+ i = 0;
+ erts_mtx_lock(&drv_ev_state.grow_lock);
+ len = erts_atomic_read_nob(&drv_ev_state.len);
+ for (i = 0; i < ERTS_CHECK_IO_DRV_EV_STATE_LOCK_CNT; i++) {
+ erts_mtx_lock(&drv_ev_state.locks[i].lck);
+ for (j = i; j < len; j+=ERTS_CHECK_IO_DRV_EV_STATE_LOCK_CNT) {
+ ErtsDrvEventState *state = get_drv_ev_state(j);
+ int pollsetid = get_pollset_id(j);
+ ASSERT(fd_mtx(j) == &drv_ev_state.locks[i].lck);
+ if (state->flags & ERTS_EV_FLAG_FALLBACK)
+ pollsetid = -1;
+ if (state->driver.select
+ && (state->type == ERTS_EV_TYPE_DRV_SEL)
+ && (is_iotask_active(&state->driver.select->iniotask)
+ || is_iotask_active(&state->driver.select->outiotask)))
+ piv[pollsetid].active_fds++;
}
+ erts_mtx_unlock(&drv_ev_state.locks[i].lck);
+ }
+ erts_mtx_unlock(&drv_ev_state.grow_lock);
- #ifdef ERTS_SYS_CONTINOUS_FD_NUMBERS
- piv[j].memory_size += sizeof(ErtsDrvEventState) * erts_atomic_read_nob(&drv_ev_state.len);
- #else
- piv[j].memory_size += safe_hash_table_sz(&drv_ev_state.tab);
- {
- SafeHashInfo hi;
- safe_hash_get_info(&hi, &drv_ev_state.tab);
- piv[j].memory_size += hi.objs * sizeof(ErtsDrvEventState);
- }
- erts_spin_lock(&drv_ev_state.prealloc_lock);
- piv[j].memory_size += drv_ev_state.num_prealloc * sizeof(ErtsDrvEventState);
- erts_spin_unlock(&drv_ev_state.prealloc_lock);
- #endif
+ piv[0].memory_size += sizeof(ErtsDrvEventState) * erts_atomic_read_nob(&drv_ev_state.len);
+#else
+ piv[0].memory_size += safe_hash_table_sz(&drv_ev_state.tab);
+ {
+ SafeHashInfo hi;
+ safe_hash_get_info(&hi, &drv_ev_state.tab);
+ piv[0].memory_size += hi.objs * sizeof(ErtsDrvEventState);
}
+ erts_spin_lock(&drv_ev_state.prealloc_lock);
+ piv[0].memory_size += drv_ev_state.num_prealloc * sizeof(ErtsDrvEventState);
+ erts_spin_unlock(&drv_ev_state.prealloc_lock);
+#endif
hpp = NULL;
szp = &sz;
sz = 0;
+ piv -= ERTS_POLL_USE_FALLBACK;
+
bld_it:
- for (j = erts_no_schedulers-1; j >= 0; j--) {
+ for (j = no_pollsets-1; j >= 0; j--) {
i = 0;
tags[i] = erts_bld_atom(hpp, szp, "name");
@@ -2378,9 +2188,6 @@ ERTS_CIO_EXPORT(erts_check_io_info)(void *proc)
tags[i] = erts_bld_atom(hpp, szp, "primary");
values[i++] = erts_bld_atom(hpp, szp, piv[j].primary);
- tags[i] = erts_bld_atom(hpp, szp, "fallback");
- values[i++] = erts_bld_atom(hpp, szp, piv[j].fallback ? piv[j].fallback : "false");
-
tags[i] = erts_bld_atom(hpp, szp, "kernel_poll");
values[i++] = erts_bld_atom(hpp, szp,
piv[j].kernel_poll ? piv[j].kernel_poll : "false");
@@ -2389,20 +2196,13 @@ ERTS_CIO_EXPORT(erts_check_io_info)(void *proc)
values[i++] = erts_bld_uint(hpp, szp, piv[j].memory_size);
tags[i] = erts_bld_atom(hpp, szp, "total_poll_set_size");
- values[i++] = erts_bld_uint(hpp, szp, (Uint) piv[j].poll_set_size);
-
- if (piv[j].fallback) {
- tags[i] = erts_bld_atom(hpp, szp, "fallback_poll_set_size");
- values[i++] = erts_bld_uint(hpp, szp, (Uint) piv[j].fallback_poll_set_size);
- }
+ values[i++] = erts_bld_uint(hpp, szp, piv[j].poll_set_size);
tags[i] = erts_bld_atom(hpp, szp, "lazy_updates");
values[i++] = piv[j].lazy_updates ? am_true : am_false;
- if (piv[j].lazy_updates) {
- tags[i] = erts_bld_atom(hpp, szp, "pending_updates");
- values[i++] = erts_bld_uint(hpp, szp, (Uint) piv[j].pending_updates);
- }
+ tags[i] = erts_bld_atom(hpp, szp, "pending_updates");
+ values[i++] = erts_bld_uint(hpp, szp, piv[j].pending_updates);
tags[i] = erts_bld_atom(hpp, szp, "batch_updates");
values[i++] = piv[j].batch_updates ? am_true : am_false;
@@ -2410,22 +2210,17 @@ ERTS_CIO_EXPORT(erts_check_io_info)(void *proc)
tags[i] = erts_bld_atom(hpp, szp, "concurrent_updates");
values[i++] = piv[j].concurrent_updates ? am_true : am_false;
+ tags[i] = erts_bld_atom(hpp, szp, "fallback");
+ values[i++] = piv[j].is_fallback ? am_true : am_false;
+
tags[i] = erts_bld_atom(hpp, szp, "max_fds");
- values[i++] = erts_bld_uint(hpp, szp, (Uint) piv[j].max_fds);
+ values[i++] = erts_bld_uint(hpp, szp, piv[j].max_fds);
tags[i] = erts_bld_atom(hpp, szp, "active_fds");
- values[i++] = erts_bld_uint(hpp, szp, (Uint) piv[j].active_fds);
-
- #ifdef ERTS_POLL_COUNT_AVOIDED_WAKEUPS
- tags[i] = erts_bld_atom(hpp, szp, "no_avoided_wakeups");
- values[i++] = erts_bld_uint(hpp, szp, (Uint) piv[j].no_avoided_wakeups);
+ values[i++] = erts_bld_uint(hpp, szp, piv[j].active_fds);
- tags[i] = erts_bld_atom(hpp, szp, "no_avoided_interrupts");
- values[i++] = erts_bld_uint(hpp, szp, (Uint) piv[j].no_avoided_interrupts);
-
- tags[i] = erts_bld_atom(hpp, szp, "no_interrupt_timed");
- values[i++] = erts_bld_uint(hpp, szp, (Uint) piv[j].no_interrupt_timed);
- #endif
+ tags[i] = erts_bld_atom(hpp, szp, "poll_threads");
+ values[i++] = erts_bld_uint(hpp, szp, piv[j].poll_threads);
res = erts_bld_2tup_list(hpp, szp, i, tags, values);
@@ -2454,6 +2249,10 @@ static ERTS_INLINE ErtsPollEvents
print_events(ErtsPollEvents ev)
{
int first = 1;
+ if(ev == ERTS_POLL_EV_NONE) {
+ erts_printf("N/A");
+ return 0;
+ }
if(ev & ERTS_POLL_EV_IN) {
ev &= ~ERTS_POLL_EV_IN;
erts_printf("%s%s", first ? "" : "|", "IN");
@@ -2486,15 +2285,40 @@ print_flags(EventStateFlags f)
erts_printf("%s","USED");
delim = "|";
}
- if(f & ERTS_EV_FLAG_DEFER_IN_EV) {
- erts_printf("%s%s", delim, "DRIN");
+ if(f & ERTS_EV_FLAG_FALLBACK) {
+ erts_printf("%s%s", delim, "FLBK");
delim = "|";
}
- if(f & ERTS_EV_FLAG_DEFER_OUT_EV) {
- erts_printf("%s%s", delim, "DROUT");
+}
+
+#ifdef DEBUG_PRINT_MODE
+
+static ERTS_INLINE char *
+drvmode2str(int mode) {
+ switch (mode) {
+ case ERL_DRV_READ|ERL_DRV_USE: return "READ|USE";
+ case ERL_DRV_WRITE|ERL_DRV_USE: return "WRITE|USE";
+ case ERL_DRV_READ|ERL_DRV_WRITE|ERL_DRV_USE: return "READ|WRITE|USE";
+ case ERL_DRV_USE: return "USE";
+ case ERL_DRV_READ: return "READ";
+ case ERL_DRV_WRITE: return "WRITE";
+ case ERL_DRV_READ|ERL_DRV_WRITE: return "READ|WRITE";
+ default: return "UNKNOWN";
+ }
+}
+
+static ERTS_INLINE char *
+nifmode2str(enum ErlNifSelectFlags mode) {
+ switch (mode) {
+ case ERL_NIF_SELECT_READ: return "READ";
+ case ERL_NIF_SELECT_WRITE: return "WRITE";
+ case ERL_NIF_SELECT_STOP: return "STOP";
+ default: return "UNKNOWN";
}
}
+#endif
+
typedef struct {
int used_fds;
int num_errors;
@@ -2506,53 +2330,32 @@ typedef struct {
#endif
} IterDebugCounters;
-static void doit_erts_check_io_debug(void *vstate, void *vcounters)
+static int erts_debug_print_checkio_state(ErtsDrvEventState *state,
+ ErtsPollEvents ep_events,
+ int internal)
{
- ErtsDrvEventState *state = (ErtsDrvEventState *) vstate;
- IterDebugCounters *counters = (IterDebugCounters *) vcounters;
- ErtsPollEvents cio_events = state->events;
- ErtsSysFdType fd = state->fd;
-#ifdef ERTS_SYS_CONTINOUS_FD_NUMBERS
- int internal = 0;
- ErtsPollEvents ep_events = counters->epep[(int) fd];
-#endif
- int err = 0;
-
#if defined(HAVE_FSTAT) && !defined(NO_FSTAT_ON_SYS_FD_TYPE)
struct stat stat_buf;
#endif
- if (state->driver.select)
- counters->no_driver_select_structs++;
- if (state->driver.nif)
- counters->no_enif_select_structs++;
-
+ ErtsSysFdType fd = state->fd;
+ ErtsPollEvents cio_events = state->events;
+ int err = 0;
#ifdef ERTS_SYS_CONTINOUS_FD_NUMBERS
- if (state->events || ep_events) {
- if (ep_events & ERTS_POLL_EV_NVAL) {
- ep_events &= ~ERTS_POLL_EV_NVAL;
- internal = 1;
- counters->internal_fds++;
- }
- else
- counters->used_fds++;
-#else
- if (state->events) {
- counters->used_fds++;
+ ErtsPollEvents aio_events = state->active_events;
#endif
-
- erts_printf("pollset=%d fd=%d ",
- (int)(state->pollset - pollsetv), (int) fd);
-
+ erts_printf("pollset=%d fd=%d ",
+ state->flags & ERTS_EV_FLAG_FALLBACK ? -1 : get_pollset_id(fd), (int) fd);
+
#if defined(HAVE_FSTAT) && !defined(NO_FSTAT_ON_SYS_FD_TYPE)
- if (fstat((int) fd, &stat_buf) < 0)
- erts_printf("type=unknown ");
- else {
- erts_printf("type=");
+ if (fstat((int) fd, &stat_buf) < 0)
+ erts_printf("type=unknown ");
+ else {
+ erts_printf("type=");
#ifdef S_ISSOCK
- if (S_ISSOCK(stat_buf.st_mode))
- erts_printf("sock ");
- else
+ if (S_ISSOCK(stat_buf.st_mode))
+ erts_printf("sock ");
+ else
#endif
#ifdef S_ISFIFO
if (S_ISFIFO(stat_buf.st_mode))
@@ -2560,196 +2363,239 @@ static void doit_erts_check_io_debug(void *vstate, void *vcounters)
else
#endif
#ifdef S_ISCHR
- if (S_ISCHR(stat_buf.st_mode))
- erts_printf("chr ");
- else
+ if (S_ISCHR(stat_buf.st_mode))
+ erts_printf("chr ");
+ else
#endif
#ifdef S_ISDIR
- if (S_ISDIR(stat_buf.st_mode))
- erts_printf("dir ");
- else
+ if (S_ISDIR(stat_buf.st_mode))
+ erts_printf("dir ");
+ else
#endif
#ifdef S_ISBLK
- if (S_ISBLK(stat_buf.st_mode))
- erts_printf("blk ");
- else
+ if (S_ISBLK(stat_buf.st_mode))
+ erts_printf("blk ");
+ else
#endif
#ifdef S_ISREG
- if (S_ISREG(stat_buf.st_mode))
- erts_printf("reg ");
- else
+ if (S_ISREG(stat_buf.st_mode))
+ erts_printf("reg ");
+ else
#endif
#ifdef S_ISLNK
- if (S_ISLNK(stat_buf.st_mode))
- erts_printf("lnk ");
- else
+ if (S_ISLNK(stat_buf.st_mode))
+ erts_printf("lnk ");
+ else
#endif
#ifdef S_ISDOOR
- if (S_ISDOOR(stat_buf.st_mode))
- erts_printf("door ");
- else
+ if (S_ISDOOR(stat_buf.st_mode))
+ erts_printf("door ");
+ else
#endif
#ifdef S_ISWHT
- if (S_ISWHT(stat_buf.st_mode))
- erts_printf("wht ");
- else
+ if (S_ISWHT(stat_buf.st_mode))
+ erts_printf("wht ");
+ else
#endif
#ifdef S_ISXATTR
- if (S_ISXATTR(stat_buf.st_mode))
- erts_printf("xattr ");
- else
+ if (S_ISXATTR(stat_buf.st_mode))
+ erts_printf("xattr ");
+ else
#endif
- erts_printf("unknown ");
- }
+ erts_printf("unknown ");
+ }
#else
- erts_printf("type=unknown ");
+ erts_printf("type=unknown ");
#endif
- if (state->type == ERTS_EV_TYPE_DRV_SEL) {
- erts_printf("driver_select ");
-
-#ifdef ERTS_SYS_CONTINOUS_FD_NUMBERS
- if (internal) {
- erts_printf("internal ");
- err = 1;
- }
-
- if (cio_events == ep_events) {
- erts_printf("ev=");
- if (print_events(cio_events) != 0)
- err = 1;
- }
- else {
- ErtsPollEvents ev = cio_events;
-#if ERTS_CIO_DEFER_ACTIVE_EVENTS
- if (state->flags & ERTS_EV_FLAG_DEFER_IN_EV)
- ev &= ~ERTS_POLL_EV_IN;
- if (state->flags & ERTS_EV_FLAG_DEFER_OUT_EV)
- ev &= ~ERTS_POLL_EV_OUT;
-#endif
- if (ev != ep_events)
- err = 1;
- erts_printf("cio_ev=");
- print_events(cio_events);
- erts_printf(" ep_ev=");
- print_events(ep_events);
- }
-#else
- if (print_events(cio_events) != 0)
- err = 1;
-#endif
- erts_printf(" ");
- if (cio_events & ERTS_POLL_EV_IN) {
- Eterm id = state->driver.select->inport;
- if (is_nil(id)) {
- erts_printf("inport=none inname=none indrv=none ");
- err = 1;
- }
- else {
- ErtsPortNames *pnp = erts_get_port_names(id, ERTS_INVALID_ERL_DRV_PORT);
- erts_printf(" inport=%T inname=%s indrv=%s ",
- id,
- pnp->name ? pnp->name : "unknown",
- (pnp->driver_name
- ? pnp->driver_name
- : "unknown"));
- erts_free_port_names(pnp);
- }
- }
- if (cio_events & ERTS_POLL_EV_OUT) {
- Eterm id = state->driver.select->outport;
- if (is_nil(id)) {
- erts_printf("outport=none outname=none outdrv=none ");
- err = 1;
- }
- else {
- ErtsPortNames *pnp = erts_get_port_names(id, ERTS_INVALID_ERL_DRV_PORT);
- erts_printf(" outport=%T outname=%s outdrv=%s ",
- id,
- pnp->name ? pnp->name : "unknown",
- (pnp->driver_name
- ? pnp->driver_name
- : "unknown"));
- erts_free_port_names(pnp);
- }
- }
- }
- else if (state->type == ERTS_EV_TYPE_NIF) {
- ErtsResource* r;
- erts_printf("enif_select ");
+ if (state->type == ERTS_EV_TYPE_DRV_SEL) {
+ erts_printf("driver_select ");
#ifdef ERTS_SYS_CONTINOUS_FD_NUMBERS
- if (internal) {
- erts_printf("internal ");
- err = 1;
- }
-
+ if (internal) {
+ erts_printf("internal ");
+ err = 1;
+ }
+ if (aio_events == cio_events) {
if (cio_events == ep_events) {
erts_printf("ev=");
if (print_events(cio_events) != 0)
err = 1;
}
else {
- err = 1;
+ ErtsPollEvents ev = cio_events;
+ if (ev != ep_events && ep_events != ERTS_POLL_EV_NONE)
+ err = 1;
erts_printf("cio_ev=");
print_events(cio_events);
erts_printf(" ep_ev=");
print_events(ep_events);
}
+ } else {
+ erts_printf("cio_ev=");
+ print_events(cio_events);
+ erts_printf(" aio_ev=");
+ print_events(aio_events);
+ if ((aio_events != ep_events && ep_events != ERTS_POLL_EV_NONE) ||
+ (aio_events != 0 && ep_events == ERTS_POLL_EV_NONE)) {
+ erts_printf(" ep_ev=");
+ print_events(ep_events);
+ err = 1;
+ }
+ }
#else
+ if (print_events(cio_events) != 0)
+ err = 1;
+#endif
+ erts_printf(" ");
+ if (cio_events & ERTS_POLL_EV_IN) {
+ Eterm id = state->driver.select->inport;
+ if (is_nil(id)) {
+ erts_printf("inport=none inname=none indrv=none ");
+ err = 1;
+ }
+ else {
+ ErtsPortNames *pnp = erts_get_port_names(id, ERTS_INVALID_ERL_DRV_PORT);
+ erts_printf(" inport=%T inname=%s indrv=%s ",
+ id,
+ pnp->name ? pnp->name : "unknown",
+ (pnp->driver_name
+ ? pnp->driver_name
+ : "unknown"));
+ erts_free_port_names(pnp);
+ }
+ }
+ if (cio_events & ERTS_POLL_EV_OUT) {
+ Eterm id = state->driver.select->outport;
+ if (is_nil(id)) {
+ erts_printf("outport=none outname=none outdrv=none ");
+ err = 1;
+ }
+ else {
+ ErtsPortNames *pnp = erts_get_port_names(id, ERTS_INVALID_ERL_DRV_PORT);
+ erts_printf(" outport=%T outname=%s outdrv=%s ",
+ id,
+ pnp->name ? pnp->name : "unknown",
+ (pnp->driver_name
+ ? pnp->driver_name
+ : "unknown"));
+ erts_free_port_names(pnp);
+ }
+ }
+ }
+ else if (state->type == ERTS_EV_TYPE_NIF) {
+ ErtsResource* r;
+ erts_printf("enif_select ");
+
+#ifdef ERTS_SYS_CONTINOUS_FD_NUMBERS
+ if (internal) {
+ erts_printf("internal ");
+ err = 1;
+ }
+
+ if (cio_events == ep_events) {
+ erts_printf("ev=");
if (print_events(cio_events) != 0)
err = 1;
+ }
+ else {
+ err = 1;
+ erts_printf("cio_ev=");
+ print_events(cio_events);
+ erts_printf(" ep_ev=");
+ print_events(ep_events);
+ }
+#else
+ if (print_events(cio_events) != 0)
+ err = 1;
#endif
- erts_printf(" inpid=%T dd_cnt=%b32d", state->driver.nif->in.pid,
- state->driver.nif->in.ddeselect_cnt);
- erts_printf(" outpid=%T dd_cnt=%b32d", state->driver.nif->out.pid,
- state->driver.nif->out.ddeselect_cnt);
- r = state->driver.stop.resource;
- erts_printf(" resource=%p(%T:%T)", r, r->type->module, r->type->name);
+ erts_printf(" inpid=%T", state->driver.nif->in.pid);
+ erts_printf(" outpid=%T", state->driver.nif->out.pid);
+ r = state->driver.stop.resource;
+ erts_printf(" resource=%p(%T:%T)", r, r->type->module, r->type->name);
+ }
+#ifdef ERTS_SYS_CONTINOUS_FD_NUMBERS
+ else if (internal) {
+ erts_printf("internal ");
+ if (cio_events) {
+ err = 1;
+ erts_printf("cio_ev=");
+ print_events(cio_events);
+ }
+ if (ep_events) {
+ erts_printf("ep_ev=");
+ print_events(ep_events);
}
+ }
+#endif
+ else {
+ err = 1;
+ erts_printf("control_type=%d ", (int)state->type);
#ifdef ERTS_SYS_CONTINOUS_FD_NUMBERS
- else if (internal) {
- erts_printf("internal ");
- if (cio_events) {
- err = 1;
- erts_printf("cio_ev=");
- print_events(cio_events);
- }
- if (ep_events) {
- erts_printf("ep_ev=");
- print_events(ep_events);
- }
- }
+ if (cio_events == ep_events) {
+ erts_printf("ev=");
+ print_events(cio_events);
+ }
+ else {
+ erts_printf("cio_ev="); print_events(cio_events);
+ erts_printf(" ep_ev="); print_events(ep_events);
+ }
+#else
+ erts_printf("ev=0x%b32x", (Uint32) cio_events);
#endif
- else {
- err = 1;
- erts_printf("control_type=%d ", (int)state->type);
+ }
+
+ erts_printf(" flags="); print_flags(state->flags);
+ if (err) {
+ erts_printf(" ERROR");
+ }
+ erts_printf("\r\n");
+ return err;
+}
+
+static void doit_erts_check_io_debug(void *vstate, void *vcounters)
+{
+ ErtsDrvEventState *state = (ErtsDrvEventState *) vstate;
+ IterDebugCounters *counters = (IterDebugCounters *) vcounters;
+ int internal = 0;
#ifdef ERTS_SYS_CONTINOUS_FD_NUMBERS
- if (cio_events == ep_events) {
- erts_printf("ev=");
- print_events(cio_events);
- }
- else {
- erts_printf("cio_ev="); print_events(cio_events);
- erts_printf(" ep_ev="); print_events(ep_events);
- }
+ ErtsSysFdType fd = state->fd;
+ ErtsPollEvents ep_events = counters->epep[(int) fd];
#else
- erts_printf("ev=0x%b32x", (Uint32) cio_events);
+ ErtsPollEvents ep_events = ERTS_POLL_EV_NONE;
#endif
- }
-
- erts_printf(" flags="); print_flags(state->flags);
- if (err) {
+ if (state->driver.select) {
+ counters->no_driver_select_structs++;
+ ASSERT(state->events || (ep_events != 0 && ep_events != ERTS_POLL_EV_NONE));
+ }
+ if (state->driver.nif) {
+ counters->no_enif_select_structs++;
+ ASSERT(state->events || (ep_events != 0 && ep_events != ERTS_POLL_EV_NONE));
+ }
+
+#ifdef ERTS_SYS_CONTINOUS_FD_NUMBERS
+ if (state->events || (ep_events != 0 && ep_events != ERTS_POLL_EV_NONE)) {
+ if (ep_events & ERTS_POLL_EV_NVAL) {
+ ep_events &= ~ERTS_POLL_EV_NVAL;
+ internal = 1;
+ counters->internal_fds++;
+ }
+ else
+ counters->used_fds++;
+#else
+ if (state->events) {
+ counters->used_fds++;
+#endif
+ if (erts_debug_print_checkio_state(state, ep_events, internal)) {
counters->num_errors++;
- erts_printf(" ERROR");
}
- erts_printf("\n");
}
}
-
+
+/* ciodpi can be NULL when called from etp-commands */
int
-ERTS_CIO_EXPORT(erts_check_io_debug)(ErtsCheckIoDebugInfo *ciodip)
+erts_check_io_debug(ErtsCheckIoDebugInfo *ciodip)
{
#ifdef ERTS_SYS_CONTINOUS_FD_NUMBERS
int fd, len, i;
@@ -2758,12 +2604,10 @@ ERTS_CIO_EXPORT(erts_check_io_debug)(ErtsCheckIoDebugInfo *ciodip)
#ifdef ERTS_SYS_CONTINOUS_FD_NUMBERS
ErtsDrvEventState null_des;
- null_des.pollset = NULL;
null_des.driver.select = NULL;
null_des.driver.nif = NULL;
null_des.driver.stop.drv_ptr = NULL;
null_des.events = 0;
- null_des.remove_cnt = 0;
null_des.type = ERTS_EV_TYPE_NONE;
null_des.flags = 0;
@@ -2771,24 +2615,36 @@ ERTS_CIO_EXPORT(erts_check_io_debug)(ErtsCheckIoDebugInfo *ciodip)
sizeof(ErtsPollEvents)*drv_ev_state.max_fds);
#endif
- erts_printf("--- fds in pollset --------------------------------------\n");
#if defined(ERTS_ENABLE_LOCK_CHECK)
erts_lc_check_exact(NULL, 0); /* No locks should be locked */
#endif
- erts_thr_progress_block(); /* stop the world to avoid messy locking */
+ if (ciodip)
+ erts_thr_progress_block(); /* stop the world to avoid messy locking */
#ifdef ERTS_SYS_CONTINOUS_FD_NUMBERS
len = erts_atomic_read_nob(&drv_ev_state.len);
- for (i = 0; i < erts_no_schedulers; i++) {
- ERTS_POLL_EXPORT(erts_poll_get_selected_events)(pollsetv[i].ps,
- counters.epep,
- drv_ev_state.max_fds);
+#if ERTS_POLL_USE_FALLBACK
+ erts_printf("--- fds in flbk pollset ---------------------------------\n");
+ erts_poll_get_selected_events_flbk(get_fallback(), counters.epep,
+ drv_ev_state.max_fds);
+ for (fd = 0; fd < len; fd++) {
+ if (drv_ev_state.v[fd].flags & ERTS_EV_FLAG_FALLBACK)
+ doit_erts_check_io_debug(&drv_ev_state.v[fd], &counters);
+ }
+#endif
+ erts_printf("--- fds in pollset --------------------------------------\n");
+
+ for (i = 0; i < erts_no_pollsets; i++) {
+ erts_poll_get_selected_events(pollsetv[i],
+ counters.epep,
+ drv_ev_state.max_fds);
for (fd = 0; fd < len; fd++) {
- if (drv_ev_state.v[fd].pollset == &pollsetv[i])
- doit_erts_check_io_debug(&drv_ev_state.v[fd], &counters);
+ if (!(drv_ev_state.v[fd].flags & ERTS_EV_FLAG_FALLBACK)
+ && get_pollset_id(fd) == i)
+ doit_erts_check_io_debug(&drv_ev_state.v[fd], &counters);
}
}
for (fd = len ; fd < drv_ev_state.max_fds; fd++) {
@@ -2799,11 +2655,15 @@ ERTS_CIO_EXPORT(erts_check_io_debug)(ErtsCheckIoDebugInfo *ciodip)
safe_hash_for_each(&drv_ev_state.tab, &doit_erts_check_io_debug,
&counters);
#endif
- erts_thr_progress_unblock();
- ciodip->no_used_fds = counters.used_fds;
- ciodip->no_driver_select_structs = counters.no_driver_select_structs;
- ciodip->no_enif_select_structs = counters.no_enif_select_structs;
+ if (ciodip)
+ erts_thr_progress_unblock();
+
+ if (ciodip) {
+ ciodip->no_used_fds = counters.used_fds;
+ ciodip->no_driver_select_structs = counters.no_driver_select_structs;
+ ciodip->no_enif_select_structs = counters.no_enif_select_structs;
+ }
erts_printf("\n");
erts_printf("used fds=%d\n", counters.used_fds);
@@ -2822,97 +2682,19 @@ ERTS_CIO_EXPORT(erts_check_io_debug)(ErtsCheckIoDebugInfo *ciodip)
}
#ifdef ERTS_ENABLE_LOCK_COUNT
-void ERTS_CIO_EXPORT(erts_lcnt_update_cio_locks)(int enable) {
+void erts_lcnt_update_cio_locks(int enable) {
+ int i;
#ifndef ERTS_SYS_CONTINOUS_FD_NUMBERS
erts_lcnt_enable_hash_lock_count(&drv_ev_state.tab, ERTS_LOCK_FLAGS_CATEGORY_IO, enable);
#else
(void)enable;
#endif
-}
-#endif /* ERTS_ENABLE_LOCK_COUNT */
-
-#ifdef ERTS_ENABLE_KERNEL_POLL
-# ifdef ERTS_KERNEL_POLL_VERSION
-
-/*
- * Compile these only once for kp/nkp
- */
-void erts_init_check_io(void)
-{
- if (erts_use_kernel_poll)
- erts_init_check_io_kp();
- else
- erts_init_check_io_nkp();
-}
-
-int
-driver_select(ErlDrvPort port, ErlDrvEvent event, int mode, int on)
-{
- return (*erts_io_funcs.select)(port, event, mode, on);
-}
-
-int enif_select(ErlNifEnv* env, ErlNifEvent event,
- enum ErlNifSelectFlags flags, void* obj, const ErlNifPid* pid, Eterm ref)
-{
- return (*erts_io_funcs.enif_select)(env, event, flags, obj, pid, ref);
-}
-
-struct pollset_info* erts_get_pollset(int sched_nr)
-{
- return (*erts_io_funcs.get_pollset)(sched_nr);
-}
-
-void erts_io_notify_port_task_executed(ErtsPortTaskHandle *pthp)
-{
- erts_io_funcs.notify_port_task_executed(pthp);
-}
-
-void erts_check_io(int do_wait)
-{
- erts_io_funcs.check_io(do_wait);
-}
-
-Uint erts_check_io_size(void)
-{
- return erts_io_funcs.size();
-}
-
-
-Eterm erts_check_io_info(void *p)
-{
- return (*erts_io_funcs.info)(p);
-}
-
-int erts_check_io_max_files(void)
-{
- return erts_io_funcs.max_files();
-}
-
-int
-erts_check_io_debug(ErtsCheckIoDebugInfo *ip)
-{
- return (*erts_io_funcs.check_io_debug)(ip);
-}
-
-void erts_check_io_interrupt(struct pollset_info* psi, int set)
-{
- erts_io_funcs.check_io_interrupt(psi, set);
-}
-
-void erts_check_io_interrupt_timed(struct pollset_info* psi, int set,
- ErtsMonotonicTime timeout_time)
-{
- erts_io_funcs.check_io_interrupt_tmd(psi, set, timeout_time);
-}
-
-#ifdef ERTS_ENABLE_LOCK_COUNT
-void erts_lcnt_update_cio_locks(int enable)
-{
- erts_io_funcs.lcnt_update_cio_locks(enable);
-}
+#if ERTS_POLL_USE_FALLBACK
+ erts_lcnt_enable_pollset_lock_count_flbk(get_fallback(), enable);
#endif
-#endif /* ERTS_KERNEL_POLL_VERSION */
-
-#endif /* !ERTS_ENABLE_KERNEL_POLL */
+ for (i = 0; i < erts_no_pollsets; i++)
+ erts_lcnt_enable_pollset_lock_count(pollsetv[i], enable);
+}
+#endif /* ERTS_ENABLE_LOCK_COUNT */
diff --git a/erts/emulator/sys/common/erl_check_io.h b/erts/emulator/sys/common/erl_check_io.h
index ab53d91756..443ef1264c 100644
--- a/erts/emulator/sys/common/erl_check_io.h
+++ b/erts/emulator/sys/common/erl_check_io.h
@@ -18,10 +18,11 @@
* %CopyrightEnd%
*/
-/*
- * Description: Check I/O
+/**
+ * @description Check I/O, a cross platform IO polling framework for ERTS
*
- * Author: Rickard Green
+ * @author Rickard Green
+ * @author Lukas Larsson
*/
#ifndef ERL_CHECK_IO_H__
@@ -30,35 +31,96 @@
#include "sys.h"
#include "erl_sys_driver.h"
-struct pollset_info;
+/** @brief a structure that is used by each polling thread */
+struct erts_poll_thread;
+/**
+ * Get the memory size of the check io framework
+ */
Uint erts_check_io_size(void);
-Eterm erts_check_io_info(void *);
-void erts_io_notify_port_task_executed(ErtsPortTaskHandle *pthp);
-void erts_check_io_async_sig_interrupt(struct pollset_info *psi);
+/**
+ * Returns an Eterm with information about all the pollsets active at the
+ * moment.
+ *
+ * @param proc the Process* to allocate the result on. It is passed as
+ * void * because of header include problems.
+ */
+Eterm erts_check_io_info(void *proc);
+/**
+ * Should be called when a port IO task has been executed in order to re-enable
+ * or clear the information about the fd.
+ *
+ * @param type The type of event that has been completed.
+ * @param handle The port task handle of the event.
+ * @param reset A function pointer to be called when the port task handle
+ * should be reset.
+ */
+void erts_io_notify_port_task_executed(ErtsPortTaskType type,
+ ErtsPortTaskHandle *handle,
+ void (*reset)(ErtsPortTaskHandle *));
+/**
+ * Returns the maximum number of fds that the check io framework can handle.
+ */
int erts_check_io_max_files(void);
-void erts_check_io(int);
-void erts_init_check_io(void);
+/**
+ * Called by any thread that should check for new IO events. This function will
+ * not return unless erts_check_io_interrupt(pt, 1) is called by another thread.
+ *
+ * @param pt the poll thread structure to use.
+ */
+void erts_check_io(struct erts_poll_thread *pt);
+/**
+ * Initialize the check io framework. This function will parse the arguments
+ * and delete any entries that it is interested in.
+ *
+ * @param argc the number of arguments
+ * @param argv an array with the arguments
+ */
+void erts_init_check_io(int *argc, char **argv);
+/**
+ * Interrupt the poll thread so that it can execute other code.
+ *
+ * Should be called with set = 0 by the waiting thread before calling
+ * erts_check_io.
+ *
+ * @param pt the poll thread to wake
+ * @param set whether to set or clear the interrupt flag
+ */
+void erts_check_io_interrupt(struct erts_poll_thread *pt, int set);
+/**
+ * Create a new poll thread structure that is associated with the number no.
+ * It is the callers responsibility that no is unique.
+ */
+struct erts_poll_thread* erts_create_pollset_thread(int no);
#ifdef ERTS_ENABLE_LOCK_COUNT
+/**
+ * Toggle lock counting on all check io locks
+ */
void erts_lcnt_update_cio_locks(int enable);
#endif
-void erts_check_io_interrupt(struct pollset_info*, int);
-void erts_check_io_interrupt_timed(struct pollset_info*, int, ErtsMonotonicTime);
-struct pollset_info* erts_get_pollset(int sched_num);
-
typedef struct {
ErtsPortTaskHandle task;
- erts_atomic_t executed_time;
- struct pollset_info *pollset;
+ ErtsSysFdType fd;
} ErtsIoTask;
+
#endif /* ERL_CHECK_IO_H__ */
#if !defined(ERL_CHECK_IO_C__) && !defined(ERTS_ALLOC_C__)
#define ERL_CHECK_IO_INTERNAL__
#endif
+#define ERTS_CHECK_IO_DRV_EV_STATE_LOCK_CNT 128
+
+/* Controls how many pollsets to allocate. Fd's are hashed into
+ each pollset based on the FD. When doing non-concurrent updates
+ there will be one pollset per thread.
+*/
+extern int erts_no_pollsets;
+extern int erts_no_poll_threads;
+
+
#ifndef ERL_CHECK_IO_INTERNAL__
#define ERL_CHECK_IO_INTERNAL__
#include "erl_poll.h"
diff --git a/erts/emulator/sys/common/erl_poll.c b/erts/emulator/sys/common/erl_poll.c
index 341370ca14..c671ea99f4 100644
--- a/erts/emulator/sys/common/erl_poll.c
+++ b/erts/emulator/sys/common/erl_poll.c
@@ -18,9 +18,8 @@
* %CopyrightEnd%
*/
-/*
- * Description: Poll interface suitable for ERTS with or without
- * SMP support.
+/**
+ * @description Poll interface suitable for ERTS
*
* The interface is currently implemented using:
* - select
@@ -29,12 +28,36 @@
* - epoll with poll or select as fallback
* - kqueue with poll or select as fallback
*
- * Some time in the future it will also be
- * implemented using Solaris ports.
*
+ * @author Rickard Green
+ * @author Lukas Larsson
+ *
+ * There are two major different implementations off IO polling in this
+ * file. The concurrent and non-concurrent implementations.
+ * When available epoll/kqueue are used to implement the concurrent
+ * versions. poll, select and dev/poll use non-concurrent updates.
+ *
+ * Concurrent version:
+ * In the concurrent version erts_poll_control directly modifies
+ * the kernel pollset without waking the thread that is waiting
+ * on events. Also the ErtsPollResFd type is directly mapped to
+ * the native event type, so no extra copying is needed. Note that
+ * as no locking at all is done, fds can be triggered that have been
+ * removed from the pollset. The check_io layer has to deal with this.
+ *
+ * Non-concurrent version:
+ * In the non-concurrent version, the pollset has an internal representation
+ * of the pollset that is updated by erts_poll_control. When an fd is updated,
+ * its number is placed in the update request queue and then the waiting thread
+ * is woken in order to see the change. The internal data in the pollset is
+ * protected by a mutex that has to be taken by both the modifying and waiting
+ * thread at different times.
*
+ * The non-concurrent pollset cannot have fd's closed in it while a thread is
+ * waiting on that fd. In order to fix this, when an ERTS_POLL_OP_DEL command
+ * is issued, the fd is marked as closing and the waiting thread is woken. The
+ * fd is then returned in the waiting threads results as ERTS_POLL_EV_NONE.
*
- * Author: Rickard Green
*/
#ifdef HAVE_CONFIG_H
@@ -62,6 +85,8 @@
# ifdef SYS_SELECT_H
# include <sys/select.h>
# endif
+#elif defined(_DARWIN_UNLIMITED_SELECT)
+# undef _DARWIN_UNLIMITED_SELECT
#endif
#ifdef NO_SYSCONF
# if ERTS_POLL_USE_SELECT
@@ -83,20 +108,35 @@
#error "Missing implementation of erts_poll()"
#endif
-#if defined(ERTS_KERNEL_POLL_VERSION) && !ERTS_POLL_USE_KERNEL_POLL
-#error "Missing kernel poll implementation of erts_poll()"
-#endif
+#if 0
+#define ERTS_POLL_DEBUG_PRINT 1
+
+#define DEBUG_PRINT(FMT, PS, ...) \
+ do { \
+ int myerrno = errno; \
+ erts_printf("%d: " FMT "\r\n", (PS)->id, ##__VA_ARGS__); \
+ errno = myerrno; \
+ } while(0)
-#if defined(ERTS_NO_KERNEL_POLL_VERSION) && ERTS_POLL_USE_KERNEL_POLL
-#error "Kernel poll used when it shouldn't be used"
+/* Define to print info about modifications done to each fd */
+#define DEBUG_PRINT_FD(FMT, PS, FD, ...) DEBUG_PRINT("%d: " FMT, PS, FD, ##__VA_ARGS__)
+/* Define to print entry and exit from erts_poll_wait (can be very spammy) */
+//#define DEBUG_PRINT_WAIT(FMT, PS, ...) DEBUG_PRINT(FMT, PS, ##__VA_ARGS__)
+
+#else
+#define ERTS_POLL_DEBUG_PRINT 0
+#define DEBUG_PRINT(...)
#endif
-#if 0
-#define ERTS_POLL_DEBUG_PRINT
+#ifndef DEBUG_PRINT_FD
+#define DEBUG_PRINT_FD(...)
+#endif
+#ifndef DEBUG_PRINT_WAIT
+#define DEBUG_PRINT_WAIT(...)
#endif
-#ifdef _DARWIN_UNLIMITED_SELECT
+#if defined(_DARWIN_UNLIMITED_SELECT) && ERTS_POLL_USE_SELECT
typedef struct {
size_t sz;
fd_set* ptr;
@@ -142,44 +182,41 @@ int ERTS_SELECT(int nfds, ERTS_fd_set *readfds, ERTS_fd_set *writefds,
# define ERTS_SELECT select
#endif
-#define ERTS_POLL_USE_BATCH_UPDATE_POLLSET (ERTS_POLL_USE_DEVPOLL \
- || ERTS_POLL_USE_KQUEUE)
+#define ERTS_POLL_IS_FALLBACK (ERTS_POLL_USE_POLL || ERTS_POLL_USE_SELECT) && ERTS_ENABLE_KERNEL_POLL
-#define ERTS_POLL_USE_CONCURRENT_UPDATE ERTS_POLL_USE_EPOLL
+#define ERTS_POLL_USE_CONCURRENT_UPDATE (ERTS_POLL_USE_EPOLL || ERTS_POLL_USE_KQUEUE)
-#define ERTS_POLL_COALESCE_KP_RES (ERTS_POLL_USE_KQUEUE || ERTS_POLL_USE_EPOLL)
+#if !ERTS_POLL_USE_CONCURRENT_UPDATE
+
+#define ERTS_POLLSET_SET_HAVE_UPDATE_REQUESTS(PS) \
+ erts_atomic32_set_nob(&(PS)->have_update_requests, (erts_aint32_t) 1)
+#define ERTS_POLLSET_UNSET_HAVE_UPDATE_REQUESTS(PS) \
+ erts_atomic32_set_nob(&(PS)->have_update_requests, (erts_aint32_t) 0)
+#define ERTS_POLLSET_HAVE_UPDATE_REQUESTS(PS) \
+ ((int) erts_atomic32_read_nob(&(PS)->have_update_requests))
#define ERTS_POLLSET_LOCK(PS) \
erts_mtx_lock(&(PS)->mtx)
#define ERTS_POLLSET_UNLOCK(PS) \
erts_mtx_unlock(&(PS)->mtx)
-#define ERTS_POLLSET_SET_POLLED_CHK(PS) \
- ((int) erts_atomic32_xchg_nob(&(PS)->polled, (erts_aint32_t) 1))
-#define ERTS_POLLSET_UNSET_POLLED(PS) \
- erts_atomic32_set_nob(&(PS)->polled, (erts_aint32_t) 0)
-#define ERTS_POLLSET_IS_POLLED(PS) \
- ((int) erts_atomic32_read_nob(&(PS)->polled))
+#else
+#define ERTS_POLLSET_SET_HAVE_UPDATE_REQUESTS(PS)
+#define ERTS_POLLSET_UNSET_HAVE_UPDATE_REQUESTS(PS)
+#define ERTS_POLLSET_HAVE_UPDATE_REQUESTS(PS) 0
-#define ERTS_POLLSET_SET_HAVE_UPDATE_REQUESTS(PS) \
- erts_atomic32_set_nob(&(PS)->have_update_requests, (erts_aint32_t) 1)
-#define ERTS_POLLSET_UNSET_HAVE_UPDATE_REQUESTS(PS) \
- erts_atomic32_set_nob(&(PS)->have_update_requests, (erts_aint32_t) 0)
-#define ERTS_POLLSET_HAVE_UPDATE_REQUESTS(PS) \
- ((int) erts_atomic32_read_nob(&(PS)->have_update_requests))
+#define ERTS_POLLSET_LOCK(PS)
+#define ERTS_POLLSET_UNLOCK(PS)
-#if ERTS_POLL_USE_FALLBACK
-# if ERTS_POLL_USE_POLL
-# define ERTS_POLL_NEED_FALLBACK(PS) ((PS)->no_poll_fds > 1)
-# elif ERTS_POLL_USE_SELECT
-# define ERTS_POLL_NEED_FALLBACK(PS) ((PS)->no_select_fds > 1)
-# endif
#endif
+
/*
* --- Data types ------------------------------------------------------------
*/
+#if !ERTS_POLL_USE_CONCURRENT_UPDATE
+
#define ERTS_POLLSET_UPDATE_REQ_BLOCK_SIZE 128
typedef struct ErtsPollSetUpdateRequestsBlock_ ErtsPollSetUpdateRequestsBlock;
@@ -189,43 +226,20 @@ struct ErtsPollSetUpdateRequestsBlock_ {
int fds[ERTS_POLLSET_UPDATE_REQ_BLOCK_SIZE];
};
-
-
# define ERTS_POLL_FD_FLG_INURQ (((unsigned short) 1) << 0)
-#if ERTS_POLL_USE_FALLBACK
-# define ERTS_POLL_FD_FLG_INFLBCK (((unsigned short) 1) << 1)
-# define ERTS_POLL_FD_FLG_USEFLBCK (((unsigned short) 1) << 2)
-#endif
-# define ERTS_POLL_FD_FLG_RST (((unsigned short) 1) << 3)
+# define ERTS_POLL_FD_FLG_RST (((unsigned short) 1) << 1)
+
typedef struct {
#if ERTS_POLL_USE_POLL
int pix;
#endif
+
ErtsPollEvents used_events;
ErtsPollEvents events;
-#if ERTS_POLL_COALESCE_KP_RES
- unsigned short res_ev_ix;
-#endif
unsigned short flags;
} ErtsFdStatus;
-
-#if ERTS_POLL_COALESCE_KP_RES
-/* res_ev_ix max value */
-#define ERTS_POLL_MAX_RES ((1 << sizeof(unsigned short)*8) - 1)
-#endif
-
-#if ERTS_POLL_USE_KQUEUE
-
-#define ERTS_POLL_KQ_OP_HANDLED 1
-#define ERTS_POLL_KQ_OP_DEL_R 2
-#define ERTS_POLL_KQ_OP_DEL_W 3
-#define ERTS_POLL_KQ_OP_ADD_R 4
-#define ERTS_POLL_KQ_OP_ADD_W 5
-#define ERTS_POLL_KQ_OP_ADD2_R 6
-#define ERTS_POLL_KQ_OP_ADD2_W 7
-
#endif
/*
@@ -233,196 +247,172 @@ typedef struct {
* get unique names in debugger for kp/nkp
*/
struct ERTS_POLL_EXPORT(erts_pollset) {
- ErtsPollSet next;
+ int id;
int internal_fd_limit;
- ErtsFdStatus *fds_status;
erts_atomic_t no_of_user_fds;
- int fds_status_len;
+
#if ERTS_POLL_USE_KERNEL_POLL
int kp_fd;
- int res_events_len;
-#if ERTS_POLL_USE_EPOLL
- struct epoll_event *res_events;
-#elif ERTS_POLL_USE_KQUEUE
- struct kevent *res_events;
-#elif ERTS_POLL_USE_DEVPOLL
- struct pollfd *res_events;
-#endif
#endif /* ERTS_POLL_USE_KERNEL_POLL */
+
#if ERTS_POLL_USE_POLL
int next_poll_fds_ix;
int no_poll_fds;
int poll_fds_len;
- struct pollfd*poll_fds;
+ struct pollfd *poll_fds;
#elif ERTS_POLL_USE_SELECT
int next_sel_fd;
int max_fd;
-#if ERTS_POLL_USE_FALLBACK
- int no_select_fds;
-#endif
ERTS_fd_set input_fds;
ERTS_fd_set res_input_fds;
ERTS_fd_set output_fds;
ERTS_fd_set res_output_fds;
+#elif ERTS_POLL_USE_DEVPOLL
+ struct pollfd *poll_fds;
+ int poll_fds_ix;
#endif
+
+#if !ERTS_POLL_USE_CONCURRENT_UPDATE
+ ErtsFdStatus *fds_status;
+ int fds_status_len;
ErtsPollSetUpdateRequestsBlock update_requests;
ErtsPollSetUpdateRequestsBlock *curr_upd_req_block;
erts_atomic32_t have_update_requests;
- erts_atomic32_t polled;
erts_mtx_t mtx;
int wake_fds[2];
-#if ERTS_POLL_USE_TIMERFD
- int timer_fd;
-#endif
-#if ERTS_POLL_USE_FALLBACK
- int fallback_used;
-#endif
erts_atomic32_t wakeup_state;
- erts_atomic64_t timeout_time;
-#ifdef ERTS_POLL_COUNT_AVOIDED_WAKEUPS
- erts_atomic_t no_avoided_wakeups;
- erts_atomic_t no_avoided_interrupts;
- erts_atomic_t no_interrupt_timed;
#endif
};
void erts_silence_warn_unused_result(long unused);
static void fatal_error(char *format, ...);
-static void fatal_error_async_signal_safe(char *error_str);
static int max_fds = -1;
-static ErtsPollSet pollsets;
-static erts_mtx_t pollsets_lock;
#if ERTS_POLL_USE_POLL
+#if !ERTS_POLL_IS_FALLBACK
+static ERTS_INLINE short ev2pollev(ErtsPollEvents ev)
+{
+ return ERTS_POLL_EV_E2N(ev);
+}
+
+static ERTS_INLINE ErtsPollEvents pollev2ev(short ev)
+{
+ return ERTS_POLL_EV_N2E(ev);
+}
+
+#else /* ERTS_POLL_IS_FALLBACK */
+
static ERTS_INLINE short
ev2pollev(ErtsPollEvents ev)
{
-#if !ERTS_POLL_USE_FALLBACK || ERTS_POLL_USE_KQUEUE
- return ERTS_POLL_EV_E2N(ev);
-#else /* Note, we only map events we are interested in */
short res_ev = (short) 0;
if (ev & ERTS_POLL_EV_IN)
- res_ev |= ERTS_POLL_EV_NKP_IN;
+ res_ev |= ERTS_POLL_EV_NKP_IN;
if (ev & ERTS_POLL_EV_OUT)
- res_ev |= ERTS_POLL_EV_NKP_OUT;
+ res_ev |= ERTS_POLL_EV_NKP_OUT;
return res_ev;
-#endif
}
static ERTS_INLINE ErtsPollEvents
pollev2ev(short ev)
{
-#if !ERTS_POLL_USE_FALLBACK || ERTS_POLL_USE_KQUEUE
- return ERTS_POLL_EV_N2E(ev);
-#else /* Note, we only map events we are interested in */
ErtsPollEvents res_ev = (ErtsPollEvents) 0;
if (ev & ERTS_POLL_EV_NKP_IN)
- res_ev |= ERTS_POLL_EV_IN;
+ res_ev |= ERTS_POLL_EV_IN;
if (ev & ERTS_POLL_EV_NKP_OUT)
- res_ev |= ERTS_POLL_EV_OUT;
+ res_ev |= ERTS_POLL_EV_OUT;
if (ev & ERTS_POLL_EV_NKP_ERR)
- res_ev |= ERTS_POLL_EV_ERR;
+ res_ev |= ERTS_POLL_EV_ERR;
if (ev & ERTS_POLL_EV_NKP_NVAL)
- res_ev |= ERTS_POLL_EV_NVAL;
- return res_ev;
-#endif
+ res_ev |= ERTS_POLL_EV_NVAL;
+ return res_ev;
}
-#endif
+#endif /* !ERTS_POLL_IS_FALLBACK */
+
+#endif /* ERTS_POLL_USE_POLL */
+
#ifdef HARD_DEBUG
static void check_poll_result(ErtsPollResFd pr[], int len);
-#if ERTS_POLL_USE_DEVPOLL
-static void check_poll_status(ErtsPollSet ps);
-#endif /* ERTS_POLL_USE_DEVPOLL */
#endif /* HARD_DEBUG */
-#ifdef ERTS_POLL_DEBUG_PRINT
+#if ERTS_POLL_USE_DEVPOLL && defined(DEBUG)
+static void check_poll_status(ErtsPollSet *ps);
+#endif /* ERTS_POLL_USE_DEVPOLL && DEBUG */
static void print_misc_debug_info(void);
+#if ERTS_POLL_USE_EPOLL
+uint32_t epoll_events(int kp_fd, int fd);
#endif
-static ERTS_INLINE void
-init_timeout_time(ErtsPollSet ps)
-{
- erts_atomic64_init_nob(&ps->timeout_time,
- (erts_aint64_t) ERTS_MONOTONIC_TIME_MAX);
-}
-
-static ERTS_INLINE void
-set_timeout_time(ErtsPollSet ps, ErtsMonotonicTime time)
-{
- erts_atomic64_set_relb(&ps->timeout_time,
- (erts_aint64_t) time);
-}
-
-static ERTS_INLINE ErtsMonotonicTime
-get_timeout_time(ErtsPollSet ps)
-{
- return (ErtsMonotonicTime) erts_atomic64_read_acqb(&ps->timeout_time);
-}
#define ERTS_POLL_NOT_WOKEN 0
#define ERTS_POLL_WOKEN -1
#define ERTS_POLL_WOKEN_INTR 1
+#if !ERTS_POLL_USE_CONCURRENT_UPDATE
static ERTS_INLINE void
-reset_wakeup_state(ErtsPollSet ps)
+reset_wakeup_state(ErtsPollSet *ps)
{
erts_atomic32_set_mb(&ps->wakeup_state, ERTS_POLL_NOT_WOKEN);
}
+#endif
static ERTS_INLINE int
-is_woken(ErtsPollSet ps)
+is_woken(ErtsPollSet *ps)
{
+#if !ERTS_POLL_USE_CONCURRENT_UPDATE
return erts_atomic32_read_acqb(&ps->wakeup_state) != ERTS_POLL_NOT_WOKEN;
+#else
+ return 0;
+#endif
}
static ERTS_INLINE int
-is_interrupted_reset(ErtsPollSet ps)
+is_interrupted_reset(ErtsPollSet *ps)
{
+#if !ERTS_POLL_USE_CONCURRENT_UPDATE
return (erts_atomic32_xchg_acqb(&ps->wakeup_state, ERTS_POLL_NOT_WOKEN)
== ERTS_POLL_WOKEN_INTR);
+#else
+ return 0;
+#endif
}
static ERTS_INLINE void
-woke_up(ErtsPollSet ps)
+woke_up(ErtsPollSet *ps)
{
+#if !ERTS_POLL_USE_CONCURRENT_UPDATE
erts_aint32_t wakeup_state = erts_atomic32_read_acqb(&ps->wakeup_state);
if (wakeup_state == ERTS_POLL_NOT_WOKEN)
(void) erts_atomic32_cmpxchg_nob(&ps->wakeup_state,
ERTS_POLL_WOKEN,
ERTS_POLL_NOT_WOKEN);
ASSERT(erts_atomic32_read_nob(&ps->wakeup_state) != ERTS_POLL_NOT_WOKEN);
+#endif
}
/*
* --- Wakeup pipe -----------------------------------------------------------
*/
+#if !ERTS_POLL_USE_CONCURRENT_UPDATE
static ERTS_INLINE void
-wake_poller(ErtsPollSet ps, int interrupted, int async_signal_safe)
+wake_poller(ErtsPollSet *ps, int interrupted)
{
int wake;
- if (async_signal_safe)
- wake = 1;
- else {
- erts_aint32_t wakeup_state;
- if (!interrupted)
- wakeup_state = erts_atomic32_cmpxchg_relb(&ps->wakeup_state,
- ERTS_POLL_WOKEN,
- ERTS_POLL_NOT_WOKEN);
- else
- wakeup_state = erts_atomic32_xchg_relb(&ps->wakeup_state,
- ERTS_POLL_WOKEN_INTR);
- wake = wakeup_state == ERTS_POLL_NOT_WOKEN;
- }
- /*
- * NOTE: This function might be called from signal handlers in the
- * non-smp case; therefore, it has to be async-signal safe in
- * the non-smp case.
- */
+ erts_aint32_t wakeup_state;
+ if (!interrupted)
+ wakeup_state = erts_atomic32_cmpxchg_relb(&ps->wakeup_state,
+ ERTS_POLL_WOKEN,
+ ERTS_POLL_NOT_WOKEN);
+ else
+ wakeup_state = erts_atomic32_xchg_relb(&ps->wakeup_state,
+ ERTS_POLL_WOKEN_INTR);
+ wake = wakeup_state == ERTS_POLL_NOT_WOKEN;
+
if (wake) {
ssize_t res;
if (ps->wake_fds[1] < 0)
@@ -432,29 +422,27 @@ wake_poller(ErtsPollSet ps, int interrupted, int async_signal_safe)
res = write(ps->wake_fds[1], "!", 1);
} while (res < 0 && errno == EINTR);
if (res <= 0 && errno != ERRNO_BLOCK) {
- if (async_signal_safe)
- fatal_error_async_signal_safe(__FILE__
- ":XXX:wake_poller(): "
- "Failed to write on wakeup pipe\n");
- else
- fatal_error("%s:%d:wake_poller(): "
- "Failed to write to wakeup pipe fd=%d: "
- "%s (%d)\n",
- __FILE__, __LINE__,
- ps->wake_fds[1],
- erl_errno_id(errno), errno);
+ fatal_error("%s:%d:wake_poller(): "
+ "Failed to write to wakeup pipe fd=%d: "
+ "%s (%d)\n",
+ __FILE__, __LINE__,
+ ps->wake_fds[1],
+ erl_errno_id(errno), errno);
}
}
}
static ERTS_INLINE void
-cleanup_wakeup_pipe(ErtsPollSet ps)
+cleanup_wakeup_pipe(ErtsPollSet *ps)
{
+ int intr = 0;
int fd = ps->wake_fds[0];
int res;
do {
char buf[32];
res = read(fd, buf, sizeof(buf));
+ if (res > 0)
+ intr = 1;
} while (res > 0 || (res < 0 && errno == EINTR));
if (res < 0 && errno != ERRNO_BLOCK) {
fatal_error("%s:%d:cleanup_wakeup_pipe(): "
@@ -464,10 +452,12 @@ cleanup_wakeup_pipe(ErtsPollSet ps)
fd,
erl_errno_id(errno), errno);
}
+ if (intr)
+ erts_atomic32_set_nob(&ps->wakeup_state, ERTS_POLL_WOKEN_INTR);
}
static void
-create_wakeup_pipe(ErtsPollSet ps)
+create_wakeup_pipe(ErtsPollSet *ps)
{
int do_wake = 0;
int wake_fds[2];
@@ -484,20 +474,13 @@ create_wakeup_pipe(ErtsPollSet ps)
SET_NONBLOCKING(wake_fds[0]);
SET_NONBLOCKING(wake_fds[1]);
-#ifdef ERTS_POLL_DEBUG_PRINT
- erts_printf("wakeup fds = {%d, %d}\n", wake_fds[0], wake_fds[1]);
-#endif
+ DEBUG_PRINT("wakeup fds = {%d, %d}", ps, wake_fds[0], wake_fds[1]);
ERTS_POLL_EXPORT(erts_poll_control)(ps,
wake_fds[0],
+ ERTS_POLL_OP_ADD,
ERTS_POLL_EV_IN,
- 1, &do_wake);
-#if ERTS_POLL_USE_FALLBACK
- /* We depend on the wakeup pipe being handled by kernel poll */
- if (ps->fds_status[wake_fds[0]].flags & ERTS_POLL_FD_FLG_INFLBCK)
- fatal_error("%s:%d:create_wakeup_pipe(): Internal error\n",
- __FILE__, __LINE__);
-#endif
+ &do_wake);
if (ps->internal_fd_limit <= wake_fds[1])
ps->internal_fd_limit = wake_fds[1] + 1;
if (ps->internal_fd_limit <= wake_fds[0])
@@ -508,80 +491,11 @@ create_wakeup_pipe(ErtsPollSet ps)
/*
- * --- timer fd -----------------------------------------------------------
- */
-
-#if ERTS_POLL_USE_TIMERFD
-
-/* We use the timerfd when using epoll_wait to get high accuracy
- timeouts, i.e. we want to sleep with < ms accuracy. */
-
-static void
-create_timerfd(ErtsPollSet ps)
-{
- int do_wake = 0;
- int timer_fd;
- timer_fd = timerfd_create(CLOCK_MONOTONIC,0);
- ERTS_POLL_EXPORT(erts_poll_control)(ps,
- timer_fd,
- ERTS_POLL_EV_IN,
- 1, &do_wake);
-#if ERTS_POLL_USE_FALLBACK
- /* We depend on the wakeup pipe being handled by kernel poll */
- if (ps->fds_status[timer_fd].flags & ERTS_POLL_FD_FLG_INFLBCK)
- fatal_error("%s:%d:create_wakeup_pipe(): Internal error\n",
- __FILE__, __LINE__);
-#endif
- if (ps->internal_fd_limit <= timer_fd)
- ps->internal_fd_limit = timer_fd + 1;
- ps->timer_fd = timer_fd;
-}
-
-static ERTS_INLINE void
-timerfd_set(ErtsPollSet ps, struct itimerspec *its)
-{
-#ifdef DEBUG
- struct itimerspec old_its;
- int res;
- res = timerfd_settime(ps->timer_fd, 0, its, &old_its);
- ASSERT(res == 0);
- ASSERT(old_its.it_interval.tv_sec == 0 &&
- old_its.it_interval.tv_nsec == 0 &&
- old_its.it_value.tv_sec == 0 &&
- old_its.it_value.tv_nsec == 0);
-
-#else
- timerfd_settime(ps->timer_fd, 0, its, NULL);
-#endif
-}
-
-static ERTS_INLINE int
-timerfd_clear(ErtsPollSet ps, int res, int max_res) {
-
- struct itimerspec its;
- /* we always have to clear the timer */
- its.it_interval.tv_sec = 0;
- its.it_interval.tv_nsec = 0;
- its.it_value.tv_sec = 0;
- its.it_value.tv_nsec = 0;
- timerfd_settime(ps->timer_fd, 0, &its, NULL);
-
- /* only timeout fd triggered */
- if (res == 1 && ps->res_events[0].data.fd == ps->timer_fd)
- return 0;
-
- return res;
-}
-
-#endif /* ERTS_POLL_USE_TIMERFD */
-
-
-/*
* --- Poll set update requests ----------------------------------------------
*/
static ERTS_INLINE void
-enqueue_update_request(ErtsPollSet ps, int fd)
+enqueue_update_request(ErtsPollSet *ps, int fd)
{
ErtsPollSetUpdateRequestsBlock *urqbp;
@@ -596,13 +510,11 @@ enqueue_update_request(ErtsPollSet ps, int fd)
urqbp = ps->curr_upd_req_block;
if (urqbp->len == ERTS_POLLSET_UPDATE_REQ_BLOCK_SIZE) {
- ASSERT(!urqbp->next);
urqbp = erts_alloc(ERTS_ALC_T_POLLSET_UPDREQ,
sizeof(ErtsPollSetUpdateRequestsBlock));
- ps->curr_upd_req_block->next = urqbp;
- ps->curr_upd_req_block = urqbp;
- urqbp->next = NULL;
+ urqbp->next = ps->curr_upd_req_block;
urqbp->len = 0;
+ ps->curr_upd_req_block = urqbp;
}
ps->fds_status[fd].flags |= ERTS_POLL_FD_FLG_INURQ;
@@ -610,28 +522,29 @@ enqueue_update_request(ErtsPollSet ps, int fd)
}
static ERTS_INLINE void
-free_update_requests_block(ErtsPollSet ps,
+free_update_requests_block(ErtsPollSet *ps,
ErtsPollSetUpdateRequestsBlock *urqbp)
{
if (urqbp != &ps->update_requests)
erts_free(ERTS_ALC_T_POLLSET_UPDREQ, (void *) urqbp);
else {
- urqbp->next = NULL;
urqbp->len = 0;
}
}
+#endif /* !ERTS_POLL_USE_CONCURRENT_UPDATE */
/*
* --- Growing poll set structures -------------------------------------------
*/
-#ifndef ERTS_KERNEL_POLL_VERSION /* only one shared implementation */
+#if !ERTS_NO_KERNEL_POLL_VERSION || !ERTS_ENABLE_KERNEL_POLL
+/* only one shared implementation */
#define ERTS_FD_TABLE_MIN_LENGTH 1024
#define ERTS_FD_TABLE_EXP_THRESHOLD (2048*1024)
-int erts_poll_new_table_len (int old_len, int need_len)
+int erts_poll_new_table_len(int old_len, int need_len)
{
int new_len;
@@ -641,7 +554,7 @@ int erts_poll_new_table_len (int old_len, int need_len)
}
else {
new_len = old_len;
- do {
+ do {
if (new_len < ERTS_FD_TABLE_EXP_THRESHOLD)
new_len *= 2;
else
@@ -654,30 +567,9 @@ int erts_poll_new_table_len (int old_len, int need_len)
}
#endif
-#if ERTS_POLL_USE_KERNEL_POLL
-static void
-grow_res_events(ErtsPollSet ps, int new_len)
-{
- size_t new_size = sizeof(
-#if ERTS_POLL_USE_EPOLL
- struct epoll_event
-#elif ERTS_POLL_USE_DEVPOLL
- struct pollfd
-#elif ERTS_POLL_USE_KQUEUE
- struct kevent
-#endif
- ) * erts_poll_new_table_len(ps->res_events_len, new_len);
- /* We do not need to save previously stored data */
- if (ps->res_events)
- erts_free(ERTS_ALC_T_POLL_RES_EVS, ps->res_events);
- ps->res_events = erts_alloc(ERTS_ALC_T_POLL_RES_EVS, new_size);
- ps->res_events_len = new_len;
-}
-#endif /* ERTS_POLL_USE_KERNEL_POLL */
-
#if ERTS_POLL_USE_POLL
static void
-grow_poll_fds(ErtsPollSet ps, int min_ix)
+grow_poll_fds(ErtsPollSet *ps, int min_ix)
{
int i;
int new_len = erts_poll_new_table_len(ps->poll_fds_len, min_ix + 1);
@@ -725,8 +617,9 @@ ensure_select_fds(int fd, ERTS_fd_set* in, ERTS_fd_set* out)
# define ensure_select_fds(fd, in, out) do {} while(0)
#endif /* _DARWIN_UNLIMITED_SELECT */
+#if !ERTS_POLL_USE_CONCURRENT_UPDATE
static void
-grow_fds_status(ErtsPollSet ps, int min_fd)
+grow_fds_status(ErtsPollSet *ps, int min_fd)
{
int i;
int new_len = erts_poll_new_table_len(ps->fds_status_len, min_fd + 1);
@@ -745,428 +638,26 @@ grow_fds_status(ErtsPollSet ps, int min_fd)
#endif
ps->fds_status[i].used_events = (ErtsPollEvents) 0;
ps->fds_status[i].events = (ErtsPollEvents) 0;
-#if ERTS_POLL_COALESCE_KP_RES
- ps->fds_status[i].res_ev_ix = (unsigned short) ERTS_POLL_MAX_RES;
-#endif
ps->fds_status[i].flags = (unsigned short) 0;
}
ps->fds_status_len = new_len;
}
+#endif
/*
* --- Selecting fd to poll on -----------------------------------------------
*/
-#if ERTS_POLL_USE_FALLBACK
-static int update_fallback_pollset(ErtsPollSet ps, int fd);
-#endif
-
-static ERTS_INLINE int
-need_update(ErtsPollSet ps, int fd)
-{
-#if ERTS_POLL_USE_KERNEL_POLL
- int reset;
-#endif
-
- ASSERT(fd < ps->fds_status_len);
-
-#if ERTS_POLL_USE_KERNEL_POLL
- reset = (int) (ps->fds_status[fd].flags & ERTS_POLL_FD_FLG_RST);
- if (reset && !ps->fds_status[fd].used_events) {
- ps->fds_status[fd].flags &= ~ERTS_POLL_FD_FLG_RST;
- reset = 0;
- }
-#else
- ps->fds_status[fd].flags &= ~ERTS_POLL_FD_FLG_RST;
-#endif
-
- if (ps->fds_status[fd].used_events != ps->fds_status[fd].events)
- return 1;
-
-#if ERTS_POLL_USE_KERNEL_POLL
- return reset;
-#else
- return 0;
-#endif
-}
-
-#if ERTS_POLL_USE_BATCH_UPDATE_POLLSET
-
-#if ERTS_POLL_USE_KQUEUE
-#define ERTS_POLL_MIN_BATCH_BUF_SIZE 128
-#else
-#define ERTS_POLL_MIN_BATCH_BUF_SIZE 64
-#endif
-
-typedef struct {
- int len;
- int size;
-#if ERTS_POLL_USE_DEVPOLL
- struct pollfd *buf;
-#elif ERTS_POLL_USE_KQUEUE
- struct kevent *buf;
- struct kevent *ebuf;
-#endif
-} ErtsPollBatchBuf;
-
-
-static ERTS_INLINE void
-setup_batch_buf(ErtsPollSet ps, ErtsPollBatchBuf *bbp)
-{
- bbp->len = 0;
-#if ERTS_POLL_USE_DEVPOLL
- bbp->size = ps->res_events_len;
- bbp->buf = ps->res_events;
-#elif ERTS_POLL_USE_KQUEUE
- bbp->size = ps->res_events_len/2;
- bbp->buf = ps->res_events;
- bbp->ebuf = bbp->buf + bbp->size;
-#endif
-}
-
-
-#if ERTS_POLL_USE_DEVPOLL
-
-static void
-write_batch_buf(ErtsPollSet ps, ErtsPollBatchBuf *bbp)
-{
- ssize_t wres;
- char *buf = (char *) bbp->buf;
- size_t buf_size = sizeof(struct pollfd)*bbp->len;
-
- while (1) {
- wres = write(ps->kp_fd, (void *) buf, buf_size);
- if (wres < 0) {
- if (errno == EINTR)
- continue;
- fatal_error("%s:%d:write_batch_buf(): "
- "Failed to write to /dev/poll: "
- "%s (%d)\n",
- __FILE__, __LINE__,
- erl_errno_id(errno), errno);
- }
- buf_size -= wres;
- if (buf_size <= 0)
- break;
- buf += wres;
- }
-
- if (buf_size < 0) {
- fatal_error("%s:%d:write_devpoll_buf(): Internal error\n",
- __FILE__, __LINE__);
- }
- bbp->len = 0;
-}
-
-#elif ERTS_POLL_USE_KQUEUE
-
-static void
-write_batch_buf(ErtsPollSet ps, ErtsPollBatchBuf *bbp)
-{
- int res;
- int len = bbp->len;
- struct kevent *buf = bbp->buf;
- struct timespec ts = {0, 0};
-
- do {
- res = kevent(ps->kp_fd, buf, len, NULL, 0, &ts);
- } while (res < 0 && errno == EINTR);
- if (res < 0) {
- int i;
- struct kevent *ebuf = bbp->ebuf;
- do {
- res = kevent(ps->kp_fd, buf, len, ebuf, len, &ts);
- } while (res < 0 && errno == EINTR);
- if (res < 0) {
- fatal_error("%s:%d: kevent() failed: %s (%d)\n",
- __FILE__, __LINE__, erl_errno_id(errno), errno);
- }
- for (i = 0; i < res; i++) {
- if (ebuf[i].flags & EV_ERROR) {
- short filter;
- int fd = (int) ebuf[i].ident;
-
- switch ((int) (long) ebuf[i].udata) {
-
- /*
- * Since we use a lazy update approach EV_DELETE will
- * frequently fail. This since kqueue automatically
- * removes a file descriptor that is closed from the
- * poll set.
- */
- case ERTS_POLL_KQ_OP_DEL_R:
- case ERTS_POLL_KQ_OP_DEL_W:
- case ERTS_POLL_KQ_OP_HANDLED:
- break;
-
- /*
- * According to the kqueue man page EVFILT_READ support
- * does not imply EVFILT_WRITE support; therefore,
- * if an EV_ADD fail, we may have to remove other
- * events on this fd in the kqueue pollset before
- * adding fd to the fallback pollset.
- */
- case ERTS_POLL_KQ_OP_ADD_W:
- if (ps->fds_status[fd].used_events & ERTS_POLL_EV_IN) {
- filter = EVFILT_READ;
- goto rm_add_fb;
- }
- goto add_fb;
- case ERTS_POLL_KQ_OP_ADD_R:
- if (ps->fds_status[fd].used_events & ERTS_POLL_EV_OUT) {
- filter = EVFILT_WRITE;
- goto rm_add_fb;
- }
- goto add_fb;
- case ERTS_POLL_KQ_OP_ADD2_W:
- case ERTS_POLL_KQ_OP_ADD2_R: {
- int j;
- for (j = i+1; j < res; j++) {
- if (fd == (int) ebuf[j].ident) {
- ebuf[j].udata = (void *) ERTS_POLL_KQ_OP_HANDLED;
- if (!(ebuf[j].flags & EV_ERROR)) {
- switch ((int) (long) ebuf[j].udata) {
- case ERTS_POLL_KQ_OP_ADD2_W:
- filter = EVFILT_WRITE;
- goto rm_add_fb;
- case ERTS_POLL_KQ_OP_ADD2_R:
- filter = EVFILT_READ;
- goto rm_add_fb;
- default:
- fatal_error("%s:%d:write_batch_buf(): "
- "Internal error",
- __FILE__, __LINE__);
- break;
- }
- }
- goto add_fb;
- }
- }
- /* The other add succeded... */
- filter = ((((int) (long) ebuf[i].udata)
- == ERTS_POLL_KQ_OP_ADD2_W)
- ? EVFILT_READ
- : EVFILT_WRITE);
- rm_add_fb:
- {
- struct kevent kev;
- struct timespec ts = {0, 0};
- EV_SET(&kev, fd, filter, EV_DELETE, 0, 0, 0);
- (void) kevent(ps->kp_fd, &kev, 1, NULL, 0, &ts);
- }
-
- add_fb:
- ps->fds_status[fd].flags |= ERTS_POLL_FD_FLG_USEFLBCK;
- ASSERT(ps->fds_status[fd].used_events);
- ps->fds_status[fd].used_events = 0;
- erts_atomic_dec_nob(&ps->no_of_user_fds);
- update_fallback_pollset(ps, fd);
- ASSERT(ps->fds_status[fd].flags & ERTS_POLL_FD_FLG_INFLBCK);
- break;
- }
- default:
- fatal_error("%s:%d:write_batch_buf(): Internal error",
- __FILE__, __LINE__);
- break;
- }
- }
- }
- }
- bbp->len = 0;
-}
-
-#endif /* ERTS_POLL_USE_KQUEUE */
-
-static ERTS_INLINE void
-batch_update_pollset(ErtsPollSet ps, int fd, ErtsPollBatchBuf *bbp)
-{
- int buf_len;
-#if ERTS_POLL_USE_DEVPOLL
- short events;
- struct pollfd *buf;
-#elif ERTS_POLL_USE_KQUEUE
- struct kevent *buf;
-#endif
-
-#ifdef ERTS_POLL_DEBUG_PRINT
- erts_printf("Doing lazy update on fd=%d\n", fd);
-#endif
-
- if (!need_update(ps, fd))
- return;
-
- /* Make sure we have room for at least maximum no of entries
- per fd */
- if (bbp->size - bbp->len < 2)
- write_batch_buf(ps, bbp);
-
- buf_len = bbp->len;
- buf = bbp->buf;
-
- ASSERT(fd < ps->fds_status_len);
-
-#if ERTS_POLL_USE_DEVPOLL
- events = ERTS_POLL_EV_E2N(ps->fds_status[fd].events);
- if (!events) {
- buf[buf_len].events = POLLREMOVE;
- erts_atomic_dec_nob(&ps->no_of_user_fds);
- }
- else if (!ps->fds_status[fd].used_events) {
- buf[buf_len].events = events;
- erts_atomic_inc_nob(&ps->no_of_user_fds);
- }
- else {
- if ((ps->fds_status[fd].flags & ERTS_POLL_FD_FLG_RST)
- || (ps->fds_status[fd].used_events & ~events)) {
- /* Reset or removed events... */
- buf[buf_len].fd = fd;
- buf[buf_len].events = POLLREMOVE;
- buf[buf_len++].revents = 0;
- }
- buf[buf_len].events = events;
- }
- buf[buf_len].fd = fd;
- buf[buf_len++].revents = 0;
-
-#elif ERTS_POLL_USE_KQUEUE
-
- if (ps->fds_status[fd].flags & ERTS_POLL_FD_FLG_INFLBCK) {
- if (ps->fds_status[fd].flags & ERTS_POLL_FD_FLG_USEFLBCK)
- update_fallback_pollset(ps, fd);
- else { /* Remove from fallback and try kqueue */
- ErtsPollEvents events = ps->fds_status[fd].events;
- ps->fds_status[fd].events = (ErtsPollEvents) 0;
- update_fallback_pollset(ps, fd);
- ASSERT(!(ps->fds_status[fd].flags & ERTS_POLL_FD_FLG_INFLBCK));
- if (events) {
- ps->fds_status[fd].events = events;
- goto try_kqueue;
- }
- }
- }
- else {
- ErtsPollEvents events, used_events;
- int mod_w, mod_r;
- try_kqueue:
- events = ERTS_POLL_EV_E2N(ps->fds_status[fd].events);
- used_events = ERTS_POLL_EV_E2N(ps->fds_status[fd].used_events);
- if (!(ps->fds_status[fd].flags & ERTS_POLL_FD_FLG_RST)) {
- if (!used_events &&
- (events & ERTS_POLL_EV_IN) && (events & ERTS_POLL_EV_OUT))
- goto do_add_rw;
- mod_r = ((events & ERTS_POLL_EV_IN)
- != (used_events & ERTS_POLL_EV_IN));
- mod_w = ((events & ERTS_POLL_EV_OUT)
- != (used_events & ERTS_POLL_EV_OUT));
- goto do_mod;
- }
- else { /* Reset */
- if ((events & ERTS_POLL_EV_IN) && (events & ERTS_POLL_EV_OUT)) {
- do_add_rw:
- EV_SET(&buf[buf_len], fd, EVFILT_READ, EV_ADD,
- 0, 0, (void *) ERTS_POLL_KQ_OP_ADD2_R);
- buf_len++;
- EV_SET(&buf[buf_len], fd, EVFILT_WRITE, EV_ADD,
- 0, 0, (void *) ERTS_POLL_KQ_OP_ADD2_W);
- buf_len++;
-
- }
- else {
- mod_r = 1;
- mod_w = 1;
- do_mod:
- if (mod_r) {
- if (events & ERTS_POLL_EV_IN) {
- EV_SET(&buf[buf_len], fd, EVFILT_READ, EV_ADD,
- 0, 0, (void *) ERTS_POLL_KQ_OP_ADD_R);
- buf_len++;
- }
- else if (used_events & ERTS_POLL_EV_IN) {
- EV_SET(&buf[buf_len], fd, EVFILT_READ, EV_DELETE,
- 0, 0, (void *) ERTS_POLL_KQ_OP_DEL_R);
- buf_len++;
- }
- }
- if (mod_w) {
- if (events & ERTS_POLL_EV_OUT) {
- EV_SET(&buf[buf_len], fd, EVFILT_WRITE, EV_ADD,
- 0, 0, (void *) ERTS_POLL_KQ_OP_ADD_W);
- buf_len++;
- }
- else if (used_events & ERTS_POLL_EV_OUT) {
- EV_SET(&buf[buf_len], fd, EVFILT_WRITE, EV_DELETE,
- 0, 0, (void *) ERTS_POLL_KQ_OP_DEL_W);
- buf_len++;
- }
- }
- }
- }
- if (used_events) {
- if (!events) {
- erts_atomic_dec_nob(&ps->no_of_user_fds);
- }
- }
- else {
- if (events)
- erts_atomic_inc_nob(&ps->no_of_user_fds);
- }
- ASSERT((events & ~(ERTS_POLL_EV_IN|ERTS_POLL_EV_OUT)) == 0);
- ASSERT((used_events & ~(ERTS_POLL_EV_IN|ERTS_POLL_EV_OUT)) == 0);
- }
-
-#endif
-
- ps->fds_status[fd].flags &= ~ERTS_POLL_FD_FLG_RST;
- ps->fds_status[fd].used_events = ps->fds_status[fd].events;
-
- bbp->len = buf_len;
-}
-
-#else /* !ERTS_POLL_USE_BATCH_UPDATE_POLLSET */
-
#if ERTS_POLL_USE_EPOLL
static int
-#if ERTS_POLL_USE_CONCURRENT_UPDATE
-conc_update_pollset(ErtsPollSet ps, int fd, int *update_fallback)
-#else
-update_pollset(ErtsPollSet ps, int fd)
-#endif
+update_pollset(ErtsPollSet *ps, int fd, ErtsPollOp op, ErtsPollEvents events)
{
int res;
- int op;
+ int epoll_op = EPOLL_CTL_MOD;
struct epoll_event epe_templ;
struct epoll_event epe;
- ASSERT(fd < ps->fds_status_len);
-
- if (!need_update(ps, fd))
- return 0;
-
-#ifdef ERTS_POLL_DEBUG_PRINT
- erts_printf("Doing update on fd=%d\n", fd);
-#endif
- if (ps->fds_status[fd].flags & ERTS_POLL_FD_FLG_INFLBCK) {
-#if ERTS_POLL_USE_CONCURRENT_UPDATE
- if (!*update_fallback) {
- *update_fallback = 1;
- return 0;
- }
-#endif
- if (ps->fds_status[fd].flags & ERTS_POLL_FD_FLG_USEFLBCK) {
- return update_fallback_pollset(ps, fd);
- }
- else { /* Remove from fallback and try epoll */
- ErtsPollEvents events = ps->fds_status[fd].events;
- ps->fds_status[fd].events = (ErtsPollEvents) 0;
- res = update_fallback_pollset(ps, fd);
- ASSERT(!(ps->fds_status[fd].flags & ERTS_POLL_FD_FLG_INFLBCK));
- if (!events)
- return res;
- ps->fds_status[fd].events = events;
- }
- }
-
- epe_templ.events = ERTS_POLL_EV_E2N(ps->fds_status[fd].events);
+ epe_templ.events = ERTS_POLL_EV_E2N(events) | EPOLLONESHOT;
epe_templ.data.fd = fd;
#ifdef VALGRIND
@@ -1174,30 +665,24 @@ update_pollset(ErtsPollSet ps, int fd)
memset((void *) &epe.data, 0, sizeof(epoll_data_t));
#endif
- if (epe_templ.events && ps->fds_status[fd].flags & ERTS_POLL_FD_FLG_RST) {
- do {
- /* We init 'epe' every time since epoll_ctl() may modify it
- (not declared const and not documented as const). */
- epe.events = epe_templ.events;
- epe.data.fd = epe_templ.data.fd;
- res = epoll_ctl(ps->kp_fd, EPOLL_CTL_DEL, fd, &epe);
- } while (res != 0 && errno == EINTR);
- erts_atomic_dec_nob(&ps->no_of_user_fds);
- ps->fds_status[fd].used_events = 0;
- }
-
- if (!epe_templ.events) {
+ switch (op) {
+ case ERTS_POLL_OP_DEL:
/* A note on EPOLL_CTL_DEL: linux kernel versions before 2.6.9
need a non-NULL event pointer even though it is ignored... */
- op = EPOLL_CTL_DEL;
+ epoll_op = EPOLL_CTL_DEL;
+ epe_templ.events = 0;
erts_atomic_dec_nob(&ps->no_of_user_fds);
- }
- else if (!ps->fds_status[fd].used_events) {
- op = EPOLL_CTL_ADD;
+ break;
+ case ERTS_POLL_OP_ADD:
+ epoll_op = EPOLL_CTL_ADD;
erts_atomic_inc_nob(&ps->no_of_user_fds);
- }
- else {
- op = EPOLL_CTL_MOD;
+ break;
+ case ERTS_POLL_OP_MOD:
+ epoll_op = EPOLL_CTL_MOD;
+ break;
+ default:
+ ASSERT(0);
+ break;
}
do {
@@ -1205,33 +690,32 @@ update_pollset(ErtsPollSet ps, int fd)
(not declared const and not documented as const). */
epe.events = epe_templ.events;
epe.data.fd = epe_templ.data.fd;
- res = epoll_ctl(ps->kp_fd, op, fd, &epe);
+ res = epoll_ctl(ps->kp_fd, epoll_op, fd, &epe);
} while (res != 0 && errno == EINTR);
-#if defined(ERTS_POLL_DEBUG_PRINT) && 1
+#if ERTS_POLL_DEBUG_PRINT
{
int saved_errno = errno;
- erts_printf("%s = epoll_ctl(%d, %s, %d, {Ox%x, %d})\n",
- res == 0 ? "0" : erl_errno_id(errno),
- ps->kp_fd,
- (op == EPOLL_CTL_ADD
- ? "EPOLL_CTL_ADD"
- : (op == EPOLL_CTL_MOD
- ? "EPOLL_CTL_MOD"
- : (op == EPOLL_CTL_DEL
- ? "EPOLL_CTL_DEL"
- : "UNKNOWN"))),
- fd,
- epe_templ.events,
- fd);
+ DEBUG_PRINT_FD("%s = epoll_ctl(%d, %s, %d, {0x%x, %d})",
+ ps, fd,
+ res == 0 ? "0" : erl_errno_id(errno),
+ ps->kp_fd,
+ (epoll_op == EPOLL_CTL_ADD
+ ? "EPOLL_CTL_ADD"
+ : (epoll_op == EPOLL_CTL_MOD
+ ? "EPOLL_CTL_MOD"
+ : (epoll_op == EPOLL_CTL_DEL
+ ? "EPOLL_CTL_DEL"
+ : "UNKNOWN"))),
+ fd,
+ epe_templ.events,
+ fd);
errno = saved_errno;
}
#endif
- if (res == 0)
- ps->fds_status[fd].used_events = ps->fds_status[fd].events;
- else {
+ if (res != 0) {
switch (op) {
- case EPOLL_CTL_MOD:
+ case ERTS_POLL_OP_MOD:
epe.events = 0;
do {
/* We init 'epe' every time since epoll_ctl() may modify it
@@ -1240,29 +724,18 @@ update_pollset(ErtsPollSet ps, int fd)
epe.data.fd = fd;
res = epoll_ctl(ps->kp_fd, EPOLL_CTL_DEL, fd, &epe);
} while (res != 0 && errno == EINTR);
- ps->fds_status[fd].used_events = 0;
/* Fall through ... */
- case EPOLL_CTL_ADD: {
- ps->fds_status[fd].flags |= ERTS_POLL_FD_FLG_USEFLBCK;
+ case ERTS_POLL_OP_ADD: {
erts_atomic_dec_nob(&ps->no_of_user_fds);
-#if ERTS_POLL_USE_CONCURRENT_UPDATE
- if (!*update_fallback) {
- *update_fallback = 1;
- return 0;
- }
-#endif
- ASSERT(!(ps->fds_status[fd].flags & ERTS_POLL_FD_FLG_INFLBCK));
- res = update_fallback_pollset(ps, fd);
- ASSERT(ps->fds_status[fd].flags & ERTS_POLL_FD_FLG_INFLBCK);
+ res = ERTS_POLL_EV_NVAL;
break;
}
- case EPOLL_CTL_DEL: {
+ case ERTS_POLL_OP_DEL: {
/*
* Since we use a lazy update approach EPOLL_CTL_DEL will
* frequently fail. This since epoll automatically removes
* a filedescriptor that is closed from the poll set.
*/
- ps->fds_status[fd].used_events = 0;
res = 0;
break;
}
@@ -1271,67 +744,277 @@ update_pollset(ErtsPollSet ps, int fd)
__FILE__, __LINE__);
break;
}
+ } else {
+ res = events;
}
- ps->fds_status[fd].flags &= ~ERTS_POLL_FD_FLG_RST;
return res;
}
-#if ERTS_POLL_USE_CONCURRENT_UPDATE
+#endif /* ERTS_POLL_USE_EPOLL */
+
+#if ERTS_POLL_USE_KQUEUE
+
+/* Some versions of the EV_SET macro used kevp multiple times,
+ so we define out own version that make sure that it is safe
+ to do kevp++ in the argument list. */
+#define ERTS_EV_SET(kevp, a, b, c, f) do { \
+ struct kevent *kevp_ = kevp; \
+ EV_SET(kevp_, a, b, c, 0, 0, f); \
+ } while(0)
+
static int
-update_pollset(ErtsPollSet ps, int fd)
+update_pollset(ErtsPollSet *ps, int fd, ErtsPollOp op, ErtsPollEvents events)
{
- int update_fallback = 1;
- return conc_update_pollset(ps, fd, &update_fallback);
-}
-#endif
-
-#endif /* ERTS_POLL_USE_EPOLL */
+ int res = 0, len = 0;
+ struct kevent evts[2];
+ struct timespec ts = {0, 0};
-#endif /* ERTS_POLL_USE_BATCH_UPDATE_POLLSET */
+#ifdef EV_DISPATCH
+ /* If we have EV_DISPATCH we use it. The kevent descriptions for both
+ read and write are added on OP_ADD and removed on OP_DEL. And then
+ after than only EV_ENABLE|EV_DISPATCH are used.
+
+ It could be possible to not modify the pollset when disabling and/or
+ deleting events, but that may cause the poll threads to be awoken
+ a lot more than they should so we take the cost here instead of
+ in the poll thread.
+
+ Note: We need to have EV_DISPATCH both when the event is enabled and
+ disabled, as otherwise the event may be triggered twice on each re-arm.
+ Not sure if this is intended or not (can't find anything about it in the
+ man page), but it seems to be the way it works...
+ */
+
+ if (op == ERTS_POLL_OP_DEL) {
+ erts_atomic_dec_nob(&ps->no_of_user_fds);
+ /* We could probably skip this delete, do we want to? */
+ ERTS_EV_SET(&evts[len++], fd, EVFILT_READ, EV_DELETE, (void *) 0);
+ ERTS_EV_SET(&evts[len++], fd, EVFILT_WRITE, EV_DELETE, (void *) 0);
+ } else if (op == ERTS_POLL_OP_ADD) {
+ uint32_t flags;
+ erts_atomic_inc_nob(&ps->no_of_user_fds);
+
+ flags = EV_ADD|EV_DISPATCH;
+ flags |= ((events & ERTS_POLL_EV_IN) ? 0 : EV_DISABLE);
+ ERTS_EV_SET(&evts[len++], fd, EVFILT_READ, flags, (void *) ERTS_POLL_EV_IN);
+
+ flags = EV_ADD|EV_DISPATCH;
+ flags |= ((events & ERTS_POLL_EV_OUT) ? 0 : EV_DISABLE);
+ ERTS_EV_SET(&evts[len++], fd, EVFILT_WRITE, flags, (void *) ERTS_POLL_EV_OUT);
+ } else {
+ uint32_t flags;
+ ASSERT(op == ERTS_POLL_OP_MOD);
-#if ERTS_POLL_USE_POLL || ERTS_POLL_USE_SELECT || ERTS_POLL_USE_FALLBACK
+ flags = EV_DISPATCH;
+ flags |= (events & ERTS_POLL_EV_IN) ? EV_ENABLE : EV_DISABLE;
+ ERTS_EV_SET(&evts[len++], fd, EVFILT_READ, flags, (void *) ERTS_POLL_EV_IN);
-#if ERTS_POLL_USE_FALLBACK
-static int update_fallback_pollset(ErtsPollSet ps, int fd)
+ flags = EV_DISPATCH;
+ flags |= (events & ERTS_POLL_EV_OUT) ? EV_ENABLE : EV_DISABLE;
+ ERTS_EV_SET(&evts[len++], fd, EVFILT_WRITE, flags, (void *) ERTS_POLL_EV_OUT);
+ }
#else
-static int update_pollset(ErtsPollSet ps, int fd)
+ uint32_t flags = EV_ADD|EV_ONESHOT;
+
+ if (op == ERTS_POLL_OP_DEL) {
+ erts_atomic_dec_nob(&ps->no_of_user_fds);
+ /* We don't do anything when a delete is issued. The fds will be removed
+ when they are triggered, or when they are closed. */
+ events = 0;
+ } else if (op == ERTS_POLL_OP_ADD) {
+ erts_atomic_inc_nob(&ps->no_of_user_fds);
+ }
+
+ if (events & ERTS_POLL_EV_IN) {
+ ERTS_EV_SET(&evts[len++], fd, EVFILT_READ, flags, (void *) ERTS_POLL_EV_IN);
+ }
+ if (events & ERTS_POLL_EV_OUT) {
+ ERTS_EV_SET(&evts[len++], fd, EVFILT_WRITE, flags, (void *) ERTS_POLL_EV_OUT);
+ }
+
#endif
-{
-#ifdef ERTS_POLL_DEBUG_PRINT
-#if ERTS_POLL_USE_FALLBACK
- erts_printf("Doing fallback update on fd=%d\n", fd);
+ if (len)
+ do {
+ res = kevent(ps->kp_fd, evts, len, NULL, 0, &ts);
+ } while (res < 0 && errno == EINTR);
+#if ERTS_POLL_DEBUG_PRINT
+ {
+ int saved_errno = errno, i;
+ char keventb[255], *keventbp = keventb;
+ if (res < 0)
+ keventbp += sprintf(keventbp,"%s = ",erl_errno_id(saved_errno));
+ else
+ keventbp += sprintf(keventbp,"%d = ",res);
+ keventbp += sprintf(keventbp, "kevent(%d, {",ps->kp_fd);
+ for (i = 0; i < len; i++) {
+ const char *flags = "UNKNOWN";
+ if (evts[i].flags == EV_DELETE) flags = "EV_DELETE";
+ if (evts[i].flags == (EV_ADD|EV_ONESHOT)) flags = "EV_ADD|EV_ONESHOT";
+#ifdef EV_DISPATCH
+ if (evts[i].flags == (EV_ADD|EV_DISPATCH)) flags = "EV_ADD|EV_DISPATCH";
+ if (evts[i].flags == (EV_ADD|EV_DISABLE)) flags = "EV_ADD|EV_DISABLE";
+ if (evts[i].flags == (EV_ENABLE|EV_DISPATCH)) flags = "EV_ENABLE|EV_DISPATCH";
+ if (evts[i].flags == EV_DISABLE) flags = "EV_DISABLE";
+ if (evts[i].flags == (EV_DISABLE|EV_DISPATCH)) flags = "EV_DISABLE|EV_DISABLE";
+#endif
+
+ keventbp += sprintf(keventbp, "%s{%lu, %s, %s}",i > 0 ? ", " : "",
+ evts[i].ident,
+ (evts[i].filter == EVFILT_READ
+ ? "EVFILT_READ"
+ : (evts[i].filter == EVFILT_WRITE
+ ? "EVFILT_WRITE"
+ : "UNKNOWN")), flags);
+ }
+ keventbp += sprintf(keventbp, "}, %d)", len);
+ DEBUG_PRINT_FD("%s", ps, fd, keventb);
+ errno = saved_errno;
+ }
+#endif
+ if (res < 0) {
+ if (op != ERTS_POLL_OP_DEL) {
+#ifdef EV_RECEIPT
+ struct kevent receipt_evts[2];
+ len = 0;
+ ERTS_EV_SET(&evts[len++], fd, EVFILT_WRITE, EV_DELETE|EV_RECEIPT, (void *) 0);
+ ERTS_EV_SET(&evts[len++], fd, EVFILT_READ, EV_DELETE|EV_RECEIPT, (void *) 0);
+ do {
+ res = kevent(ps->kp_fd, evts, len, receipt_evts, 2, &ts);
+ } while (res < 0 && errno == EINTR);
#else
- erts_printf("Doing update on fd=%d\n", fd);
+ ERTS_EV_SET(&evts[0], fd, EVFILT_WRITE, EV_DELETE, (void *) 0);
+ do {
+ res = kevent(ps->kp_fd, evts, 1, NULL, 0, &ts);
+ } while (res < 0 && errno == EINTR);
+ ERTS_EV_SET(&evts[0], fd, EVFILT_READ, EV_DELETE, (void *) 0);
+ do {
+ res = kevent(ps->kp_fd, evts, 1, NULL, 0, &ts);
+ } while (res < 0 && errno == EINTR);
+#endif
+ if (op == ERTS_POLL_OP_ADD)
+ erts_atomic_dec_nob(&ps->no_of_user_fds);
+ return ERTS_POLL_EV_NVAL;
+ }
+ return 0;
+ }
+ return events;
+}
+
+#endif /* ERTS_POLL_USE_KQUEUE */
+
+#if !ERTS_POLL_USE_CONCURRENT_UPDATE
+
+static ERTS_INLINE void
+init_batch_update(ErtsPollSet *ps, int len)
+{
+#if ERTS_POLL_USE_DEVPOLL
+ ASSERT(ps->poll_fds == NULL);
+ ps->poll_fds = erts_alloc(ERTS_ALC_T_TMP, sizeof(struct pollfd) * len);
+ ps->poll_fds_ix = 0;
#endif
+}
+
+static ERTS_INLINE void
+write_batch_update(ErtsPollSet *ps)
+{
+#if ERTS_POLL_USE_DEVPOLL
+ ssize_t wres;
+ char *buf = (char *) ps->poll_fds;
+ size_t buf_size = sizeof(struct pollfd)*ps->poll_fds_ix;
+
+ while (1) {
+ wres = write(ps->kp_fd, (void *) buf, buf_size);
+ if (wres < 0) {
+ if (errno == EINTR)
+ continue;
+ fatal_error("%s:%d:write_batch_buf(): "
+ "Failed to write to /dev/poll: "
+ "%s (%d)\n",
+ __FILE__, __LINE__,
+ erl_errno_id(errno), errno);
+ }
+#if ERTS_POLL_DEBUG_PRINT
+ {
+ int saved_errno = errno, i;
+ char devpollb[2048], *devpollbp = devpollb;
+ devpollbp += sprintf(devpollbp, "%d = devpoll(%d, {", wres, ps->kp_fd);
+ for (i = 0; i < wres / sizeof(struct pollfd); i++) {
+ if (devpollbp == devpollb)
+ devpollbp += sprintf(devpollbp, "%d = devpoll(%d, {", wres, ps->kp_fd);
+ devpollbp += sprintf(devpollbp, "%s{fd = %d, events = %s}",
+ i > 0 ? ", " : "",
+ ps->poll_fds[i].fd,
+ ev2str(ps->poll_fds[i].events));
+ if (devpollbp - devpollb > 512) {
+ devpollbp += sprintf(devpollbp, "}, %d)", ps->poll_fds_ix);
+ DEBUG_PRINT("%s", ps, devpollb);
+ devpollbp = devpollb;
+ }
+ }
+ devpollbp += sprintf(devpollbp, "}, %d)", ps->poll_fds_ix);
+ DEBUG_PRINT("%s", ps, devpollb);
+ errno = saved_errno;
+ }
#endif
- ASSERT(fd < ps->fds_status_len);
-#if ERTS_POLL_USE_FALLBACK
- ASSERT(ps->fds_status[fd].used_events
- ? (ps->fds_status[fd].flags & ERTS_POLL_FD_FLG_INFLBCK)
- : (ps->fds_status[fd].flags & ERTS_POLL_FD_FLG_USEFLBCK));
+ buf_size -= wres;
+ if (buf_size <= 0)
+ break;
+ buf += wres;
+ }
+
+ if (buf_size < 0) {
+ fatal_error("%s:%d:write_devpoll_buf(): Internal error\n",
+ __FILE__, __LINE__);
+ }
+ erts_free(ERTS_ALC_T_TMP, ps->poll_fds);
+ ps->poll_fds = NULL;
#endif
+}
- if (!need_update(ps, fd))
- return 0;
+static ERTS_INLINE int
+need_update(ErtsPollSet *ps, int fd, int *resetp)
+{
+ int reset;
+ ASSERT(fd < ps->fds_status_len);
-#if ERTS_POLL_USE_FALLBACK
+ reset = (int) (ps->fds_status[fd].flags & ERTS_POLL_FD_FLG_RST);
ps->fds_status[fd].flags &= ~ERTS_POLL_FD_FLG_RST;
-#endif
+
+ *resetp = reset;
+
+ if (reset || ps->fds_status[fd].used_events != ps->fds_status[fd].events)
+ return 1;
+
+ return 0;
+}
+
+static int update_pollset(ErtsPollSet *ps, ErtsPollResFd pr[], int fd)
+{
+ int res = 0, reset = 0;
+ ErtsPollEvents events = ps->fds_status[fd].events;
+ ASSERT(fd < ps->fds_status_len);
+
+ if (!need_update(ps, fd, &reset))
+ return res;
#if ERTS_POLL_USE_POLL /* --- poll -------------------------------- */
- if (!ps->fds_status[fd].events) {
+ if (!events) {
int pix = ps->fds_status[fd].pix;
int last_pix;
+
+ if (reset) {
+ /* When a fd has been reset, we tell the caller of erts_poll_wait
+ this by setting the fd as ERTS_POLL_EV_NONE */
+ ERTS_POLL_RES_SET_FD(&pr[res], fd);
+ ERTS_POLL_RES_SET_EVTS(&pr[res], ERTS_POLL_EV_NONE);
+ DEBUG_PRINT_FD("trig %s (poll)", ps, fd, ev2str(ERTS_POLL_EV_NONE));
+ res++;
+ }
+
if (pix < 0) {
-#if ERTS_POLL_USE_FALLBACK
- ASSERT(!(ps->fds_status[fd].flags & ERTS_POLL_FD_FLG_INFLBCK));
-#endif
- return -1;
+ return res;
}
-#if ERTS_POLL_USE_FALLBACK
- ASSERT(ps->fds_status[fd].flags & ERTS_POLL_FD_FLG_INFLBCK);
-#endif
erts_atomic_dec_nob(&ps->no_of_user_fds);
last_pix = --ps->no_poll_fds;
if (pix != last_pix) {
@@ -1348,126 +1031,151 @@ static int update_pollset(ErtsPollSet ps, int fd)
/* Clear this fd status */
ps->fds_status[fd].pix = -1;
ps->fds_status[fd].used_events = (ErtsPollEvents) 0;
-#if ERTS_POLL_USE_FALLBACK
- ps->fds_status[fd].flags &= ~ERTS_POLL_FD_FLG_INFLBCK;
-#endif
+
}
else {
int pix = ps->fds_status[fd].pix;
if (pix < 0) {
-#if ERTS_POLL_USE_FALLBACK
- ASSERT(!(ps->fds_status[fd].flags & ERTS_POLL_FD_FLG_INFLBCK)
- || fd == ps->kp_fd);
-#endif
erts_atomic_inc_nob(&ps->no_of_user_fds);
ps->fds_status[fd].pix = pix = ps->no_poll_fds++;
if (pix >= ps->poll_fds_len)
grow_poll_fds(ps, pix);
ps->poll_fds[pix].fd = fd;
ps->fds_status[fd].pix = pix;
-#if ERTS_POLL_USE_FALLBACK
- ps->fds_status[fd].flags |= ERTS_POLL_FD_FLG_INFLBCK;
-#endif
}
-#if ERTS_POLL_USE_FALLBACK
- ASSERT(ps->fds_status[fd].flags & ERTS_POLL_FD_FLG_INFLBCK);
-#endif
-
/* Events to be used in next poll */
- ps->poll_fds[pix].events = ev2pollev(ps->fds_status[fd].events);
+ ps->poll_fds[pix].events = ev2pollev(events);
if (ps->poll_fds[pix].revents) {
/* Remove result events that we should not poll for anymore */
ps->poll_fds[pix].revents
&= ev2pollev(~(~ps->fds_status[fd].used_events
- & ps->fds_status[fd].events));
+ & events));
}
/* Save events to be used in next poll */
- ps->fds_status[fd].used_events = ps->fds_status[fd].events;
+ ps->fds_status[fd].used_events = events;
}
- return 0;
+ return res;
#elif ERTS_POLL_USE_SELECT /* --- select ------------------------------ */
- {
- ErtsPollEvents events = ps->fds_status[fd].events;
+ if (!events) {
+
+ if (reset) {
+ /* When a fd has been reset, we tell the caller of erts_poll_wait
+ this by setting the fd as ERTS_POLL_EV_NONE */
+ ERTS_POLL_RES_SET_FD(&pr[res], fd);
+ ERTS_POLL_RES_SET_EVTS(&pr[res], ERTS_POLL_EV_NONE);
+ DEBUG_PRINT_FD("trig %s (select)", ps, fd, ev2str(ERTS_POLL_EV_NONE));
+ res++;
+ }
+
+ ERTS_FD_CLR(fd, &ps->input_fds);
+ ERTS_FD_CLR(fd, &ps->output_fds);
+
+ if (ps->fds_status[fd].used_events) {
+ erts_atomic_dec_nob(&ps->no_of_user_fds);
+ ps->fds_status[fd].used_events = (ErtsPollEvents) 0;
+ }
+
+ if (fd == ps->max_fd) {
+ int max = ps->max_fd;
+ for (max = ps->max_fd; max >= 0; max--)
+ if (ps->fds_status[max].used_events)
+ break;
+ ps->max_fd = max;
+ }
+
+ } else {
+
ensure_select_fds(fd, &ps->input_fds, &ps->output_fds);
- if ((ERTS_POLL_EV_IN & events)
- != (ERTS_POLL_EV_IN & ps->fds_status[fd].used_events)) {
- if (ERTS_POLL_EV_IN & events) {
- ERTS_FD_SET(fd, &ps->input_fds);
- }
- else {
- ERTS_FD_CLR(fd, &ps->input_fds);
- }
- }
- if ((ERTS_POLL_EV_OUT & events)
- != (ERTS_POLL_EV_OUT & ps->fds_status[fd].used_events)) {
- if (ERTS_POLL_EV_OUT & events) {
- ERTS_FD_SET(fd, &ps->output_fds);
- }
- else {
- ERTS_FD_CLR(fd, &ps->output_fds);
- }
- }
- if (!ps->fds_status[fd].used_events) {
- ASSERT(events);
- erts_atomic_inc_nob(&ps->no_of_user_fds);
-#if ERTS_POLL_USE_FALLBACK
- ps->no_select_fds++;
- ps->fds_status[fd].flags |= ERTS_POLL_FD_FLG_INFLBCK;
-#endif
- }
- else if (!events) {
- ASSERT(ps->fds_status[fd].used_events);
- erts_atomic_dec_nob(&ps->no_of_user_fds);
- ps->fds_status[fd].events = events;
-#if ERTS_POLL_USE_FALLBACK
- ps->no_select_fds--;
- ps->fds_status[fd].flags &= ~ERTS_POLL_FD_FLG_INFLBCK;
-#endif
- }
+ if (!ps->fds_status[fd].used_events)
+ erts_atomic_inc_nob(&ps->no_of_user_fds);
+
+ if (events & ERTS_POLL_EV_IN)
+ ERTS_FD_SET(fd, &ps->input_fds);
+ else
+ ERTS_FD_CLR(fd, &ps->input_fds);
+
+ if (events & ERTS_POLL_EV_OUT)
+ ERTS_FD_SET(fd, &ps->output_fds);
+ else
+ ERTS_FD_CLR(fd, &ps->output_fds);
ps->fds_status[fd].used_events = events;
- if (events && fd > ps->max_fd)
- ps->max_fd = fd;
- else if (!events && fd == ps->max_fd) {
- int max = ps->max_fd;
- for (max = ps->max_fd; max >= 0; max--)
- if (ps->fds_status[max].used_events)
- break;
- ps->max_fd = max;
- }
+ if (fd > ps->max_fd)
+ ps->max_fd = fd;
}
- return 0;
-#endif
-}
-#endif /* ERTS_POLL_USE_POLL || ERTS_POLL_USE_SELECT || ERTS_POLL_USE_FALLBACK */
+ return res;
+#elif ERTS_POLL_USE_DEVPOLL
+ if (!events) {
-static void
-handle_update_requests(ErtsPollSet ps)
-{
- ErtsPollSetUpdateRequestsBlock *urqbp = &ps->update_requests;
-#if ERTS_POLL_USE_BATCH_UPDATE_POLLSET
- ErtsPollBatchBuf bb;
- setup_batch_buf(ps, &bb);
+ if (reset) {
+ /* When a fd has been reset, we tell the caller of erts_poll_wait
+ this by setting the fd as ERTS_POLL_EV_NONE */
+ ERTS_POLL_RES_SET_FD(&pr[res], fd);
+ ERTS_POLL_RES_SET_EVTS(&pr[res], ERTS_POLL_EV_NONE);
+ DEBUG_PRINT_FD("trig %s (devpoll)", ps, fd, ev2str(ERTS_POLL_EV_NONE));
+ res++;
+ }
+
+ ps->poll_fds[ps->poll_fds_ix].fd = fd;
+ ps->poll_fds[ps->poll_fds_ix].revents = 0;
+ ps->poll_fds[ps->poll_fds_ix++].events = POLLREMOVE;
+
+ if (ps->fds_status[fd].used_events) {
+ erts_atomic_dec_nob(&ps->no_of_user_fds);
+ ps->fds_status[fd].used_events = 0;
+ }
+
+ } else {
+ if (!ps->fds_status[fd].used_events) {
+ erts_atomic_inc_nob(&ps->no_of_user_fds);
+ }
+ ps->poll_fds[ps->poll_fds_ix].fd = fd;
+ ps->poll_fds[ps->poll_fds_ix].revents = 0;
+ ps->poll_fds[ps->poll_fds_ix++].events = ERTS_POLL_EV_E2N(events);
+ ps->fds_status[fd].used_events = ps->fds_status[fd].events;
+ }
+
+ return res;
#endif
+}
+
+static int
+handle_update_requests(ErtsPollSet *ps, ErtsPollResFd pr[], int no_fds)
+{
+ int res = 0;
+ ErtsPollSetUpdateRequestsBlock *urqbp = ps->curr_upd_req_block;
while (urqbp) {
ErtsPollSetUpdateRequestsBlock *free_urqbp = urqbp;
int i;
int len = urqbp->len;
+
+ init_batch_update(ps, len);
+
for (i = 0; i < len; i++) {
int fd = urqbp->fds[i];
ASSERT(fd < ps->fds_status_len);
- ps->fds_status[fd].flags &= ~ERTS_POLL_FD_FLG_INURQ;
-#if ERTS_POLL_USE_BATCH_UPDATE_POLLSET
- batch_update_pollset(ps, fd, &bb);
-#else
- update_pollset(ps, fd);
-#endif
+ ASSERT(ps->fds_status[fd].flags & ERTS_POLL_FD_FLG_INURQ);
+
+ /* We have run out of PollResFd slots to put results in,
+ so we yield here and return later for more. */
+ if (res == no_fds && pr != NULL) {
+ memmove(urqbp->fds, urqbp->fds+i, sizeof(int) * (len - i));
+ urqbp->len -= i;
+ ps->curr_upd_req_block = urqbp;
+ write_batch_update(ps);
+ return res;
+ }
+
+ if (ps->fds_status[fd].flags & ERTS_POLL_FD_FLG_INURQ) {
+ ps->fds_status[fd].flags &= ~ERTS_POLL_FD_FLG_INURQ;
+ res += update_pollset(ps, pr + res, fd);
+ }
}
free_urqbp = urqbp;
@@ -1475,12 +1183,9 @@ handle_update_requests(ErtsPollSet ps)
free_update_requests_block(ps, free_urqbp);
- }
+ write_batch_update(ps);
-#if ERTS_POLL_USE_BATCH_UPDATE_POLLSET
- if (bb.len)
- write_batch_buf(ps, &bb);
-#endif
+ }
ps->curr_upd_req_block = &ps->update_requests;
@@ -1489,16 +1194,19 @@ handle_update_requests(ErtsPollSet ps)
#endif
ERTS_POLLSET_UNSET_HAVE_UPDATE_REQUESTS(ps);
+ return res;
}
+#endif /* !ERTS_POLL_USE_CONCURRENT_UPDATE */
static ERTS_INLINE ErtsPollEvents
-poll_control(ErtsPollSet ps, int fd, ErtsPollEvents events, int on, int *do_wake)
+poll_control(ErtsPollSet *ps, int fd, ErtsPollOp op,
+ ErtsPollEvents events, int *do_wake)
{
ErtsPollEvents new_events;
if (fd < ps->internal_fd_limit || fd >= max_fds) {
- if (fd < 0) {
+ if (fd < 0 || fd >= max_fds) {
new_events = ERTS_POLL_EV_ERR;
goto done;
}
@@ -1508,115 +1216,55 @@ poll_control(ErtsPollSet ps, int fd, ErtsPollEvents events, int on, int *do_wake
goto done;
}
#endif
+#if !ERTS_POLL_USE_CONCURRENT_UPDATE
if (fd == ps->wake_fds[0] || fd == ps->wake_fds[1]) {
new_events = ERTS_POLL_EV_NVAL;
goto done;
}
-#if ERTS_POLL_USE_TIMERFD
- if (fd == ps->timer_fd) {
- new_events = ERTS_POLL_EV_NVAL;
- goto done;
- }
#endif
}
+#if ERTS_POLL_USE_CONCURRENT_UPDATE
+
+ new_events = update_pollset(ps, fd, op, events);
+
+#else /* !ERTS_POLL_USE_CONCURRENT_UPDATE */
if (fd >= ps->fds_status_len)
grow_fds_status(ps, fd);
ASSERT(fd < ps->fds_status_len);
- new_events = ps->fds_status[fd].events;
-
- if (events == 0) {
- *do_wake = 0;
- goto done;
- }
-
- if (on)
- new_events |= events;
- else
- new_events &= ~events;
-
- if (new_events == (ErtsPollEvents) 0) {
- ps->fds_status[fd].flags |= ERTS_POLL_FD_FLG_RST;
-#if ERTS_POLL_USE_FALLBACK
- ps->fds_status[fd].flags &= ~ERTS_POLL_FD_FLG_USEFLBCK;
-#endif
- }
-
- ps->fds_status[fd].events = new_events;
-
- if (new_events == ps->fds_status[fd].used_events
- && !(ps->fds_status[fd].flags & ERTS_POLL_FD_FLG_RST)
- ) {
- *do_wake = 0;
- goto done;
- }
-
-
-#if ERTS_POLL_USE_CONCURRENT_UPDATE
- if (ERTS_POLLSET_IS_POLLED(ps)) {
- int update_fallback = 0;
- conc_update_pollset(ps, fd, &update_fallback);
- if (!update_fallback) {
- *do_wake = 0; /* no need to wake kernel poller */
- goto done;
- }
+ if (op == ERTS_POLL_OP_DEL) {
+ ps->fds_status[fd].flags |= ERTS_POLL_FD_FLG_RST;
+ ps->fds_status[fd].events = 0;
+ *do_wake = 1;
+ } else if (op == ERTS_POLL_OP_ADD) {
+ ASSERT(ps->fds_status[fd].events == 0);
+ ps->fds_status[fd].events = events;
+ *do_wake = 1;
+ } else {
+ ASSERT(op == ERTS_POLL_OP_MOD);
+ ps->fds_status[fd].events = events;
+ *do_wake = 1;
}
-#endif
+ new_events = ps->fds_status[fd].events;
enqueue_update_request(ps, fd);
-
- /*
- * If new events have been added, we need to wake up the
- * polling thread, but if events have been removed we don't.
- */
- if ((new_events && (ps->fds_status[fd].flags & ERTS_POLL_FD_FLG_RST))
- || (~ps->fds_status[fd].used_events & new_events))
- *do_wake = 1;
+#endif /* !ERTS_POLL_USE_CONCURRENT_UPDATE */
done:
-#ifdef ERTS_POLL_DEBUG_PRINT
- erts_printf("0x%x = poll_control(ps, %d, 0x%x, %s) do_wake=%d\n",
- (int) new_events, fd, (int) events, (on ? "on" : "off"), *do_wake);
-#endif
+ DEBUG_PRINT_FD("%s = %s(%p, %d, %s, %s) do_wake=%d",
+ ps, fd, ev2str(new_events), __FUNCTION__, ps,
+ fd, op2str(op), ev2str(events), *do_wake);
return new_events;
}
-void
-ERTS_POLL_EXPORT(erts_poll_controlv)(ErtsPollSet ps,
- ErtsPollControlEntry pcev[],
- int len)
-{
- int i;
- int do_wake;
- int final_do_wake = 0;
-
- ERTS_POLLSET_LOCK(ps);
-
- for (i = 0; i < len; i++) {
- do_wake = 0;
- pcev[i].events = poll_control(ps,
- pcev[i].fd,
- pcev[i].events,
- pcev[i].on,
- &do_wake);
- final_do_wake |= do_wake;
- }
-
- ERTS_POLLSET_UNLOCK(ps);
-
- if (final_do_wake)
- wake_poller(ps, 0, 0);
-
-}
-
ErtsPollEvents
-ERTS_POLL_EXPORT(erts_poll_control)(ErtsPollSet ps,
+ERTS_POLL_EXPORT(erts_poll_control)(ErtsPollSet *ps,
ErtsSysFdType fd,
+ ErtsPollOp op,
ErtsPollEvents events,
- int on,
int* do_wake) /* In: Wake up polling thread */
/* Out: Poller is woken */
{
@@ -1624,13 +1272,15 @@ ERTS_POLL_EXPORT(erts_poll_control)(ErtsPollSet ps,
ERTS_POLLSET_LOCK(ps);
- res = poll_control(ps, fd, events, on, do_wake);
+ res = poll_control(ps, fd, op, events, do_wake);
ERTS_POLLSET_UNLOCK(ps);
+#if !ERTS_POLL_USE_CONCURRENT_UPDATE
if (*do_wake) {
- wake_poller(ps, 0, 0);
+ wake_poller(ps, 0);
}
+#endif
return res;
}
@@ -1642,180 +1292,60 @@ ERTS_POLL_EXPORT(erts_poll_control)(ErtsPollSet ps,
#if ERTS_POLL_USE_KERNEL_POLL
static ERTS_INLINE int
-save_kp_result(ErtsPollSet ps, ErtsPollResFd pr[], int max_res, int chk_fds_res)
+ERTS_POLL_EXPORT(save_result)(ErtsPollSet *ps, ErtsPollResFd pr[], int max_res, int chk_fds_res, int ebadf)
{
- int res = 0;
- int i;
- int n = chk_fds_res < max_res ? chk_fds_res : max_res;
+#if !ERTS_POLL_USE_CONCURRENT_UPDATE || ERTS_POLL_DEBUG_PRINT
+ int n = chk_fds_res < max_res ? chk_fds_res : max_res, i;
+ int res = n;
+#if !ERTS_POLL_USE_CONCURRENT_UPDATE
int wake_fd = ps->wake_fds[0];
-#if ERTS_POLL_USE_TIMERFD
- int timer_fd = ps->timer_fd;
#endif
for (i = 0; i < n; i++) {
-
-#if ERTS_POLL_USE_EPOLL /* --- epoll ------------------------------- */
-
- if (ps->res_events[i].events) {
- int fd = ps->res_events[i].data.fd;
- int ix;
- ErtsPollEvents revents;
- if (fd == wake_fd) {
- cleanup_wakeup_pipe(ps);
- continue;
- }
-#if ERTS_POLL_USE_TIMERFD
- if (fd == timer_fd) {
- continue;
- }
+ int fd = ERTS_POLL_RES_GET_FD(&pr[i]);
+#ifdef DEBUG_PRINT_MODE
+ ErtsPollEvents evts = ERTS_POLL_RES_GET_EVTS(pr+i);
#endif
- ASSERT(!(ps->fds_status[fd].flags & ERTS_POLL_FD_FLG_INFLBCK));
- /* epoll_wait() can repeat the same fd in result array... */
- ix = (int) ps->fds_status[fd].res_ev_ix;
- ASSERT(ix >= 0);
- if (ix >= res || pr[ix].fd != fd) {
- ix = res;
- pr[ix].fd = fd;
- pr[ix].events = (ErtsPollEvents) 0;
- }
-
- revents = ERTS_POLL_EV_N2E(ps->res_events[i].events);
- pr[ix].events |= revents;
- if (revents) {
- if (res == ix) {
- ps->fds_status[fd].res_ev_ix = (unsigned short) ix;
- res++;
- }
- }
- }
-
-#elif ERTS_POLL_USE_KQUEUE /* --- kqueue ------------------------------ */
-
- struct kevent *ev;
- int fd;
- int ix;
-
- ev = &ps->res_events[i];
- fd = (int) ev->ident;
- ASSERT(fd < ps->fds_status_len);
- ASSERT(!(ps->fds_status[fd].flags & ERTS_POLL_FD_FLG_INFLBCK));
- ix = (int) ps->fds_status[fd].res_ev_ix;
-
- ASSERT(ix >= 0);
- if (ix >= res || pr[ix].fd != fd) {
- ix = res;
- pr[ix].fd = (int) ev->ident;
- pr[ix].events = (ErtsPollEvents) 0;
- }
-
- if (ev->filter == EVFILT_READ) {
- if (fd == wake_fd) {
- cleanup_wakeup_pipe(ps);
- continue;
- }
- pr[ix].events |= ERTS_POLL_EV_IN;
- }
- else if (ev->filter == EVFILT_WRITE)
- pr[ix].events |= ERTS_POLL_EV_OUT;
- if (ev->flags & (EV_ERROR|EV_EOF)) {
- if ((ev->flags & EV_ERROR) && (((int) ev->data) == EBADF))
- pr[ix].events |= ERTS_POLL_EV_NVAL;
- else
- pr[ix].events |= ERTS_POLL_EV_ERR;
- }
- if (pr[ix].events) {
- if (res == ix) {
- ps->fds_status[fd].res_ev_ix = (unsigned short) ix;
- res++;
- }
- }
-
-#elif ERTS_POLL_USE_DEVPOLL /* --- devpoll ----------------------------- */
-
- if (ps->res_events[i].revents) {
- int fd = ps->res_events[i].fd;
- ErtsPollEvents revents;
- if (fd == wake_fd) {
- cleanup_wakeup_pipe(ps);
- continue;
- }
-#if ERTS_POLL_USE_TIMERFD
- if (fd == timer_fd) {
- continue;
- }
-#endif
- revents = ERTS_POLL_EV_N2E(ps->res_events[i].events);
- pr[res].fd = fd;
- pr[res].events = revents;
- res++;
- }
-#endif
-
- }
-
- return res;
-}
-
-#endif /* ERTS_POLL_USE_KERNEL_POLL */
-
-#if ERTS_POLL_USE_FALLBACK
-
-static int
-get_kp_results(ErtsPollSet ps, ErtsPollResFd pr[], int max_res)
-{
- int res;
+ DEBUG_PRINT_FD("trig %s (%s)", ps, fd,
+ ev2str(evts),
#if ERTS_POLL_USE_KQUEUE
- struct timespec ts = {0, 0};
+ "kqueue"
+#elif ERTS_POLL_USE_EPOLL
+ "epoll"
+#else
+ "/dev/poll"
#endif
+ );
- if (max_res > ps->res_events_len)
- grow_res_events(ps, max_res);
+#if !ERTS_POLL_USE_CONCURRENT_UPDATE
- do {
-#if ERTS_POLL_USE_EPOLL
- res = epoll_wait(ps->kp_fd, ps->res_events, max_res, 0);
-#elif ERTS_POLL_USE_KQUEUE
- res = kevent(ps->kp_fd, NULL, 0, ps->res_events, max_res, &ts);
+ if (fd == wake_fd) {
+ cleanup_wakeup_pipe(ps);
+ ERTS_POLL_RES_SET_EVTS(&pr[i], ERTS_POLL_EV_NONE);
+ } else {
+ /* Reset the events to emulate ONESHOT semantics */
+ ps->fds_status[fd].events = 0;
+ enqueue_update_request(ps, fd);
+ }
#endif
- } while (res < 0 && errno == EINTR);
-
- if (res < 0) {
- fatal_error("%s:%d: %s() failed: %s (%d)\n",
- __FILE__, __LINE__,
-#if ERTS_POLL_USE_EPOLL
- "epoll_wait",
-#elif ERTS_POLL_USE_KQUEUE
- "kevent",
-#endif
- erl_errno_id(errno), errno);
}
- return save_kp_result(ps, pr, max_res, res);
+ return res;
+#else
+ ASSERT(chk_fds_res <= max_res);
+ return chk_fds_res;
+#endif
}
-#endif /* ERTS_POLL_USE_FALLBACK */
-
-
+#else /* !ERTS_POLL_USE_KERNEL_POLL */
static ERTS_INLINE int
-save_poll_result(ErtsPollSet ps, ErtsPollResFd pr[], int max_res,
- int chk_fds_res, int ebadf)
+ERTS_POLL_EXPORT(save_result)(ErtsPollSet *ps, ErtsPollResFd pr[], int max_res, int chk_fds_res, int ebadf)
{
-#if ERTS_POLL_USE_DEVPOLL
- return save_kp_result(ps, pr, max_res, chk_fds_res);
-#elif ERTS_POLL_USE_FALLBACK
- if (!ps->fallback_used)
- return save_kp_result(ps, pr, max_res, chk_fds_res);
- else
-#endif /* ERTS_POLL_USE_FALLBACK */
- {
-
#if ERTS_POLL_USE_POLL /* --- poll -------------------------------- */
int res = 0;
-#if !ERTS_POLL_USE_FALLBACK
int wake_fd = ps->wake_fds[0];
-#endif
int i, first_ix, end_ix;
/*
@@ -1832,23 +1362,30 @@ save_poll_result(ErtsPollSet ps, ErtsPollResFd pr[], int max_res,
if (ps->poll_fds[i].revents != (short) 0) {
int fd = ps->poll_fds[i].fd;
ErtsPollEvents revents;
-#if ERTS_POLL_USE_FALLBACK
- if (fd == ps->kp_fd) {
- res += get_kp_results(ps, &pr[res], max_res-res);
- i++;
- continue;
- }
-#else
if (fd == wake_fd) {
cleanup_wakeup_pipe(ps);
i++;
continue;
}
-#endif
revents = pollev2ev(ps->poll_fds[i].revents);
- pr[res].fd = fd;
- pr[res].events = revents;
+ ERTS_POLL_RES_SET_FD(&pr[res], fd);
+ ERTS_POLL_RES_SET_EVTS(&pr[res], revents);
+
+ /* If an fd returns as error, we may want to check the
+ update_requests queue to see if it has been reset
+ before delivering the result?!?! This should allow
+ the user to do driver_dselect + close without waiting
+ for stop_select... */
+
+ DEBUG_PRINT_FD("trig %s (poll)", ps, ERTS_POLL_RES_GET_FD(&pr[res]),
+ ev2str(ERTS_POLL_RES_GET_EVTS(&pr[res])));
+
res++;
+
+ /* Clear the events for this fd in order to mimic
+ how epoll ONESHOT works */
+ ps->fds_status[fd].events = 0;
+ enqueue_update_request(ps, fd);
}
i++;
}
@@ -1864,9 +1401,7 @@ save_poll_result(ErtsPollSet ps, ErtsPollResFd pr[], int max_res,
#elif ERTS_POLL_USE_SELECT /* --- select ------------------------------ */
int res = 0;
-#if !ERTS_POLL_USE_FALLBACK
int wake_fd = ps->wake_fds[0];
-#endif
int fd, first_fd, end_fd;
/*
@@ -1879,29 +1414,23 @@ save_poll_result(ErtsPollSet ps, ErtsPollResFd pr[], int max_res,
if (!ebadf) {
while (1) {
while (fd < end_fd && res < max_res) {
-
- pr[res].events = (ErtsPollEvents) 0;
+ ErtsPollEvents events = 0;
if (ERTS_FD_ISSET(fd, &ps->res_input_fds)) {
-#if ERTS_POLL_USE_FALLBACK
- if (fd == ps->kp_fd) {
- res += get_kp_results(ps, &pr[res], max_res-res);
- fd++;
- continue;
- }
-#else
if (fd == wake_fd) {
cleanup_wakeup_pipe(ps);
fd++;
continue;
}
-#endif
- pr[res].events |= ERTS_POLL_EV_IN;
+ events |= ERTS_POLL_EV_IN;
}
if (ERTS_FD_ISSET(fd, &ps->res_output_fds))
- pr[res].events |= ERTS_POLL_EV_OUT;
- if (pr[res].events) {
- pr[res].fd = fd;
+ events |= ERTS_POLL_EV_OUT;
+ if (events) {
+ ERTS_POLL_RES_SET_FD(&pr[res], fd);
+ ERTS_POLL_RES_SET_EVTS(&pr[res], events);
res++;
+ ps->fds_status[fd].events = 0;
+ enqueue_update_request(ps, fd);
}
fd++;
}
@@ -1934,7 +1463,7 @@ save_poll_result(ErtsPollSet ps, ErtsPollResFd pr[], int max_res,
if (ps->fds_status[fd].events & ERTS_POLL_EV_OUT) {
oset = &ps->res_output_fds;
ERTS_FD_ZERO(oset);
- ERTS_FD_SET(fd, oset);
+ ERTS_FD_SET(fd, oset);
}
do {
/* Initiate 'tv' each time;
@@ -1943,49 +1472,31 @@ save_poll_result(ErtsPollSet ps, ErtsPollResFd pr[], int max_res,
sres = ERTS_SELECT(ps->max_fd+1, iset, oset, NULL, &tv);
} while (sres < 0 && errno == EINTR);
if (sres < 0) {
-#if ERTS_POLL_USE_FALLBACK
- if (fd == ps->kp_fd) {
- res += get_kp_results(ps,
- &pr[res],
- max_res-res);
- fd++;
- continue;
- }
-#else
if (fd == wake_fd) {
cleanup_wakeup_pipe(ps);
fd++;
continue;
}
-#endif
- pr[res].fd = fd;
- pr[res].events = ERTS_POLL_EV_NVAL;
+ ERTS_POLL_RES_SET_FD(&pr[res], fd);
+ ERTS_POLL_RES_SET_EVTS(&pr[res], ERTS_POLL_EV_NVAL);
res++;
}
else if (sres > 0) {
- pr[res].fd = fd;
+ ErtsPollEvents events = 0;
+ ERTS_POLL_RES_SET_FD(&pr[res], fd);
if (iset && ERTS_FD_ISSET(fd, iset)) {
-#if ERTS_POLL_USE_FALLBACK
- if (fd == ps->kp_fd) {
- res += get_kp_results(ps,
- &pr[res],
- max_res-res);
- fd++;
- continue;
- }
-#else
if (fd == wake_fd) {
cleanup_wakeup_pipe(ps);
fd++;
continue;
}
-#endif
- pr[res].events |= ERTS_POLL_EV_IN;
+ events |= ERTS_POLL_EV_IN;
}
if (oset && ERTS_FD_ISSET(fd, oset)) {
- pr[res].events |= ERTS_POLL_EV_OUT;
+ events |= ERTS_POLL_EV_OUT;
}
- ASSERT(pr[res].events);
+ ASSERT(events);
+ ERTS_POLL_RES_SET_EVTS(&pr[res], events);
res++;
}
}
@@ -2000,353 +1511,145 @@ save_poll_result(ErtsPollSet ps, ErtsPollResFd pr[], int max_res,
}
ps->next_sel_fd = fd;
return res;
-#endif
- }
+#endif /* ERTS_POLL_USE_SELECT */
}
-static ERTS_INLINE ErtsMonotonicTime
-get_timeout(ErtsPollSet ps,
- int resolution,
- ErtsMonotonicTime timeout_time)
-{
- ErtsMonotonicTime timeout, save_timeout_time;
-
- if (timeout_time == ERTS_POLL_NO_TIMEOUT) {
- save_timeout_time = ERTS_MONOTONIC_TIME_MIN;
- timeout = 0;
- }
- else {
- ErtsMonotonicTime diff_time, current_time;
- current_time = erts_get_monotonic_time(NULL);
- diff_time = timeout_time - current_time;
- if (diff_time <= 0) {
- save_timeout_time = ERTS_MONOTONIC_TIME_MIN;
- timeout = 0;
- }
- else {
- save_timeout_time = current_time;
- switch (resolution) {
- case 1000:
- /* Round up to nearest even milli second */
- timeout = ERTS_MONOTONIC_TO_MSEC(diff_time - 1) + 1;
- if (timeout > (ErtsMonotonicTime) INT_MAX)
- timeout = (ErtsMonotonicTime) INT_MAX;
- save_timeout_time += ERTS_MSEC_TO_MONOTONIC(timeout);
- timeout -= ERTS_PREMATURE_TIMEOUT(timeout, 1000);
- break;
- case 1000000:
- /* Round up to nearest even micro second */
- timeout = ERTS_MONOTONIC_TO_USEC(diff_time - 1) + 1;
- save_timeout_time += ERTS_USEC_TO_MONOTONIC(timeout);
- timeout -= ERTS_PREMATURE_TIMEOUT(timeout, 1000*1000);
- break;
- case 1000000000:
- /* Round up to nearest even nano second */
- timeout = ERTS_MONOTONIC_TO_NSEC(diff_time - 1) + 1;
- save_timeout_time += ERTS_NSEC_TO_MONOTONIC(timeout);
- timeout -= ERTS_PREMATURE_TIMEOUT(timeout, 1000*1000*1000);
- break;
- default:
- ERTS_INTERNAL_ERROR("Invalid resolution");
- timeout = 0;
- save_timeout_time = 0;
- break;
- }
- }
- }
- set_timeout_time(ps, save_timeout_time);
- return timeout;
-}
-
-#if ERTS_POLL_USE_SELECT
+#endif /* !ERTS_POLL_USE_KERNEL_POLL */
static ERTS_INLINE int
-get_timeout_timeval(ErtsPollSet ps,
- SysTimeval *tvp,
- ErtsMonotonicTime timeout_time)
-{
- ErtsMonotonicTime timeout = get_timeout(ps,
- 1000*1000,
- timeout_time);
-
- if (!timeout) {
- tvp->tv_sec = 0;
- tvp->tv_usec = 0;
-
- return 0;
- }
- else {
- ErtsMonotonicTime sec = timeout/(1000*1000);
- tvp->tv_sec = sec;
- tvp->tv_usec = timeout - sec*(1000*1000);
-
- ASSERT(tvp->tv_sec >= 0);
- ASSERT(tvp->tv_usec >= 0);
- ASSERT(tvp->tv_usec < 1000*1000);
-
- return !0;
- }
-
-}
-
-#endif
-
-#if ERTS_POLL_USE_KQUEUE || (ERTS_POLL_USE_POLL && defined(HAVE_PPOLL)) || ERTS_POLL_USE_TIMERFD
-
-static ERTS_INLINE int
-get_timeout_timespec(ErtsPollSet ps,
- struct timespec *tsp,
- ErtsMonotonicTime timeout_time)
-{
- ErtsMonotonicTime timeout = get_timeout(ps,
- 1000*1000*1000,
- timeout_time);
-
- if (!timeout) {
- tsp->tv_sec = 0;
- tsp->tv_nsec = 0;
- return 0;
- }
- else {
- ErtsMonotonicTime sec = timeout/(1000*1000*1000);
- tsp->tv_sec = sec;
- tsp->tv_nsec = timeout - sec*(1000*1000*1000);
-
- ASSERT(tsp->tv_sec >= 0);
- ASSERT(tsp->tv_nsec >= 0);
- ASSERT(tsp->tv_nsec < 1000*1000*1000);
-
- return !0;
- }
-}
-
-#endif
-
-#if ERTS_POLL_USE_TIMERFD
-
-static ERTS_INLINE int
-get_timeout_itimerspec(ErtsPollSet ps,
- struct itimerspec *itsp,
- ErtsMonotonicTime timeout_time)
-{
-
- itsp->it_interval.tv_sec = 0;
- itsp->it_interval.tv_nsec = 0;
-
- return get_timeout_timespec(ps, &itsp->it_value, timeout_time);
-}
-
-#endif
-
-static ERTS_INLINE int
-check_fd_events(ErtsPollSet ps, ErtsMonotonicTime timeout_time, int max_res)
+check_fd_events(ErtsPollSet *ps, ErtsPollResFd pr[], int do_wait, int max_res)
{
int res;
- ERTS_MSACC_PUSH_STATE_M();
- if (erts_atomic_read_nob(&ps->no_of_user_fds) == 0
- && timeout_time == ERTS_POLL_NO_TIMEOUT) {
- /* Nothing to poll and zero timeout; done... */
- return 0;
- }
- else {
- int timeout;
-#if ERTS_POLL_USE_FALLBACK
- if (!(ps->fallback_used = ERTS_POLL_NEED_FALLBACK(ps))) {
-
-#if ERTS_POLL_USE_EPOLL /* --- epoll ------------------------------- */
- if (max_res > ps->res_events_len)
- grow_res_events(ps, max_res);
-#if ERTS_POLL_USE_TIMERFD
- {
- struct itimerspec its;
- timeout = get_timeout_itimerspec(ps, &its, timeout_time);
- if (timeout) {
- erts_thr_progress_prepare_wait(NULL);
- ERTS_MSACC_SET_STATE_CACHED_M(ERTS_MSACC_STATE_SLEEP);
- timerfd_set(ps, &its);
- res = epoll_wait(ps->kp_fd, ps->res_events, max_res, -1);
- res = timerfd_clear(ps, res, max_res);
- } else {
- res = epoll_wait(ps->kp_fd, ps->res_events, max_res, 0);
- }
- }
-#else /* !ERTS_POLL_USE_TIMERFD */
- timeout = (int) get_timeout(ps, 1000, timeout_time);
- if (timeout) {
- erts_thr_progress_prepare_wait(NULL);
- ERTS_MSACC_SET_STATE_CACHED_M(ERTS_MSACC_STATE_SLEEP);
- }
- res = epoll_wait(ps->kp_fd, ps->res_events, max_res, timeout);
-#endif /* !ERTS_POLL_USE_TIMERFD */
-#elif ERTS_POLL_USE_KQUEUE /* --- kqueue ------------------------------ */
- struct timespec ts;
- if (max_res > ps->res_events_len)
- grow_res_events(ps, max_res);
- timeout = get_timeout_timespec(ps, &ts, timeout_time);
- if (timeout) {
- erts_thr_progress_prepare_wait(NULL);
- ERTS_MSACC_SET_STATE_CACHED_M(ERTS_MSACC_STATE_SLEEP);
- }
- res = kevent(ps->kp_fd, NULL, 0, ps->res_events, max_res, &ts);
-#endif /* ----------------------------------------- */
- }
- else /* use fallback (i.e. poll() or select()) */
-#endif /* ERTS_POLL_USE_FALLBACK */
- {
+ int timeout = do_wait ? -1 : 0;
+ DEBUG_PRINT_WAIT("Entering check_fd_events(), do_wait=%d", ps, do_wait);
+ {
+#if ERTS_POLL_USE_EPOLL /* --- epoll ------------------------------- */
+ res = epoll_wait(ps->kp_fd, pr, max_res, timeout);
+
+#elif ERTS_POLL_USE_KQUEUE /* --- kqueue ------------------------------ */
+ struct timespec ts = {0, 0};
+ struct timespec *tsp = timeout ? NULL : &ts;
+ res = kevent(ps->kp_fd, NULL, 0, pr, max_res, tsp);
+#elif ERTS_POLL_USE_DEVPOLL /* --- devpoll ----------------------------- */
+ /*
+ * The ioctl() will fail with EINVAL on Solaris 10 if dp_nfds
+ * is set too high. dp_nfds should not be set greater than
+ * the maximum number of file descriptors in the poll set.
+ */
+ struct dvpoll poll_res;
+ int nfds = (int) erts_atomic_read_nob(&ps->no_of_user_fds) + 1 /* wakeup pipe */;
+ poll_res.dp_nfds = nfds < max_res ? nfds : max_res;
+ poll_res.dp_fds = pr;
+ poll_res.dp_timeout = timeout;
+ res = ioctl(ps->kp_fd, DP_POLL, &poll_res);
-#if ERTS_POLL_USE_DEVPOLL /* --- devpoll ----------------------------- */
- /*
- * The ioctl() will fail with EINVAL on Solaris 10 if dp_nfds
- * is set too high. dp_nfds should not be set greater than
- * the maximum number of file descriptors in the poll set.
- */
- struct dvpoll poll_res;
- int nfds = (int) erts_atomic_read_nob(&ps->no_of_user_fds);
- nfds++; /* Wakeup pipe */
- timeout = (int) get_timeout(ps, 1000, timeout_time);
- poll_res.dp_nfds = nfds < max_res ? nfds : max_res;
- if (poll_res.dp_nfds > ps->res_events_len)
- grow_res_events(ps, poll_res.dp_nfds);
- poll_res.dp_fds = ps->res_events;
- if (timeout) {
- erts_thr_progress_prepare_wait(NULL);
- ERTS_MSACC_SET_STATE_CACHED_M(ERTS_MSACC_STATE_SLEEP);
- }
- poll_res.dp_timeout = timeout;
- res = ioctl(ps->kp_fd, DP_POLL, &poll_res);
-#elif ERTS_POLL_USE_POLL && defined(HAVE_PPOLL) /* --- ppoll ---------------- */
- struct timespec ts;
- timeout = get_timeout_timespec(ps, &ts, timeout_time);
- if (timeout) {
- erts_thr_progress_prepare_wait(NULL);
- ERTS_MSACC_SET_STATE_CACHED_M(ERTS_MSACC_STATE_SLEEP);
- }
- res = ppoll(ps->poll_fds, ps->no_poll_fds, &ts, NULL);
#elif ERTS_POLL_USE_POLL /* --- poll --------------------------------- */
- timeout = (int) get_timeout(ps, 1000, timeout_time);
- if (timeout) {
- erts_thr_progress_prepare_wait(NULL);
- ERTS_MSACC_SET_STATE_CACHED_M(ERTS_MSACC_STATE_SLEEP);
- }
- res = poll(ps->poll_fds, ps->no_poll_fds, timeout);
+ res = poll(ps->poll_fds, ps->no_poll_fds, timeout);
+
#elif ERTS_POLL_USE_SELECT /* --- select ------------------------------ */
- SysTimeval to;
- timeout = get_timeout_timeval(ps, &to, timeout_time);
+ SysTimeval tv = {0, 0};
+ SysTimeval *tvp = timeout ? NULL : &tv;
- ERTS_FD_COPY(&ps->input_fds, &ps->res_input_fds);
- ERTS_FD_COPY(&ps->output_fds, &ps->res_output_fds);
+ ERTS_FD_COPY(&ps->input_fds, &ps->res_input_fds);
+ ERTS_FD_COPY(&ps->output_fds, &ps->res_output_fds);
- if (timeout) {
- erts_thr_progress_prepare_wait(NULL);
- ERTS_MSACC_SET_STATE_CACHED_M(ERTS_MSACC_STATE_SLEEP);
- }
- res = ERTS_SELECT(ps->max_fd + 1,
- &ps->res_input_fds,
- &ps->res_output_fds,
- NULL,
- &to);
- if (timeout) {
- erts_thr_progress_finalize_wait(NULL);
- ERTS_MSACC_POP_STATE_M();
- }
- if (res < 0
- && errno == EBADF
- && ERTS_POLLSET_HAVE_UPDATE_REQUESTS(ps)) {
- /*
- * This may have happened because another thread deselected
- * a fd in our poll set and then closed it, i.e. the driver
- * behaved correctly. We wan't to avoid looking for a bad
- * fd, that may even not exist anymore. Therefore, handle
- * update requests and try again.
- *
- * We don't know how much of the timeout is left; therfore,
- * we use a zero timeout. If no error occur and no events
- * have triggered, we fake an EAGAIN error and let the caller
- * restart us.
- */
- to.tv_sec = 0;
- to.tv_usec = 0;
- ERTS_POLLSET_LOCK(ps);
- handle_update_requests(ps);
- ERTS_POLLSET_UNLOCK(ps);
- res = ERTS_SELECT(ps->max_fd + 1,
- &ps->res_input_fds,
- &ps->res_output_fds,
- NULL,
- &to);
- if (res == 0) {
- errno = EAGAIN;
- res = -1;
- }
- }
- return res;
+ res = ERTS_SELECT(ps->max_fd + 1,
+ &ps->res_input_fds,
+ &ps->res_output_fds,
+ NULL,
+ tvp);
#endif /* ----------------------------------------- */
- }
- if (timeout) {
- erts_thr_progress_finalize_wait(NULL);
- ERTS_MSACC_POP_STATE_M();
- }
- return res;
}
+ DEBUG_PRINT_WAIT("Leaving check_fd_events(), res=%d", ps, res);
+ return res;
}
int
-ERTS_POLL_EXPORT(erts_poll_wait)(ErtsPollSet ps,
+ERTS_POLL_EXPORT(erts_poll_wait)(ErtsPollSet *ps,
ErtsPollResFd pr[],
- int *len,
- ErtsMonotonicTime timeout_time)
+ int *len)
{
- ErtsMonotonicTime to;
- int res, no_fds;
+ int res, no_fds, used_fds = 0;
int ebadf = 0;
+ int do_wait;
int ps_locked = 0;
+ ERTS_MSACC_DECLARE_CACHE();
no_fds = *len;
-#ifdef ERTS_POLL_MAX_RES
- if (no_fds >= ERTS_POLL_MAX_RES)
- no_fds = ERTS_POLL_MAX_RES;
-#endif
-
*len = 0;
+ ASSERT(no_fds > 0);
-#ifdef ERTS_POLL_DEBUG_PRINT
- erts_printf("Entering erts_poll_wait(), timeout_time=%bps\n",
- timeout_time);
-#endif
+#if !ERTS_POLL_USE_CONCURRENT_UPDATE
+ if (ERTS_POLLSET_HAVE_UPDATE_REQUESTS(ps)) {
+ ERTS_POLLSET_LOCK(ps);
+ used_fds = handle_update_requests(ps, pr, no_fds);
+ ERTS_POLLSET_UNLOCK(ps);
- if (ERTS_POLLSET_SET_POLLED_CHK(ps)) {
- res = EINVAL; /* Another thread is in erts_poll_wait()
- on this pollset... */
- goto done;
+ if (used_fds == no_fds) {
+ *len = used_fds;
+ return 0;
+ }
}
+#endif
- to = (is_woken(ps)
- ? ERTS_POLL_NO_TIMEOUT /* Use zero timeout */
- : timeout_time);
+ do_wait = !is_woken(ps) && used_fds == 0;
- if (ERTS_POLLSET_HAVE_UPDATE_REQUESTS(ps)) {
- ERTS_POLLSET_LOCK(ps);
- handle_update_requests(ps);
- ERTS_POLLSET_UNLOCK(ps);
+ DEBUG_PRINT_WAIT("Entering %s(), do_wait=%d", ps, __FUNCTION__, do_wait);
+
+ if (do_wait) {
+ erts_thr_progress_prepare_wait(NULL);
+ ERTS_MSACC_SET_STATE_CACHED(ERTS_MSACC_STATE_SLEEP);
}
while (1) {
- res = check_fd_events(ps, to, no_fds);
+ res = check_fd_events(ps, pr + used_fds, do_wait, no_fds - used_fds);
+
+#if !ERTS_POLL_USE_CONCURRENT_UPDATE
+ if (res < 0
+ && errno == EBADF
+ && ERTS_POLLSET_HAVE_UPDATE_REQUESTS(ps)) {
+ /*
+ * This may have happened because another thread deselected
+ * a fd in our poll set and then closed it, i.e. the driver
+ * behaved correctly. We wan't to avoid looking for a bad
+ * fd, that may even not exist anymore. Therefore, handle
+ * update requests and try again. This behaviour should only
+ * happen when using SELECT as the polling mechanism.
+ */
+ ERTS_POLLSET_LOCK(ps);
+ used_fds += handle_update_requests(ps, pr + used_fds, no_fds - used_fds);
+ if (used_fds == no_fds) {
+ *len = used_fds;
+ ERTS_POLLSET_UNLOCK(ps);
+ return 0;
+ }
+ res = check_fd_events(ps, pr + used_fds, 0, no_fds - used_fds);
+ /* Keep the lock over the non-blocking poll in order to not
+ get any nasty races happening. */
+ ERTS_POLLSET_UNLOCK(ps);
+ if (res == 0) {
+ errno = EAGAIN;
+ res = -1;
+ }
+ }
+#endif
+
if (res != 0)
break;
- if (to == ERTS_POLL_NO_TIMEOUT)
- break;
- if (erts_get_monotonic_time(NULL) >= timeout_time)
- break;
+ if (!do_wait)
+ break;
+ }
+
+ if (do_wait) {
+ erts_thr_progress_finalize_wait(NULL);
+ ERTS_MSACC_UPDATE_CACHE();
+ ERTS_MSACC_SET_STATE_CACHED(ERTS_MSACC_STATE_CHECK_IO);
}
woke_up(ps);
- if (res == 0) {
- res = ETIMEDOUT;
- }
- else if (res < 0) {
+ if (res < 0) {
#if ERTS_POLL_USE_SELECT
if (errno == EBADF) {
ebadf = 1;
@@ -2363,26 +1666,21 @@ ERTS_POLL_EXPORT(erts_poll_wait)(ErtsPollSet ps,
ps_locked = 1;
ERTS_POLLSET_LOCK(ps);
- no_fds = save_poll_result(ps, pr, no_fds, res, ebadf);
+ used_fds += ERTS_POLL_EXPORT(save_result)(ps, pr + used_fds, no_fds - used_fds, res, ebadf);
#ifdef HARD_DEBUG
- check_poll_result(pr, no_fds);
+ check_poll_result(pr, used_fds);
#endif
- res = (no_fds == 0 ? (is_interrupted_reset(ps) ? EINTR : EAGAIN) : 0);
- *len = no_fds;
+ res = (used_fds == 0 ? (is_interrupted_reset(ps) ? EINTR : EAGAIN) : 0);
+ *len = used_fds;
}
if (ps_locked)
ERTS_POLLSET_UNLOCK(ps);
- ERTS_POLLSET_UNSET_POLLED(ps);
- done:
- set_timeout_time(ps, ERTS_MONOTONIC_TIME_MAX);
-#ifdef ERTS_POLL_DEBUG_PRINT
- erts_printf("Leaving %s = erts_poll_wait()\n",
- res == 0 ? "0" : erl_errno_id(res));
-#endif
+ DEBUG_PRINT_WAIT("Leaving %s = %s(len = %d)", ps,
+ res == 0 ? "0" : erl_errno_id(res), __FUNCTION__, *len);
return res;
}
@@ -2392,40 +1690,14 @@ ERTS_POLL_EXPORT(erts_poll_wait)(ErtsPollSet ps,
*/
void
-ERTS_POLL_EXPORT(erts_poll_interrupt)(ErtsPollSet ps, int set)
+ERTS_POLL_EXPORT(erts_poll_interrupt)(ErtsPollSet *ps, int set)
{
+#if !ERTS_POLL_USE_CONCURRENT_UPDATE
if (!set)
reset_wakeup_state(ps);
else
- wake_poller(ps, 1, 0);
-}
-
-
-/*
- * erts_poll_interrupt_timed():
- * If 'set' != 0, interrupt thread blocked in erts_poll_wait() if it
- * is not guaranteed that it will timeout before 'msec' milli seconds.
- */
-void
-ERTS_POLL_EXPORT(erts_poll_interrupt_timed)(ErtsPollSet ps,
- int set,
- ErtsMonotonicTime timeout_time)
-{
- if (!set)
- reset_wakeup_state(ps);
- else {
- ErtsMonotonicTime max_wait_time = get_timeout_time(ps);
- if (max_wait_time > timeout_time)
- wake_poller(ps, 1, 0);
-#ifdef ERTS_POLL_COUNT_AVOIDED_WAKEUPS
- else {
- if (ERTS_POLLSET_IS_POLLED(ps))
- erts_atomic_inc_nob(&ps->no_avoided_wakeups);
- erts_atomic_inc_nob(&ps->no_avoided_interrupts);
- }
- erts_atomic_inc_nob(&ps->no_interrupt_timed);
+ wake_poller(ps, 1);
#endif
- }
}
int
@@ -2438,14 +1710,19 @@ ERTS_POLL_EXPORT(erts_poll_max_fds)(void)
*/
void
-ERTS_POLL_EXPORT(erts_poll_init)(void)
+ERTS_POLL_EXPORT(erts_poll_init)(int *concurrent_updates)
{
- erts_mtx_init(&pollsets_lock, "pollsets_lock", NIL,
- ERTS_LOCK_FLAGS_PROPERTY_STATIC | ERTS_LOCK_FLAGS_CATEGORY_IO);
- pollsets = NULL;
errno = 0;
+ if (concurrent_updates) {
+#if ERTS_POLL_USE_CONCURRENT_UPDATE
+ *concurrent_updates = 1;
+#else
+ *concurrent_updates = 0;
+#endif
+ }
+
#if !defined(NO_SYSCONF)
max_fds = sysconf(_SC_OPEN_MAX);
#elif ERTS_POLL_USE_SELECT
@@ -2464,37 +1741,28 @@ ERTS_POLL_EXPORT(erts_poll_init)(void)
fatal_error("erts_poll_init(): Failed to get max number of files: %s\n",
erl_errno_id(errno));
-#ifdef ERTS_POLL_DEBUG_PRINT
print_misc_debug_info();
-#endif
}
-ErtsPollSet
-ERTS_POLL_EXPORT(erts_poll_create_pollset)(void)
+ErtsPollSet *
+ERTS_POLL_EXPORT(erts_poll_create_pollset)(int id)
{
#if ERTS_POLL_USE_KERNEL_POLL
int kp_fd;
#endif
- ErtsPollSet ps = erts_alloc(ERTS_ALC_T_POLLSET,
+ ErtsPollSet *ps = erts_alloc(ERTS_ALC_T_POLLSET,
sizeof(struct ERTS_POLL_EXPORT(erts_pollset)));
+ ps->id = id;
ps->internal_fd_limit = 0;
- ps->fds_status = NULL;
- ps->fds_status_len = 0;
erts_atomic_init_nob(&ps->no_of_user_fds, 0);
#if ERTS_POLL_USE_KERNEL_POLL
ps->kp_fd = -1;
#if ERTS_POLL_USE_EPOLL
kp_fd = epoll_create(256);
- ps->res_events_len = 0;
- ps->res_events = NULL;
#elif ERTS_POLL_USE_DEVPOLL
kp_fd = open("/dev/poll", O_RDWR);
- ps->res_events_len = 0;
- ps->res_events = NULL;
#elif ERTS_POLL_USE_KQUEUE
kp_fd = kqueue();
- ps->res_events_len = 0;
- ps->res_events = NULL;
#endif
if (kp_fd < 0)
fatal_error("erts_poll_create_pollset(): Failed to "
@@ -2508,10 +1776,6 @@ ERTS_POLL_EXPORT(erts_poll_create_pollset)(void)
": %s (%d)\n",
erl_errno_id(errno), errno);
#endif /* ERTS_POLL_USE_KERNEL_POLL */
-#if ERTS_POLL_USE_BATCH_UPDATE_POLLSET
- /* res_events is also used as write buffer */
- grow_res_events(ps, ERTS_POLL_MIN_BATCH_BUF_SIZE);
-#endif
#if ERTS_POLL_USE_POLL
ps->next_poll_fds_ix = 0;
ps->no_poll_fds = 0;
@@ -2520,9 +1784,6 @@ ERTS_POLL_EXPORT(erts_poll_create_pollset)(void)
#elif ERTS_POLL_USE_SELECT
ps->next_sel_fd = 0;
ps->max_fd = -1;
-#if ERTS_POLL_USE_FALLBACK
- ps->no_select_fds = 0;
-#endif
#ifdef _DARWIN_UNLIMITED_SELECT
ps->input_fds.sz = 0;
ps->input_fds.ptr = NULL;
@@ -2539,140 +1800,48 @@ ERTS_POLL_EXPORT(erts_poll_create_pollset)(void)
ERTS_FD_ZERO(&ps->res_output_fds);
#endif
#endif
+#if !ERTS_POLL_USE_CONCURRENT_UPDATE
+ ps->fds_status = NULL;
+ ps->fds_status_len = 0;
ps->update_requests.next = NULL;
ps->update_requests.len = 0;
ps->curr_upd_req_block = &ps->update_requests;
erts_atomic32_init_nob(&ps->have_update_requests, 0);
- erts_atomic32_init_nob(&ps->polled, 0);
erts_mtx_init(&ps->mtx, "pollset", NIL, ERTS_LOCK_FLAGS_CATEGORY_IO);
- erts_atomic32_init_nob(&ps->wakeup_state, (erts_aint32_t) 0);
- create_wakeup_pipe(ps);
-#if ERTS_POLL_USE_TIMERFD
- create_timerfd(ps);
-#endif
-#if ERTS_POLL_USE_FALLBACK
- if (kp_fd >= ps->fds_status_len)
- grow_fds_status(ps, kp_fd);
- /* Force kernel poll fd into fallback (poll/select) set */
- ps->fds_status[kp_fd].flags
- |= ERTS_POLL_FD_FLG_INFLBCK|ERTS_POLL_FD_FLG_USEFLBCK;
- {
- int do_wake = 0;
- ERTS_POLL_EXPORT(erts_poll_control)(ps, kp_fd, ERTS_POLL_EV_IN, 1,
- &do_wake);
- }
#endif
#if ERTS_POLL_USE_KERNEL_POLL
if (ps->internal_fd_limit <= kp_fd)
ps->internal_fd_limit = kp_fd + 1;
ps->kp_fd = kp_fd;
#endif
- init_timeout_time(ps);
-#ifdef ERTS_POLL_COUNT_AVOIDED_WAKEUPS
- erts_atomic_init_nob(&ps->no_avoided_wakeups, 0);
- erts_atomic_init_nob(&ps->no_avoided_interrupts, 0);
- erts_atomic_init_nob(&ps->no_interrupt_timed, 0);
-#endif
- handle_update_requests(ps);
-#if ERTS_POLL_USE_FALLBACK
- ps->fallback_used = 0;
+#if !ERTS_POLL_USE_CONCURRENT_UPDATE
+ erts_atomic32_init_nob(&ps->wakeup_state, (erts_aint32_t) 0);
+ create_wakeup_pipe(ps);
+ handle_update_requests(ps, NULL, 0);
+ cleanup_wakeup_pipe(ps);
#endif
erts_atomic_set_nob(&ps->no_of_user_fds, 0); /* Don't count wakeup pipe and fallback fd */
- erts_mtx_lock(&pollsets_lock);
- ps->next = pollsets;
- pollsets = ps;
- erts_mtx_unlock(&pollsets_lock);
-
return ps;
}
-void
-ERTS_POLL_EXPORT(erts_poll_destroy_pollset)(ErtsPollSet ps)
-{
-
- if (ps->fds_status)
- erts_free(ERTS_ALC_T_FD_STATUS, (void *) ps->fds_status);
-
-#if ERTS_POLL_USE_EPOLL
- if (ps->kp_fd >= 0)
- close(ps->kp_fd);
- if (ps->res_events)
- erts_free(ERTS_ALC_T_POLL_RES_EVS, (void *) ps->res_events);
-#elif ERTS_POLL_USE_DEVPOLL
- if (ps->kp_fd >= 0)
- close(ps->kp_fd);
- if (ps->res_events)
- erts_free(ERTS_ALC_T_POLL_RES_EVS, (void *) ps->res_events);
-#elif ERTS_POLL_USE_POLL
- if (ps->poll_fds)
- erts_free(ERTS_ALC_T_POLL_FDS, (void *) ps->poll_fds);
-#elif ERTS_POLL_USE_SELECT
-#ifdef _DARWIN_UNLIMITED_SELECT
- if (ps->input_fds.ptr)
- erts_free(ERTS_ALC_T_SELECT_FDS, (void *) ps->input_fds.ptr);
- if (ps->res_input_fds.ptr)
- erts_free(ERTS_ALC_T_SELECT_FDS, (void *) ps->res_input_fds.ptr);
- if (ps->output_fds.ptr)
- erts_free(ERTS_ALC_T_SELECT_FDS, (void *) ps->output_fds.ptr);
- if (ps->res_output_fds.ptr)
- erts_free(ERTS_ALC_T_SELECT_FDS, (void *) ps->res_output_fds.ptr);
-#endif
-#endif
- {
- ErtsPollSetUpdateRequestsBlock *urqbp = ps->update_requests.next;
- while (urqbp) {
- ErtsPollSetUpdateRequestsBlock *free_urqbp = urqbp;
- urqbp = urqbp->next;
- free_update_requests_block(ps, free_urqbp);
- }
- }
- erts_mtx_destroy(&ps->mtx);
- if (ps->wake_fds[0] >= 0)
- close(ps->wake_fds[0]);
- if (ps->wake_fds[1] >= 0)
- close(ps->wake_fds[1]);
-#if ERTS_POLL_USE_TIMERFD
- if (ps->timer_fd >= 0)
- close(ps->timer_fd);
-#endif
-
- erts_mtx_lock(&pollsets_lock);
- if (ps == pollsets)
- pollsets = pollsets->next;
- else {
- ErtsPollSet prev_ps;
- for (prev_ps = pollsets; ps != prev_ps->next; prev_ps = prev_ps->next)
- ;
- ASSERT(ps == prev_ps->next);
- prev_ps->next = ps->next;
- }
- erts_mtx_unlock(&pollsets_lock);
-
- erts_free(ERTS_ALC_T_POLLSET, (void *) ps);
-}
-
/*
* --- Info ------------------------------------------------------------------
*/
void
-ERTS_POLL_EXPORT(erts_poll_info)(ErtsPollSet ps, ErtsPollInfo *pip)
+ERTS_POLL_EXPORT(erts_poll_info)(ErtsPollSet *ps, ErtsPollInfo *pip)
{
+#if !ERTS_POLL_USE_CONCURRENT_UPDATE
int pending_updates;
+#endif
Uint size = 0;
ERTS_POLLSET_LOCK(ps);
size += sizeof(struct ERTS_POLL_EXPORT(erts_pollset));
+#if !ERTS_POLL_USE_CONCURRENT_UPDATE
size += ps->fds_status_len*sizeof(ErtsFdStatus);
-
-#if ERTS_POLL_USE_EPOLL
- size += ps->res_events_len*sizeof(struct epoll_event);
-#elif ERTS_POLL_USE_DEVPOLL
- size += ps->res_events_len*sizeof(struct pollfd);
-#elif ERTS_POLL_USE_KQUEUE
- size += ps->res_events_len*sizeof(struct kevent);
#endif
#if ERTS_POLL_USE_POLL
@@ -2684,6 +1853,7 @@ ERTS_POLL_EXPORT(erts_poll_info)(ErtsPollSet ps, ErtsPollInfo *pip)
#endif
#endif
+#if !ERTS_POLL_USE_CONCURRENT_UPDATE
{
ErtsPollSetUpdateRequestsBlock *urqbp = ps->update_requests.next;
pending_updates = ps->update_requests.len;
@@ -2693,8 +1863,9 @@ ERTS_POLL_EXPORT(erts_poll_info)(ErtsPollSet ps, ErtsPollInfo *pip)
urqbp = urqbp->next;
}
}
+#endif
- pip->primary =
+ pip->primary =
#if ERTS_POLL_USE_KQUEUE
"kqueue"
#elif ERTS_POLL_USE_EPOLL
@@ -2708,17 +1879,7 @@ ERTS_POLL_EXPORT(erts_poll_info)(ErtsPollSet ps, ErtsPollInfo *pip)
#endif
;
- pip->fallback =
-#if !ERTS_POLL_USE_FALLBACK
- NULL
-#elif ERTS_POLL_USE_POLL
- "poll"
-#elif ERTS_POLL_USE_SELECT
- "select"
-#endif
- ;
-
- pip->kernel_poll =
+ pip->kernel_poll =
#if !ERTS_POLL_USE_KERNEL_POLL
NULL
#elif ERTS_POLL_USE_KQUEUE
@@ -2733,40 +1894,21 @@ ERTS_POLL_EXPORT(erts_poll_info)(ErtsPollSet ps, ErtsPollInfo *pip)
pip->memory_size = size;
pip->poll_set_size = (int) erts_atomic_read_nob(&ps->no_of_user_fds);
+#if !ERTS_POLL_USE_CONCURRENT_UPDATE
pip->poll_set_size++; /* Wakeup pipe */
-#if ERTS_POLL_USE_TIMERFD
- pip->poll_set_size++; /* timerfd */
-#endif
-
- pip->fallback_poll_set_size =
-#if !ERTS_POLL_USE_FALLBACK
- 0
-#elif ERTS_POLL_USE_POLL
- ps->no_poll_fds
-#elif ERTS_POLL_USE_SELECT
- ps->no_select_fds
-#endif
- ;
-
-#if ERTS_POLL_USE_FALLBACK
- /* If only kp_fd is in fallback poll set we don't use fallback... */
- if (pip->fallback_poll_set_size == 1)
- pip->fallback_poll_set_size = 0;
- else
- pip->poll_set_size++; /* kp_fd */
#endif
pip->lazy_updates =
+#if !ERTS_POLL_USE_CONCURRENT_UPDATE
1
+#else
+ 0
+#endif
;
pip->pending_updates =
+#if !ERTS_POLL_USE_CONCURRENT_UPDATE
pending_updates
- ;
-
- pip->batch_updates =
-#if ERTS_POLL_USE_BATCH_UPDATE_POLLSET
- 1
#else
0
#endif
@@ -2780,13 +1922,15 @@ ERTS_POLL_EXPORT(erts_poll_info)(ErtsPollSet ps, ErtsPollInfo *pip)
#endif
;
- pip->max_fds = max_fds;
-
-#ifdef ERTS_POLL_COUNT_AVOIDED_WAKEUPS
- pip->no_avoided_wakeups = erts_atomic_read_nob(&ps->no_avoided_wakeups);
- pip->no_avoided_interrupts = erts_atomic_read_nob(&ps->no_avoided_interrupts);
- pip->no_interrupt_timed = erts_atomic_read_nob(&ps->no_interrupt_timed);
+ pip->is_fallback =
+#if ERTS_POLL_IS_FALLBACK
+ 1
+#else
+ 0
#endif
+ ;
+
+ pip->max_fds = max_fds;
ERTS_POLLSET_UNLOCK(ps);
@@ -2822,35 +1966,60 @@ fatal_error(char *format, ...)
abort();
}
-static void
-fatal_error_async_signal_safe(char *error_str)
+/*
+ * --- Debug -----------------------------------------------------------------
+ */
+
+#if ERTS_POLL_USE_EPOLL
+uint32_t epoll_events(int kp_fd, int fd)
{
- if (ERTS_SOMEONE_IS_CRASH_DUMPING || ERTS_GOT_SIGUSR1) {
- /* See comment above in fatal_error() */
- return;
+ /* For epoll we read the information about what is selected upon from the proc fs.*/
+ char fname[30];
+ FILE *f;
+ unsigned int pos, flags, mnt_id;
+ int line = 0;
+ sprintf(fname,"/proc/%d/fdinfo/%d",getpid(), kp_fd);
+ f = fopen(fname,"r");
+ if (!f) {
+ fprintf(stderr,"failed to open file %s, errno = %d\n", fname, errno);
+ ASSERT(0);
+ return 0;
}
- if (error_str) {
- int len = 0;
- while (error_str[len])
- len++;
- if (len) {
- /* async signal safe */
- erts_silence_warn_unused_result(write(2, error_str, len));
- }
+ if (fscanf(f,"pos:\t%x\nflags:\t%x", &pos, &flags) != 2) {
+ fprintf(stderr,"failed to parse file %s, errno = %d\n", fname, errno);
+ ASSERT(0);
+ return 0;
}
- abort();
+ if (fscanf(f,"\nmnt_id:\t%x\n", &mnt_id));
+ line += 3;
+ while (!feof(f)) {
+ /* tfd: 10 events: 40000019 data: 180000000a */
+ int ev_fd;
+ uint32_t events;
+ uint64_t data;
+ if (fscanf(f,"tfd:%d events:%x data:%lx\n", &ev_fd, &events, &data) != 3) {
+ fprintf(stderr,"failed to parse file %s on line %d, errno = %d\n", fname,
+ line,
+ errno);
+ return 0;
+ }
+ if (fd == ev_fd) {
+ fclose(f);
+ return events;
+ }
+ }
+ fclose(f);
+ return 0;
}
-
-/*
- * --- Debug -----------------------------------------------------------------
- */
+#endif
void
-ERTS_POLL_EXPORT(erts_poll_get_selected_events)(ErtsPollSet ps,
+ERTS_POLL_EXPORT(erts_poll_get_selected_events)(ErtsPollSet *ps,
ErtsPollEvents ev[],
int len)
{
int fd;
+#if !ERTS_POLL_USE_CONCURRENT_UPDATE
ERTS_POLLSET_LOCK(ps);
for (fd = 0; fd < len; fd++) {
if (fd >= ps->fds_status_len)
@@ -2859,9 +2028,6 @@ ERTS_POLL_EXPORT(erts_poll_get_selected_events)(ErtsPollSet ps,
ev[fd] = ps->fds_status[fd].events;
if (
fd == ps->wake_fds[0] || fd == ps->wake_fds[1] ||
-#if ERTS_POLL_USE_TIMERFD
- fd == ps->timer_fd ||
-#endif
#if ERTS_POLL_USE_KERNEL_POLL
fd == ps->kp_fd ||
#endif
@@ -2870,7 +2036,50 @@ ERTS_POLL_EXPORT(erts_poll_get_selected_events)(ErtsPollSet ps,
}
}
ERTS_POLLSET_UNLOCK(ps);
+#elif ERTS_POLL_USE_EPOLL
+ /* For epoll we read the information about what is selected upon from the proc fs.*/
+ char fname[30];
+ FILE *f;
+ unsigned int pos, flags, mnt_id;
+ int line = 0;
+ sprintf(fname,"/proc/%d/fdinfo/%d",getpid(), ps->kp_fd);
+ for (fd = 0; fd < len; fd++)
+ ev[fd] = ERTS_POLL_EV_NONE;
+ f = fopen(fname,"r");
+ if (!f) {
+ fprintf(stderr,"failed to open file %s, errno = %d\n", fname, errno);
+ return;
+ }
+ if (fscanf(f,"pos:\t%x\nflags:\t%x", &pos, &flags) != 2) {
+ fprintf(stderr,"failed to parse file %s, errno = %d\n", fname, errno);
+ ASSERT(0);
+ return;
+ }
+ if (fscanf(f,"\nmnt_id:\t%x\n", &mnt_id));
+ line += 3;
+ while (!feof(f)) {
+ /* tfd: 10 events: 40000019 data: 180000000a */
+ int fd;
+ uint32_t events;
+ uint64_t data;
+ if (fscanf(f,"tfd:%d events:%x data:%lx\n", &fd, &events, &data) != 3) {
+ fprintf(stderr,"failed to parse file %s on line %d, errno = %d\n",
+ fname, line, errno);
+ ASSERT(0);
+ return;
+ }
+ data &= 0xFFFFFFFF;
+ ASSERT(fd == data);
+ /* Events are the events that are being monitored, which of course include
+ error and hup events, but we are only interested in IN/OUT events */
+ ev[fd] = (ERTS_POLL_EV_IN|ERTS_POLL_EV_OUT) & ERTS_POLL_EV_N2E(events);
+ line++;
+ }
+#else
+ for (fd = 0; fd < len; fd++)
+ ev[fd] = ERTS_POLL_EV_NONE;
+#endif
}
#ifdef HARD_DEBUG
@@ -2890,10 +2099,10 @@ check_poll_result(ErtsPollResFd pr[], int len)
}
-#if ERTS_POLL_USE_DEVPOLL
+#if ERTS_POLL_USE_DEVPOLL && defined(DEBUG)
static void
-check_poll_status(ErtsPollSet ps)
+check_poll_status(ErtsPollSet *ps)
{
int i;
for (i = 0; i < ps->fds_status_len; i++) {
@@ -2925,30 +2134,24 @@ check_poll_status(ErtsPollSet ps)
#endif /* ERTS_POLL_USE_DEVPOLL */
#endif /* HARD_DEBUG */
-#ifdef ERTS_POLL_DEBUG_PRINT
static void
print_misc_debug_info(void)
{
- erts_printf("erts_poll using: %s lazy_updates:%s batch_updates:%s\n",
+#if ERTS_POLL_DEBUG_PRINT
+ erts_printf("erts_poll using: %s lazy_updates:%s\n",
#if ERTS_POLL_USE_KQUEUE
"kqueue"
#elif ERTS_POLL_USE_EPOLL
"epoll"
#elif ERTS_POLL_USE_DEVPOLL
"/dev/poll"
-#endif
-#if ERTS_POLL_USE_FALLBACK
- "-"
-#endif
-#if ERTS_POLL_USE_POLL
+#elif ERTS_POLL_USE_POLL
"poll"
#elif ERTS_POLL_USE_SELECT
"select"
#endif
,
- "true"
- ,
-#if ERTS_POLL_USE_BATCH_UPDATE_POLLSET
+#if !ERTS_POLL_USE_CONCURRENT_UPDATE
"true"
#else
"false"
@@ -2967,29 +2170,20 @@ print_misc_debug_info(void)
#ifdef FD_SETSIZE
erts_printf("FD_SETSIZE=%d\n", FD_SETSIZE);
#endif
-}
-
#endif
+}
#ifdef ERTS_ENABLE_LOCK_COUNT
-static void erts_lcnt_enable_pollset_lock_count(ErtsPollSet pollset, int enable) {
+void ERTS_POLL_EXPORT(erts_lcnt_enable_pollset_lock_count)(ErtsPollSet *pollset, int enable)
+{
+#if !ERTS_POLL_USE_CONCURRENT_UPDATE
if(enable) {
erts_lcnt_install_new_lock_info(&pollset->mtx.lcnt, "pollset_rm", NIL,
ERTS_LOCK_TYPE_MUTEX | ERTS_LOCK_FLAGS_CATEGORY_IO);
} else {
erts_lcnt_uninstall(&pollset->mtx.lcnt);
}
-}
-
-void ERTS_POLL_EXPORT(erts_lcnt_update_pollset_locks)(int enable) {
- ErtsPollSet iterator;
-
- erts_mtx_lock(&pollsets_lock);
-
- for(iterator = pollsets; iterator != NULL; iterator = iterator->next) {
- erts_lcnt_enable_pollset_lock_count(iterator, enable);
- }
-
- erts_mtx_unlock(&pollsets_lock);
+#endif
+ return;
}
#endif
diff --git a/erts/emulator/sys/common/erl_poll.h b/erts/emulator/sys/common/erl_poll.h
index 6c961205fe..a1e2709a9f 100644
--- a/erts/emulator/sys/common/erl_poll.h
+++ b/erts/emulator/sys/common/erl_poll.h
@@ -18,11 +18,31 @@
* %CopyrightEnd%
*/
-/*
- * Description: Poll interface suitable for ERTS with or without
- * SMP support.
+/**
+ * @description: Poll interface suitable for ERTS with SMP support.
+ *
+ * @author: Rickard Green
+ * @author: Lukas Larsson
+ *
+ * This header file exports macros and functions that are used to
+ * react to I/O polling events from file descriptors or wait-able
+ * objects. The API exported is the following:
*
- * Author: Rickard Green
+ * defines:
+ * ERTS_POLL_EV_NONE - No events have been set. This is not the same as 0.
+ * ERTS_POLL_EV_IN - Represent an IN event
+ * ERTS_POLL_EV_OUT - Represent an OUT event
+ * ERTS_POLL_EV_ERR - Represent an error event
+ * ERTS_POLL_EV_NVAL - Represent an invalid event
+ *
+ * macro functions:
+ * ErtsSysFdType ERTS_POLL_RES_GET_FD(ErtsPollResFd *evt);
+ * void ERTS_POLL_RES_SET_FD(ErtsPollResFd *evt, ErtsSysFdType fd);
+ * ErtsPollEvents ERTS_POLL_RES_GET_EVTS(ErtsPollResFd *evt)
+ * void ERTS_POLL_RES_SET_EVTS(ErtsPollResFd *evt, ErtsPollEvents fd);
+ *
+ * functions:
+ * See erl_poll_api.h
*/
#ifndef ERL_POLL_H__
@@ -32,33 +52,27 @@
#define ERTS_POLL_NO_TIMEOUT ERTS_MONOTONIC_TIME_MIN
-#if 0
-#define ERTS_POLL_COUNT_AVOIDED_WAKEUPS
-#endif
-
#ifdef ERTS_ENABLE_KERNEL_POLL
-# if defined(ERTS_KERNEL_POLL_VERSION)
-# define ERTS_POLL_EXPORT(FUNC) FUNC ## _kp
+# undef ERTS_ENABLE_KERNEL_POLL
+# define ERTS_ENABLE_KERNEL_POLL 1
+# if defined(ERTS_NO_KERNEL_POLL_VERSION)
+# define ERTS_POLL_EXPORT(FUNC) FUNC ## _flbk
+# undef ERTS_NO_KERNEL_POLL_VERSION
+# define ERTS_NO_KERNEL_POLL_VERSION 1
+# define ERTS_KERNEL_POLL_VERSION 0
# else
-# define ERTS_POLL_EXPORT(FUNC) FUNC ## _nkp
-# undef ERTS_POLL_DISABLE_KERNEL_POLL
-# define ERTS_POLL_DISABLE_KERNEL_POLL
+# undef ERTS_KERNEL_POLL_VERSION
+# define ERTS_KERNEL_POLL_VERSION 1
+# define ERTS_NO_KERNEL_POLL_VERSION 0
+# define ERTS_POLL_EXPORT(FUNC) FUNC
# endif
#else
# define ERTS_POLL_EXPORT(FUNC) FUNC
-# undef ERTS_POLL_DISABLE_KERNEL_POLL
-# define ERTS_POLL_DISABLE_KERNEL_POLL
-#endif
-
-#ifdef ERTS_POLL_DISABLE_KERNEL_POLL
-# undef HAVE_SYS_EPOLL_H
-# undef HAVE_SYS_EVENT_H
-# undef HAVE_SYS_DEVPOLL_H
+# define ERTS_ENABLE_KERNEL_POLL 0
+# define ERTS_NO_KERNEL_POLL_VERSION 1
+# define ERTS_KERNEL_POLL_VERSION 0
#endif
-#undef ERTS_POLL_USE_KERNEL_POLL
-#define ERTS_POLL_USE_KERNEL_POLL 0
-
#undef ERTS_POLL_USE_KQUEUE
#define ERTS_POLL_USE_KQUEUE 0
#undef ERTS_POLL_USE_EPOLL
@@ -70,68 +84,96 @@
#undef ERTS_POLL_USE_SELECT
#define ERTS_POLL_USE_SELECT 0
-#if defined(HAVE_SYS_EVENT_H)
-# undef ERTS_POLL_USE_KQUEUE
-# define ERTS_POLL_USE_KQUEUE 1
-# undef ERTS_POLL_USE_KERNEL_POLL
-# define ERTS_POLL_USE_KERNEL_POLL 1
-#elif defined(HAVE_SYS_EPOLL_H)
-# undef ERTS_POLL_USE_EPOLL
-# define ERTS_POLL_USE_EPOLL 1
-# undef ERTS_POLL_USE_KERNEL_POLL
-# define ERTS_POLL_USE_KERNEL_POLL 1
-#elif defined(HAVE_SYS_DEVPOLL_H)
-# undef ERTS_POLL_USE_DEVPOLL
-# define ERTS_POLL_USE_DEVPOLL 1
-# undef ERTS_POLL_USE_KERNEL_POLL
-# define ERTS_POLL_USE_KERNEL_POLL 1
+/* Defines which structure that erts_poll_wait should use to wait with
+ and how events should be represented */
+#define ERTS_POLL_USE_EPOLL_EVS 0
+#define ERTS_POLL_USE_KQUEUE_EVS 0
+#define ERTS_POLL_USE_DEVPOLL_EVS 0
+#define ERTS_POLL_USE_POLL_EVS 0
+#define ERTS_POLL_USE_SELECT_EVS 0
+
+#define ERTS_POLL_USE_KERNEL_POLL ERTS_KERNEL_POLL_VERSION
+
+#if ERTS_ENABLE_KERNEL_POLL
+# if defined(HAVE_SYS_EVENT_H)
+# undef ERTS_POLL_USE_KQUEUE_EVS
+# define ERTS_POLL_USE_KQUEUE_EVS 1
+# undef ERTS_POLL_USE_KQUEUE
+# define ERTS_POLL_USE_KQUEUE ERTS_KERNEL_POLL_VERSION
+# elif defined(HAVE_SYS_EPOLL_H)
+# undef ERTS_POLL_USE_EPOLL_EVS
+# define ERTS_POLL_USE_EPOLL_EVS 1
+# undef ERTS_POLL_USE_EPOLL
+# define ERTS_POLL_USE_EPOLL ERTS_KERNEL_POLL_VERSION
+# elif defined(HAVE_SYS_DEVPOLL_H)
+# undef ERTS_POLL_USE_DEVPOLL_EVS
+# define ERTS_POLL_USE_DEVPOLL_EVS 1
+# undef ERTS_POLL_USE_DEVPOLL
+# define ERTS_POLL_USE_DEVPOLL ERTS_KERNEL_POLL_VERSION
+# else
+# error "Missing kernel poll implementation of erts_poll()"
+# endif
#endif
-#define ERTS_POLL_USE_FALLBACK (ERTS_POLL_USE_KQUEUE || ERTS_POLL_USE_EPOLL)
-
-#if !ERTS_POLL_USE_KERNEL_POLL || ERTS_POLL_USE_FALLBACK
+#if ERTS_NO_KERNEL_POLL_VERSION
# if defined(ERTS_USE_POLL)
+# undef ERTS_POLL_USE_POLL_EVS
+# define ERTS_POLL_USE_POLL_EVS 1
# undef ERTS_POLL_USE_POLL
# define ERTS_POLL_USE_POLL 1
# elif !defined(__WIN32__)
+# undef ERTS_POLL_USE_SELECT_EVS
+# define ERTS_POLL_USE_SELECT_EVS 1
# undef ERTS_POLL_USE_SELECT
# define ERTS_POLL_USE_SELECT 1
# endif
#endif
-#define ERTS_POLL_USE_TIMERFD 0
+#define ERTS_POLL_USE_FALLBACK (ERTS_POLL_USE_KQUEUE || ERTS_POLL_USE_EPOLL)
typedef Uint32 ErtsPollEvents;
-#undef ERTS_POLL_EV_E2N
+
+typedef enum {
+ ERTS_POLL_OP_ADD = 0, /* Add the FD to the pollset */
+ ERTS_POLL_OP_MOD = 1, /* Modify the FD in the pollset */
+ ERTS_POLL_OP_DEL = 2 /* Delete the FD from the pollset */
+} ErtsPollOp;
+
+#define op2str(op) (op == ERTS_POLL_OP_ADD ? "add" : \
+ (op == ERTS_POLL_OP_MOD ? "mod" : "del"))
#if defined(__WIN32__) /* --- win32 --------------------------------------- */
-#define ERTS_POLL_EV_IN 1
-#define ERTS_POLL_EV_OUT 2
-#define ERTS_POLL_EV_ERR 4
-#define ERTS_POLL_EV_NVAL 8
+#define ERTS_POLL_EV_IN 1
+#define ERTS_POLL_EV_OUT 2
+#define ERTS_POLL_EV_ERR 4
+#define ERTS_POLL_EV_NVAL 8
-#elif ERTS_POLL_USE_EPOLL /* --- epoll ------------------------------- */
+#define ERTS_POLL_EV_E2N(EV) (EV)
+#define ERTS_POLL_EV_N2E(EV) (EV)
-#include <sys/epoll.h>
+#elif ERTS_POLL_USE_EPOLL_EVS /* --- epoll ------------------------------- */
-#ifdef HAVE_SYS_TIMERFD_H
-#include <sys/timerfd.h>
-#undef ERTS_POLL_USE_TIMERFD
-#define ERTS_POLL_USE_TIMERFD 1
-#endif
+#include <sys/epoll.h>
#define ERTS_POLL_EV_E2N(EV) \
((uint32_t) (EV))
#define ERTS_POLL_EV_N2E(EV) \
- ((ErtsPollEvents) (EV))
+ ((ErtsPollEvents) (EV) & ~EPOLLONESHOT)
#define ERTS_POLL_EV_IN ERTS_POLL_EV_N2E(EPOLLIN)
#define ERTS_POLL_EV_OUT ERTS_POLL_EV_N2E(EPOLLOUT)
#define ERTS_POLL_EV_NVAL ERTS_POLL_EV_N2E(EPOLLET)
#define ERTS_POLL_EV_ERR ERTS_POLL_EV_N2E(EPOLLERR|EPOLLHUP)
-#elif ERTS_POLL_USE_DEVPOLL /* --- devpoll ----------------------------- */
+typedef struct epoll_event ErtsPollResFd;
+
+#define ERTS_POLL_RES_GET_FD(evt) ((ErtsSysFdType)((evt)->data.fd))
+#define ERTS_POLL_RES_SET_FD(evt, ident) (evt)->data.fd = ident
+#define ERTS_POLL_RES_GET_EVTS(evt) ERTS_POLL_EV_N2E((evt)->events)
+#define ERTS_POLL_RES_SET_EVTS(evt, evts) (evt)->events = ERTS_POLL_EV_E2N(evts)
+
+#elif ERTS_POLL_USE_DEVPOLL_EVS /* --- devpoll ----------------------------- */
#include <sys/devpoll.h>
@@ -145,12 +187,37 @@ typedef Uint32 ErtsPollEvents;
#define ERTS_POLL_EV_NVAL ERTS_POLL_EV_N2E(POLLNVAL)
#define ERTS_POLL_EV_ERR ERTS_POLL_EV_N2E(POLLERR|POLLHUP)
-#elif ERTS_POLL_USE_KQUEUE /* --- kqueue ------------------------------ */
+typedef struct pollfd ErtsPollResFd;
+
+#define ERTS_POLL_RES_GET_FD(evt) ((ErtsSysFdType)((evt)->fd))
+#define ERTS_POLL_RES_SET_FD(evt, ident) (evt)->fd = ident
+#define ERTS_POLL_RES_GET_EVTS(evt) ERTS_POLL_EV_N2E((evt)->revents)
+#define ERTS_POLL_RES_SET_EVTS(evt, evts) (evt)->revents = ERTS_POLL_EV_E2N(evts)
+
+#elif ERTS_POLL_USE_KQUEUE_EVS /* --- kqueue ------------------------------ */
/* Kqueue use fallback defines (poll() or select()) */
+
+#include <sys/event.h>
+
+#ifdef ERTS_USE_POLL
+# undef ERTS_POLL_USE_POLL_EVS
+# define ERTS_POLL_USE_POLL_EVS 1
+#elif !defined(__WIN32__)
+# undef ERTS_POLL_USE_SELECT_EVS
+# define ERTS_POLL_USE_SELECT_EVS 1
#endif
-#if ERTS_POLL_USE_POLL /* --- poll -------------------------------- */
+typedef struct kevent ErtsPollResFd;
+
+#define ERTS_POLL_RES_GET_FD(evt) ((ErtsSysFdType)((evt)->ident))
+#define ERTS_POLL_RES_SET_FD(evt, fd) (evt)->ident = fd
+#define ERTS_POLL_RES_GET_EVTS(evt) ERTS_POLL_EV_N2E((ErtsPollEvents)(evt)->udata)
+#define ERTS_POLL_RES_SET_EVTS(evt, evts) (evt)->udata = (void*)(UWord)(ERTS_POLL_EV_E2N(evts))
+
+#endif
+#if ERTS_POLL_USE_POLL_EVS
+ /* --- poll -------------------------------- */
#include <poll.h>
#define ERTS_POLL_EV_NKP_E2N(EV) \
@@ -169,7 +236,7 @@ typedef Uint32 ErtsPollEvents;
#define ERTS_POLL_EV_NKP_NVAL ERTS_POLL_EV_N2E(POLLNVAL)
#define ERTS_POLL_EV_NKP_ERR ERTS_POLL_EV_N2E(POLLERR|POLLHUP)
-#elif ERTS_POLL_USE_SELECT /* --- select ------------------------------ */
+#elif ERTS_POLL_USE_SELECT_EVS /* --- select ------------------------------ */
#define ERTS_POLL_EV_NKP_E2N(EV) (EV)
#define ERTS_POLL_EV_NKP_N2E(EV) (EV)
@@ -195,71 +262,65 @@ typedef Uint32 ErtsPollEvents;
#endif
-typedef struct ERTS_POLL_EXPORT(erts_pollset) *ErtsPollSet;
+#if !ERTS_ENABLE_KERNEL_POLL
-typedef struct {
- ErtsSysFdType fd;
- ErtsPollEvents events;
- int on;
-} ErtsPollControlEntry;
-
-typedef struct {
+typedef struct _ErtsPollResFd {
ErtsSysFdType fd;
ErtsPollEvents events;
} ErtsPollResFd;
+#define ERTS_POLL_RES_GET_FD(evt) (evt)->fd
+#define ERTS_POLL_RES_SET_FD(evt, ident) (evt)->fd = (ident)
+#define ERTS_POLL_RES_GET_EVTS(evt) ERTS_POLL_EV_N2E((evt)->events)
+#define ERTS_POLL_RES_SET_EVTS(evt, evts) (evt)->events = ERTS_POLL_EV_E2N(evts)
+
+#endif
+
+#define ERTS_POLL_EV_NONE (UINT_MAX & ~(ERTS_POLL_EV_IN|ERTS_POLL_EV_OUT|ERTS_POLL_EV_NVAL|ERTS_POLL_EV_ERR))
+
+#define ev2str(ev) \
+ (((ev) == 0 || (ev) == ERTS_POLL_EV_NONE) ? "NONE" : \
+ ((ev) == ERTS_POLL_EV_IN ? "IN" : \
+ ((ev) == ERTS_POLL_EV_OUT ? "OUT" : \
+ ((ev) == (ERTS_POLL_EV_IN|ERTS_POLL_EV_OUT) ? "IN|OUT" : \
+ ((ev) & ERTS_POLL_EV_ERR ? "ERR" : \
+ ((ev) & ERTS_POLL_EV_NVAL ? "NVAL" : "OTHER"))))))
+
+
+typedef struct ERTS_POLL_EXPORT(erts_pollset) ErtsPollSet;
+
typedef struct {
char *primary;
- char *fallback;
char *kernel_poll;
Uint memory_size;
- int poll_set_size;
- int fallback_poll_set_size;
+ Uint poll_set_size;
int lazy_updates;
- int pending_updates;
+ Uint pending_updates;
int batch_updates;
int concurrent_updates;
- int max_fds;
-#ifdef ERTS_POLL_COUNT_AVOIDED_WAKEUPS
- long no_avoided_wakeups;
- long no_avoided_interrupts;
- long no_interrupt_timed;
-#endif
- int active_fds;
+ int is_fallback;
+ Uint max_fds;
+ Uint active_fds;
+ Uint poll_threads;
} ErtsPollInfo;
-void ERTS_POLL_EXPORT(erts_poll_interrupt)(ErtsPollSet,
- int);
-void ERTS_POLL_EXPORT(erts_poll_interrupt_timed)(ErtsPollSet,
- int,
- ErtsMonotonicTime);
-ErtsPollEvents ERTS_POLL_EXPORT(erts_poll_control)(ErtsPollSet,
- ErtsSysFdType,
- ErtsPollEvents,
- int on,
- int* wake_poller
- );
-void ERTS_POLL_EXPORT(erts_poll_controlv)(ErtsPollSet,
- ErtsPollControlEntry [],
- int on);
-int ERTS_POLL_EXPORT(erts_poll_wait)(ErtsPollSet,
- ErtsPollResFd [],
- int *,
- ErtsMonotonicTime);
-int ERTS_POLL_EXPORT(erts_poll_max_fds)(void);
-void ERTS_POLL_EXPORT(erts_poll_info)(ErtsPollSet,
- ErtsPollInfo *);
-ErtsPollSet ERTS_POLL_EXPORT(erts_poll_create_pollset)(void);
-void ERTS_POLL_EXPORT(erts_poll_destroy_pollset)(ErtsPollSet);
-void ERTS_POLL_EXPORT(erts_poll_init)(void);
-void ERTS_POLL_EXPORT(erts_poll_get_selected_events)(ErtsPollSet,
- ErtsPollEvents [],
- int);
+#if defined(ERTS_POLL_USE_FALLBACK) && ERTS_KERNEL_POLL_VERSION
+# undef ERTS_POLL_EXPORT
+# define ERTS_POLL_EXPORT(FUNC) FUNC ## _flbk
+# include "erl_poll_api.h"
+# undef ERTS_POLL_EXPORT
+# define ERTS_POLL_EXPORT(FUNC) FUNC
+#elif !defined(ERTS_POLL_USE_FALLBACK)
+# define ERTS_POLL_USE_FALLBACK 0
+#endif
-int erts_poll_new_table_len(int old_len, int need_len);
+#include "erl_poll_api.h"
-#ifdef ERTS_ENABLE_LOCK_COUNT
-void ERTS_POLL_EXPORT(erts_lcnt_update_pollset_locks)(int enable);
-#endif
+/**
+ * Get the next size of the array that holds the file descriptors.
+ * This function is used in order for the check io array and the
+ * pollset array to be of the same size.
+ */
+int erts_poll_new_table_len(int old_len, int need_len);
#endif /* #ifndef ERL_POLL_H__ */
diff --git a/erts/emulator/sys/common/erl_poll_api.h b/erts/emulator/sys/common/erl_poll_api.h
new file mode 100644
index 0000000000..04beb37d1c
--- /dev/null
+++ b/erts/emulator/sys/common/erl_poll_api.h
@@ -0,0 +1,122 @@
+/*
+ * %CopyrightBegin%
+ *
+ * Copyright Ericsson AB 2006-2016. All Rights Reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ *
+ * %CopyrightEnd%
+ */
+/**
+ * @description: Poll interface functions
+ * @author Lukas Larsson
+ *
+ * The functions in the header are used to interact with the poll
+ * implementation. Iff the kernel-poll implementation needs a fallback
+ * pollset, then all functions are exported twice. Once with a _flbk
+ * suffix and once without any suffix. If no fallback is needed, then
+ * only the non-suffix version is exported.
+ */
+
+/**
+ * Initialize the poll implementation. Has to be called before any other function.
+ * @param[out] concurrent_waiters if not NULL, set to 1 if more then one thread
+ * is allowed to wait in the pollsets at the same time.
+ */
+void ERTS_POLL_EXPORT(erts_poll_init)(int *concurrent_waiters);
+/**
+ * @brief Create a new pollset.
+ * @param id The unique debug id of this pollset.
+ */
+ErtsPollSet *ERTS_POLL_EXPORT(erts_poll_create_pollset)(int id);
+
+/**
+ * Modify the contents of a pollset. This function can be called while one
+ * (or possibly more) thread is waiting in the pollset.
+ *
+ * @param ps the pollset to modify
+ * @param fd the file descriptor to modify
+ * @param op the type of operation to do. Normal usage is ADD,MOD...MOD,DEL.
+ * @param evts the events that we are changing interest to. Ignored if op is DEL.
+ * @param[in] wake_poller if set to 1 any thread waiting in the pollset will be woken.
+ * This parameter is ignored if the pollset supports concurrent waiters.
+ * @param[out] wake_poller set to 1 if the waiting thread was woken.
+ * @return The events set, or ERTS_POLL_EV_NVAL if it was not possible to add the
+ * fd to the pollset.
+ */
+ErtsPollEvents ERTS_POLL_EXPORT(erts_poll_control)(ErtsPollSet *ps,
+ ErtsSysFdType fd,
+ ErtsPollOp op,
+ ErtsPollEvents evts,
+ int *wake_poller);
+
+/**
+ * Wait for events to be ready in the pollset. If the erts_poll_init call
+ * set concurrent_waiters to 1, then multiple threads are allowed to call
+ * this function at the same time.
+ *
+ * When an event has been triggered on a fd, that event is disabled. To
+ * re-enable it the implementation has to call erts_poll_control again.
+ *
+ * @param ps the pollset to wait for events in
+ * @param res an array of fd results that the ready fds are put in.
+ * @param[in] length the length of the res array
+ * @param[out] length the number of ready events returned in res
+ * @return 0 on success, else the ERRNO of the error that happened.
+ */
+int ERTS_POLL_EXPORT(erts_poll_wait)(ErtsPollSet *ps,
+ ErtsPollResFd res[],
+ int *length);
+/**
+ * Interrupt the thread waiting in the pollset. This function should be called
+ * with set = 0 before any thread calls erts_poll_wait in order to clear any
+ * interrupts that have happened while the thread was awake.
+ *
+ * This function has no effect on pollsets that support concurrent waiters.
+ *
+ * @param ps the pollset to wake
+ * @param set if 1, interrupt the pollset, if 0 clear the interrupt flag.
+ */
+void ERTS_POLL_EXPORT(erts_poll_interrupt)(ErtsPollSet *ps, int set);
+
+/* Debug functions */
+
+/**
+ * Get the maximum number of fds supported by the pollset
+ */
+int ERTS_POLL_EXPORT(erts_poll_max_fds)(void);
+/**
+ * Get information about the given pollset
+ */
+void ERTS_POLL_EXPORT(erts_poll_info)(ErtsPollSet *ps,
+ ErtsPollInfo *info);
+/**
+ * Get information about which events are currently selected.
+ *
+ * The unix fd is used to index into the array, so naturally this function does
+ * not work on windows. If the pollset cannot figure out what the selected
+ * events for a given fd is, it is set to ERTS_POLL_EV_NONE.
+ *
+ * @param ps the pollset to get events from
+ * @param evts an array of which events are selected on.
+ */
+void ERTS_POLL_EXPORT(erts_poll_get_selected_events)(ErtsPollSet *ps,
+ ErtsPollEvents evts[],
+ int length);
+
+#ifdef ERTS_ENABLE_LOCK_COUNT
+/**
+ * Enable lock counting of any locks within the pollset.
+ */
+void ERTS_POLL_EXPORT(erts_lcnt_enable_pollset_lock_count)(ErtsPollSet *, int enable);
+#endif