kqueue(2) on osx doesn't work (emits EINVAL error) with specific fds
(i.e. /dev/tty, /dev/null, etc). When given such descriptors - start
select(2) watcher thread that will emit io events.
This reverts commit 5da380a5ca.
Contains a bug that effectively makes the select() thread busy-loop. The file
descriptor is polled for both reading and writing, regardless of what events
the main thread wants to receive. Fixing that requires proper synchronization
between the two threads.
See #614.
This reverts commit 209abbab27.
Fixes the following SIGSEGV:
(gdb) f 1
#1 0x00007fc084683aec in uv__async_io (loop=0x7fc0848e0b40,
handle=0x7fc0848e0c78, events=1) at src/unix/async.c:175
175 ASYNC_CB(h)
(gdb) list
170
171 /* If we need to sweep all handles anyway - skip this loop */
172 if (!loop->async_sweep_needed) {
173 for (i = 0; i < end; i += sizeof(h)) {
174 h = *((uv_async_t**) (buf + i));
175 ASYNC_CB(h)
176 }
177 }
178
179 bytes -= end;
(gdb) print *h
$1 = {close_cb = 0x184e1b0, data = 0x18d9520, loop = 0x7fc0848e0b40,
type = 49, handle_queue = {prev = 0x18dae10, next = 0x7860c0}, flags = 32,
next_closing = 0x1863b40, pending = 0, async_cb = 0x31,
queue = {prev = 0x18dae50, next = 0x7860c0}}
(gdb)
It looks like the async handle gets closed or otherwise becomes invalid before
the sweep is executed.
Fixes#603.
Benchmarks demonstrated that the idle timer handle approach didn't balance the
load quite fair enough, the majority of new connections still ended up in one
or two processes.
The new approach voluntarily gives up a scheduler timeslice by calling
nanosleep() with a one nanosecond timeout.
Why not sched_yield()? Because on Linux (and this is probably true for other
Unices as well), sched_yield() only yields if there are other processes running
on the same CPU.
nanosleep() on the other hand always forces the process to sleep, which gives
other processes a chance to accept our pending connections.
Implement a best effort approach to mitigating accept() EMFILE errors.
We have a spare file descriptor stashed away that we close to get below
the EMFILE limit. Next, we accept all pending connections and close them
immediately to signal the clients that we're overloaded - and we are, but
we still keep on trucking.
There is one caveat: it's not reliable in a multi-threaded environment.
The file descriptor limit is per process. Our party trick fails if another
thread opens a file or creates a socket in the time window between us
calling close() and accept().
Fixes#315.
kqueue(2) on osx doesn't work (emits EINVAL error) with specific fds
(i.e. /dev/tty, /dev/null, etc). When given such descriptors - start
select(2) watcher thread that will emit io events.
OS X has no public API for fdatasync. And as pointed out in `man fsync(2)`:
For applications that require tighter guarantees about the integrity of
their data, Mac OS X provides the F_FULLFSYNC fcntl. The F_FULLFSYNC
fcntl asks the drive to flush all buffered data to permanent storage.
Applications, such as databases, that require a strict ordering of writes
should use F_FULLFSYNC to ensure that their data is written in the order
they expect. Please see fcntl(2) for more detail.
Problem: registering two uv_fs_event_t watchers for the same path, then closing
them, caused a segmentation fault. While active, the watchers didn't work right
either, only one would receive events.
Cause: each watcher has a wd (watch descriptor) that's used as its key in a
binary tree. When you call inotify_watch_add() twice with the same path, the
second call doesn't return a new wd - it returns the existing one. That in turn
resulted in the first handle getting ousted from the binary tree, leaving
dangling pointers.
This commit addresses that by storing the watchers in a queue and storing the
queue in the binary tree instead of storing the watchers directly in the tree.
Fixesjoyent/node#3789.
* the callback gets called only once on error, not repeatedly...
* ...unless the error reason changes from e.g. UV_ENOENT to UV_EACCES
* the callback receives pointers to uv_statbuf_t objects so it can inspect what
changed
* replace libev backed timers with a pure libuv implementation
* gut ev_run() and make it take a timeout instead of flags
Incidentally speeds up the loop_count_timed benchmark by about 100%.
Makes the uv__io code a little more obscure but has the advantage that
sizeof(uv__io_t) == sizeof(ev_io), i.e. the sizes of embedding handles
don't change.
The new idle watcher was temporarily disabled in 073a48d due to some semantic
incompatibilities with the previous implementation. This commit resolves those
issues and reactivates the new implementation.
One outstanding bug is that idle watchers can run in a different order
(relative to other handle types) than the old implementation, e.g. (timer, idle)
instead of the expected (idle, timer). This will be fixed in an upcoming commit.
This commit changes how the event loop determines if it needs to stay alive.
Previously, an internal counter was increased whenever a handle got created
and decreased again when the handle was closed.
While conceptually simple, it turned out hard to work with: you often want
to keep the event loop alive only if the handle is actually doing something.
Stopped or inactive handles were a frequent source of hanging event loops.
That's why this commit changes the reference counting scheme to a model where
a handle only references the event loop when it's active. 'Active' means
different things for different handle types, e.g.:
* timers: ticking
* sockets: reading, writing or listening
* processes: always active (for now, subject to change)
* idle, check, prepare: only active when started
This commit also changes how the uv_ref() and uv_unref() functions work: they
now operate on the level of individual handles, not the whole event loop.
The Windows implementation was done by Bert Belder.
Instead of using one port per watch, use one port for all the watches.
This is a cherry-pick of commit 7326962 from v0.6 into master.
Conflicts:
include/uv-private/uv-unix.h
src/unix/core.c
src/unix/sunos.c
