Skip to content
This repository has been archived by the owner on May 4, 2018. It is now read-only.

Commit

Permalink
linux: share inotify fd across event loop
Browse files Browse the repository at this point in the history
Previously, a new inotify fd was created for each watcher, making it quite easy
to run into the system-wide fs.inotify.max_user_instances limit (usually 128).

Fixes #300.
  • Loading branch information
bnoordhuis committed Feb 23, 2012
1 parent 09c722e commit d3efefb
Show file tree
Hide file tree
Showing 8 changed files with 374 additions and 252 deletions.
2 changes: 1 addition & 1 deletion config-unix.mk
Expand Up @@ -63,7 +63,7 @@ EIO_CONFIG=config_linux.h
CSTDFLAG += -D_GNU_SOURCE
CPPFLAGS += -Isrc/ares/config_linux
LINKFLAGS+=-lrt
OBJS += src/unix/linux.o
OBJS += src/unix/linux/core.o src/unix/linux/inotify.o
endif

ifeq (FreeBSD,$(uname_S))
Expand Down
28 changes: 24 additions & 4 deletions include/uv-private/uv-unix.h
Expand Up @@ -55,6 +55,18 @@ typedef pthread_rwlock_t uv_rwlock_t;
typedef void* uv_lib_t;
#define UV_DYNAMIC /* empty */

#if __linux__
# define UV_LOOP_PRIVATE_PLATFORM_FIELDS \
/* RB_HEAD(uv__inotify_watchers, uv_fs_event_s) */ \
struct uv__inotify_watchers { \
struct uv_fs_event_s* rbh_root; \
} inotify_watchers; \
ev_io inotify_read_watcher; \
int inotify_fd;
#else
# define UV_LOOP_PRIVATE_PLATFORM_FIELDS
#endif

#define UV_LOOP_PRIVATE_FIELDS \
ares_channel channel; \
/* \
Expand All @@ -65,7 +77,8 @@ typedef void* uv_lib_t;
ev_timer timer; \
/* Poll result queue */ \
eio_channel uv_eio_channel; \
struct ev_loop* ev;
struct ev_loop* ev; \
UV_LOOP_PRIVATE_PLATFORM_FIELDS

#define UV_REQ_BUFSML_SIZE (4)

Expand Down Expand Up @@ -195,9 +208,16 @@ typedef void* uv_lib_t;
/* UV_FS_EVENT_PRIVATE_FIELDS */
#if defined(__linux__)

#define UV_FS_EVENT_PRIVATE_FIELDS \
ev_io read_watcher; \
uv_fs_event_cb cb; \
#define UV_FS_EVENT_PRIVATE_FIELDS \
/* RB_ENTRY(fs_event_s) node; */ \
struct { \
struct uv_fs_event_s* rbe_left; \
struct uv_fs_event_s* rbe_right; \
struct uv_fs_event_s* rbe_parent; \
int rbe_color; \
} node; \
ev_io read_watcher; \
uv_fs_event_cb cb;

#elif (defined(__MAC_OS_X_VERSION_MIN_REQUIRED) && __MAC_OS_X_VERSION_MIN_REQUIRED >= 1060) \
|| defined(__FreeBSD__) \
Expand Down
2 changes: 2 additions & 0 deletions include/uv.h
Expand Up @@ -1504,6 +1504,8 @@ struct uv_loop_s {
#undef UV_FS_REQ_PRIVATE_FIELDS
#undef UV_WORK_PRIVATE_FIELDS
#undef UV_FS_EVENT_PRIVATE_FIELDS
#undef UV_LOOP_PRIVATE_FIELDS
#undef UV_LOOP_PRIVATE_PLATFORM_FIELDS

#ifdef __cplusplus
}
Expand Down
6 changes: 3 additions & 3 deletions src/unix/core.c
Expand Up @@ -157,6 +157,7 @@ static int uv__loop_init(uv_loop_t* loop,
#endif
ev_set_userdata(loop->ev, loop);
eio_channel_init(&loop->uv_eio_channel, loop);
uv__loop_platform_init(loop);
return 0;
}

Expand All @@ -179,11 +180,10 @@ uv_loop_t* uv_loop_new(void) {
void uv_loop_delete(uv_loop_t* loop) {
uv_ares_destroy(loop, loop->channel);
ev_loop_destroy(loop->ev);

uv__loop_platform_delete(loop);
#ifndef NDEBUG
memset(loop, 0, sizeof *loop);
memset(loop, -1, sizeof *loop);
#endif

if (loop == default_loop_ptr)
default_loop_ptr = NULL;
else
Expand Down
10 changes: 10 additions & 0 deletions src/unix/internal.h
Expand Up @@ -217,4 +217,14 @@ void uv__fs_event_destroy(uv_fs_event_t* handle);
int uv__make_socketpair(int fds[2], int flags);
int uv__make_pipe(int fds[2], int flags);

#if __linux__
void uv__inotify_loop_init(uv_loop_t* loop);
void uv__inotify_loop_delete(uv_loop_t* loop);
# define uv__loop_platform_init(loop) uv__inotify_loop_init(loop)
# define uv__loop_platform_delete(loop) uv__inotify_loop_delete(loop)
#else
# define uv__loop_platform_init(loop)
# define uv__loop_platform_delete(loop)
#endif

#endif /* UV_UNIX_INTERNAL_H_ */
244 changes: 1 addition & 243 deletions src/unix/linux.c → src/unix/linux/core.c
Expand Up @@ -19,7 +19,7 @@
*/

#include "uv.h"
#include "internal.h"
#include "../internal.h"

#include <stdint.h>
#include <stdio.h>
Expand Down Expand Up @@ -48,91 +48,6 @@
#undef NANOSEC
#define NANOSEC 1000000000

#undef HAVE_INOTIFY_INIT
#undef HAVE_INOTIFY_INIT1
#undef HAVE_INOTIFY_ADD_WATCH
#undef HAVE_INOTIFY_RM_WATCH

#if __NR_inotify_init
# define HAVE_INOTIFY_INIT 1
#endif
#if __NR_inotify_init1
# define HAVE_INOTIFY_INIT1 1
#endif
#if __NR_inotify_add_watch
# define HAVE_INOTIFY_ADD_WATCH 1
#endif
#if __NR_inotify_rm_watch
# define HAVE_INOTIFY_RM_WATCH 1
#endif

#if HAVE_INOTIFY_INIT || HAVE_INOTIFY_INIT1
# undef IN_ACCESS
# undef IN_MODIFY
# undef IN_ATTRIB
# undef IN_CLOSE_WRITE
# undef IN_CLOSE_NOWRITE
# undef IN_OPEN
# undef IN_MOVED_FROM
# undef IN_MOVED_TO
# undef IN_CREATE
# undef IN_DELETE
# undef IN_DELETE_SELF
# undef IN_MOVE_SELF
# define IN_ACCESS 0x001
# define IN_MODIFY 0x002
# define IN_ATTRIB 0x004
# define IN_CLOSE_WRITE 0x008
# define IN_CLOSE_NOWRITE 0x010
# define IN_OPEN 0x020
# define IN_MOVED_FROM 0x040
# define IN_MOVED_TO 0x080
# define IN_CREATE 0x100
# define IN_DELETE 0x200
# define IN_DELETE_SELF 0x400
# define IN_MOVE_SELF 0x800
struct inotify_event {
int32_t wd;
uint32_t mask;
uint32_t cookie;
uint32_t len;
/* char name[0]; */
};
#endif /* HAVE_INOTIFY_INIT || HAVE_INOTIFY_INIT1 */

#undef IN_CLOEXEC
#undef IN_NONBLOCK

#if HAVE_INOTIFY_INIT1
# define IN_CLOEXEC O_CLOEXEC
# define IN_NONBLOCK O_NONBLOCK
#endif /* HAVE_INOTIFY_INIT1 */

#if HAVE_INOTIFY_INIT
inline static int inotify_init(void) {
return syscall(__NR_inotify_init);
}
#endif /* HAVE_INOTIFY_INIT */

#if HAVE_INOTIFY_INIT1
inline static int inotify_init1(int flags) {
return syscall(__NR_inotify_init1, flags);
}
#endif /* HAVE_INOTIFY_INIT1 */

#if HAVE_INOTIFY_ADD_WATCH
inline static int inotify_add_watch(int fd, const char* path, uint32_t mask) {
return syscall(__NR_inotify_add_watch, fd, path, mask);
}
#endif /* HAVE_INOTIFY_ADD_WATCH */

#if HAVE_INOTIFY_RM_WATCH
inline static int inotify_rm_watch(int fd, uint32_t wd) {
return syscall(__NR_inotify_rm_watch, fd, wd);
}
#endif /* HAVE_INOTIFY_RM_WATCH */


static char buf[MAXPATHLEN + 1];

static struct {
Expand All @@ -141,13 +56,6 @@ static struct {
} process_title;


/* Don't look aghast, this is exactly how glibc's basename() works. */
static char* basename_r(const char* path) {
char* s = strrchr(path, '/');
return s ? (s + 1) : (char*)path;
}


/*
* There's probably some way to get time from Linux than gettimeofday(). What
* it is, I don't know.
Expand Down Expand Up @@ -550,153 +458,3 @@ void uv_free_interface_addresses(uv_interface_address_t* addresses,

free(addresses);
}

#if HAVE_INOTIFY_INIT || HAVE_INOTIFY_INIT1

static int new_inotify_fd(void) {
#if HAVE_INOTIFY_INIT1
return inotify_init1(IN_NONBLOCK | IN_CLOEXEC);
#else
int fd;

if ((fd = inotify_init()) == -1)
return -1;

if (uv__cloexec(fd, 1) || uv__nonblock(fd, 1)) {
SAVE_ERRNO(uv__close(fd));
fd = -1;
}

return fd;
#endif
}


static void uv__inotify_read(EV_P_ ev_io* w, int revents) {
struct inotify_event* e;
uv_fs_event_t* handle;
const char* filename;
ssize_t size;
int events;
char *p;
/* needs to be large enough for sizeof(inotify_event) + strlen(filename) */
char buf[4096];

handle = container_of(w, uv_fs_event_t, read_watcher);

do {
do {
size = read(handle->fd, buf, sizeof buf);
}
while (size == -1 && errno == EINTR);

if (size == -1) {
assert(errno == EAGAIN || errno == EWOULDBLOCK);
break;
}

assert(size > 0); /* pre-2.6.21 thing, size=0 == read buffer too small */

/* Now we have one or more inotify_event structs. */
for (p = buf; p < buf + size; p += sizeof(*e) + e->len) {
e = (void*)p;

events = 0;
if (e->mask & (IN_ATTRIB|IN_MODIFY))
events |= UV_CHANGE;
if (e->mask & ~(IN_ATTRIB|IN_MODIFY))
events |= UV_RENAME;

/* inotify does not return the filename when monitoring a single file
* for modifications. Repurpose the filename for API compatibility.
* I'm not convinced this is a good thing, maybe it should go.
*/
filename = e->len ? (const char*) (e + 1) : basename_r(handle->filename);

handle->cb(handle, filename, events, 0);

if (handle->fd == -1)
break;
}
}
while (handle->fd != -1); /* handle might've been closed by callback */
}


int uv_fs_event_init(uv_loop_t* loop,
uv_fs_event_t* handle,
const char* filename,
uv_fs_event_cb cb,
int flags) {
int events;
int fd;

loop->counters.fs_event_init++;

/* We don't support any flags yet. */
assert(!flags);

/*
* TODO share a single inotify fd across the event loop?
* We'll run into fs.inotify.max_user_instances if we
* keep creating new inotify fds.
*/
if ((fd = new_inotify_fd()) == -1) {
uv__set_sys_error(loop, errno);
return -1;
}

events = IN_ATTRIB
| IN_CREATE
| IN_MODIFY
| IN_DELETE
| IN_DELETE_SELF
| IN_MOVED_FROM
| IN_MOVED_TO;

if (inotify_add_watch(fd, filename, events) == -1) {
uv__set_sys_error(loop, errno);
uv__close(fd);
return -1;
}

uv__handle_init(loop, (uv_handle_t*)handle, UV_FS_EVENT);
handle->filename = strdup(filename); /* this should go! */
handle->cb = cb;
handle->fd = fd;

ev_io_init(&handle->read_watcher, uv__inotify_read, fd, EV_READ);
ev_io_start(loop->ev, &handle->read_watcher);
ev_unref(loop->ev);

return 0;
}


void uv__fs_event_destroy(uv_fs_event_t* handle) {
ev_ref(handle->loop->ev);
ev_io_stop(handle->loop->ev, &handle->read_watcher);
uv__close(handle->fd);
handle->fd = -1;
free(handle->filename);
handle->filename = NULL;
}

#else /* !HAVE_INOTIFY_INIT || HAVE_INOTIFY_INIT1 */

int uv_fs_event_init(uv_loop_t* loop,
uv_fs_event_t* handle,
const char* filename,
uv_fs_event_cb cb,
int flags) {
loop->counters.fs_event_init++;
uv__set_sys_error(loop, ENOSYS);
return -1;
}


void uv__fs_event_destroy(uv_fs_event_t* handle) {
UNREACHABLE();
}

#endif /* HAVE_INOTIFY_INIT || HAVE_INOTIFY_INIT1 */

0 comments on commit d3efefb

Please sign in to comment.