sunos: workaround OS bug to prevent fs.watch() from spinning
This is a back-port of commit cfb06db from the master branch.
Fixes joyent/node#3768.
This commit is contained in:
parent
06e03193d5
commit
d6a96de4e3
@ -110,7 +110,7 @@ static void uv__fs_event_rearm(uv_fs_event_t *handle) {
|
||||
|
||||
|
||||
static void uv__fs_event_read(EV_P_ ev_io* w, int revents) {
|
||||
uv_fs_event_t *handle;
|
||||
uv_fs_event_t *handle = NULL;
|
||||
uv_loop_t *loop_;
|
||||
timespec_t timeout;
|
||||
port_event_t pe;
|
||||
@ -120,15 +120,25 @@ static void uv__fs_event_read(EV_P_ ev_io* w, int revents) {
|
||||
loop_ = container_of(w, uv_loop_t, fs_event_watcher);
|
||||
|
||||
do {
|
||||
/* TODO use port_getn() */
|
||||
uint_t n = 1;
|
||||
|
||||
/*
|
||||
* Note that our use of port_getn() here (and not port_get()) is deliberate:
|
||||
* there is a bug in event ports (Sun bug 6456558) whereby a zeroed timeout
|
||||
* causes port_get() to return success instead of ETIME when there aren't
|
||||
* actually any events (!); by using port_getn() in lieu of port_get(),
|
||||
* we can at least workaround the bug by checking for zero returned events
|
||||
* and treating it as we would ETIME.
|
||||
*/
|
||||
do {
|
||||
memset(&timeout, 0, sizeof timeout);
|
||||
r = port_get(loop_->fs_fd, &pe, &timeout);
|
||||
r = port_getn(loop_->fs_fd, &pe, 1, &n, &timeout);
|
||||
}
|
||||
while (r == -1 && errno == EINTR);
|
||||
|
||||
if (r == -1 && errno == ETIME)
|
||||
if ((r == -1 && errno == ETIME) || n == 0)
|
||||
break;
|
||||
|
||||
handle = (uv_fs_event_t *)pe.portev_user;
|
||||
assert((r == 0) && "unexpected port_get() error");
|
||||
|
||||
@ -143,7 +153,7 @@ static void uv__fs_event_read(EV_P_ ev_io* w, int revents) {
|
||||
}
|
||||
while (handle->fd != PORT_DELETED);
|
||||
|
||||
if (handle->fd != PORT_DELETED)
|
||||
if (handle != NULL && handle->fd != PORT_DELETED)
|
||||
uv__fs_event_rearm(handle);
|
||||
}
|
||||
|
||||
|
||||
Loading…
Reference in New Issue
Block a user