Split this off from endgame, so that it can be handled separately and
earlier, rather than trying to detect inside endgame which case we are
in. There appear to be some race conditions on the `handle` field still
however, which this does not yet attempt to address.
If `uv_close` was called while a connect was pending, we would fail to
release the resources for the connection, since we had not yet set the
type of the struct.
Fix a thread data-race on slow connect path code: only permitted to
write to `req` on threads, as anything else causes data race
corruption.
There seemed be a small variety of other resource management bugs in
edge cases, which turned out to make this a lot larger than initially
expected.
Refs: https://github.com/libuv/libuv/pull/3598#issuecomment-1111513567
On some streams (notably TTYs), it is permitted to continue reading
after getting EOF. So still stop reading on EOF, but allow the user to
reset the stream and try to read again (which may just get EOF).
This relaxes the constraint added in ce15b8405e.
Refs: https://github.com/libuv/libuv/pull/3006
This assertion was added when req->write_buffer was a pointer. It was
then checking that write_buffer itself was not NULL. Checking that .base
is not NULL is superfluous because WriteFile will return error 998
(ERROR_NO_ACCESS) if the input buffer is invalid. This assertion fires
on zero-length writes when base==NULL&&len==0.
Equivalents of `pipe` and `socketpair` for cross-platform use.
PR-URL: https://github.com/libuv/libuv/pull/2953
Reviewed-By: Santiago Gimeno <santiago.gimeno@gmail.com>
Previously, this field was being used uninitialized, leading to the
uv_write call returning failure. This only happens in the rare fallback
path where UV_HANDLE_EMULATE_IOCP is set, meaning that the OVERLAPPED
opened pipe already was associated with an IOCP.
Fixes: https://github.com/libuv/libuv/issues/629
PR-URL: https://github.com/libuv/libuv/pull/2620
Reviewed-By: Saúl Ibarra Corretgé <s@saghul.net>
Guard against sending the handle over the UNIX domain socket twice
when the first sendmsg() didn't write all bytes.
The changes to src/win partially undo changes made earlier this year,
see the referenced pull request for details.
Libuv never made promises about the value of `req->send_handle` at
different points in time so this should be a safe, non-breaking change.
No tests because this particular condition is hard to hit reliably
across platforms. I spent a lot of time trying to write one but it
turned out hideously complex, and worse, flaky.
Fixes: https://github.com/libuv/libuv/issues/2086
PR-URL: https://github.com/libuv/libuv/pull/2097
Refs: https://github.com/libuv/libuv/pull/1843
Reviewed-By: Santiago Gimeno <santiago.gimeno@gmail.com>
Some long overdue refactoring that unifies more of the UNIX and Windows
backends.
PR-URL: https://github.com/libuv/libuv/pull/1904
Reviewed-By: Colin Ihrig <cjihrig@gmail.com>
Reviewed-By: Saúl Ibarra Corretgé <saghul@gmail.com>
This macro is not guaranteed to be present in stdlib.h and the
macro may be disabled on Windows in some situations (e.g. when
NOMINMAX is defined).
PR-URL: https://github.com/libuv/libuv/pull/1891
Reviewed-By: Colin Ihrig <cjihrig@gmail.com>
Reviewed-By: Bartosz Sosnowski <bartosz@janeasystems.com>
Reviewed-By: Ben Noordhuis <info@bnoordhuis.nl>
* Don't silently ignore ERROR_OPERATION_ABORTED.
Code to silently ignore this error was added in c42a4ca, under the
false premise that this error is somehow equivalent to EINTR on posix
platforms. This isn't true; ERROR_OPERATION_ABORTED doesn't happen
unless the application explicitly aborts an I/O operation.
Silently ignoring this error elsewhere could potentially hide bugs,
hence libuv shouldn't do it. Instead, explicitly deal with it where
it is expected.
* Don't mark aborted reads as successful.
The worker thread used to call ReadFile() on synchronous pipes would
incorrectly mark cancelled read requests as successful, leading to
issues later on.
* Rely on main thread to restart aborted reads.
After a blocking ReadFile() call was cancelled, the worker thread
would previously attempt to restart it immediately, making
synchronization logic needlessly complex. Instead, we simply cancel
the operation, and leave it to the main loop to restart it if so
desired.
Since we now realy on the main thread to restart interrupted
ReadFile() calls, we can now have a single function interrupts a
synchronous read until the event loop restarts it again.
* Clean up uv__pipe_read_stop().
A single function to interrupt blocking reads also allows us to
remove weird logic in uv__pipe_read_stop() that quickly pauses and
unpauses a read operation in the hope of cancelling it.
* Assume CancelIo() and CancelSynchronousIo() are always available.
Since libuv doesn't support windows XP and Server 2003 any more, we
can assume that these APIs are always available, and do away with
branching around them.
PR-URL: https://github.com/libuv/libuv/pull/1843
Reviewed-By: Anna Henningsen <anna@addaleax.net>
Reviewed-By: Bartosz Sosnowski <bartosz@janeasystems.com>
Reviewed-By: Gireesh Punathil <gpunathi@in.ibm.com>
Back in the day I wrote comments in a really unusual way. Nowadays it
makes my eyes bleed, and clang-format doesn't know how to deal with it.
PR-URL: https://github.com/libuv/libuv/pull/1853
Reviewed-By: Bartosz Sosnowski <bartosz@janeasystems.com>
Reviewed-By: Ben Noordhuis <info@bnoordhuis.nl>
Reviewed-By: Colin Ihrig <cjihrig@gmail.com>
Reviewed-By: Gireesh Punathil <gpunathi@in.ibm.com>
- Remove the `active_reqs` queue, which is never used. There are more
efficient per-stream queues that `libuv` uses whenever it needs this
information, so duplicating it and managing it here seems
like unnecessary extra space and work.
- Unix `uv_loop_init` didn't explicitly initialize.
`loop->active_handles` (although it did memset the whole struct
to 0, so it wasn't wrong previously, just inconsistent).
- Consolidate repeated code for `uv__has_active_reqs`.
- Change `uv__loop_alive` to use the helper functions (mirroring the
unix copy of the same function).
- Initialize some more uv_stream_t fields in init, rather than waiting
for the connection callback. This helps surface bugs in libuv or the
caller better, since it ensures libuv doesn't see uninitialized
memory if asked to look at these fields before it thinks the socket
is connected.
- Fixes what appears to be a double-counting of `handle->reqs_pending`,
in the highly-unlikely event that the code wants to emulate IOCP,
but `RegisterWaitForSingleObject` somehow managed to fail.
PR-URL: https://github.com/libuv/libuv/pull/1746
Reviewed-By: Ben Noordhuis <info@bnoordhuis.nl>
Reviewed-By: Colin Ihrig <cjihrig@gmail.com>
Reviewed-By: Santiago Gimeno <santiago.gimeno@gmail.com>
Cross-compiling with mingw64 was broken because the header files are
called `accctrl.h` and `aclapi.h` whereas libuv was trying to include
`AccCtrl.h` and `AclAPI.h`.
PR-URL: https://github.com/libuv/libuv/pull/1629
Reviewed-By: Colin Ihrig <cjihrig@gmail.com>
Adds new uv_pipe_chmod function which can be used to make the pipe
writable or readable by all users.
PR-URL: https://github.com/libuv/libuv/pull/1386
Reviewed-By: Santiago Gimeno <santiago.gimeno@gmail.com>
Reviewed-By: Colin Ihrig <cjihrig@gmail.com>
Some functions, such as uv_fs_scandir() can be called with
a NULL loop argument. In this case, it's possible that the
Windows API hasn't been initialized, leading to a crash.
This commit adds additional uv__once_init() calls to mitigate
this situation.
Fixes: https://github.com/libuv/libuv/issues/1488
PR-URL: https://github.com/libuv/libuv/pull/1512
Reviewed-By: Colin Ihrig <cjihrig@gmail.com>
Reviewed-By: Santiago Gimeno <santiago.gimeno@gmail.com>
Fixes a race condition where if uv_read_stop was called shortly
after uv_read_start or a successful read and before the
uv_pipe_zero_readfile_thread_proc thread started,
that thread would call the blocking ReadFile call after the
HANDLE_READING flag had already been cleared.
Also ignores EINTR to be more consistent with unix
(although we generally don't expect to see this condition on windows).
PR-URL: https://github.com/libuv/libuv/pull/1322
Reviewed-By: Saúl Ibarra Corretgé <saghul@gmail.com>
Reviewed-By: Bartosz Sosnowski <bartosz@janeasystems.com>
Fixes a TODO in src/threadpool.c. Updates the Windows code to drop the
unused `loop` parameter in calls to uv_req_init().
PR-URL: https://github.com/libuv/libuv/pull/1091
Reviewed-By: Colin Ihrig <cjihrig@gmail.com>
Reviewed-By: Santiago Gimeno <santiago.gimeno@gmail.com>
Reviewed-By: Saúl Ibarra Corretgé <saghul@gmail.com>
Treat both the base being NULL or the length being 0 as ENOBUFS.
PR-URL: https://github.com/libuv/libuv/pull/997
Reviewed-By: Ben Noordhuis <info@bnoordhuis.nl>
If an API function returns UV_ENOBUFS make sure we count the terminating
null, which we need space for. In case of success the null byte is not
included in the count, but the buffer *is* null terminated.
PR-URL: https://github.com/libuv/libuv/pull/690
Reviewed-By: Colin Ihrig <cjihrig@gmail.com>
libuv has multiple functions which return buffers. Make them consistent
with the following rules: the returned size *does not* include the null
byte, but the buffer *is* null terminated.
There is only one exception to the above: Linux abstract sockets,
because null bytes are not used as string terminators in those.
Refs: https://github.com/libuv/libuv/pull/674
PR-URL: https://github.com/libuv/libuv/pull/690
Reviewed-By: Colin Ihrig <cjihrig@gmail.com>
On Windows the pipe implementation could read more from a stream than
was available and it would create an assertion failure. This change
will make it so we read the minimum of the available data or the length
of the data.
To test this, I took the existing ipc_send_recv_tcp test and updated it
to do two writes and two read on each side of the pipe since that was the
reproduction recipe used by the reporter. This approach reproduced the
issue on Windows and the committed fix resolved the issue.
Fixes: https://github.com/libuv/libuv/issues/505
PR-URL: https://github.com/libuv/libuv/pull/549
Reviewed-By: Saúl Ibarra Corretgé <saghul@gmail.com>
When duplicating the socket handle being sent the target process
id is defaulted to the current process id. This enables uv_write2
to be used for thread-clustering in addition to process-clustering on
multi-threaded programming languages.
The ipc tests are updated to run in two modes. In the _inproc mode
the echo work is done on a second thread instead of in a second
process.
An internal function int uv_current_pid() is added to the windows
specific code which caches the value of GetCurrentProcessId(). This
means uv_write2 does not call GetCurrentProcessId() every inprocess
send.
Refs: https://github.com/joyent/libuv/issues/926
PR-URL: https://github.com/libuv/libuv/pull/540
Reviewed-By: Saúl Ibarra Corretgé <saghul@gmail.com>
It's an error to do so, so just ignore it. The test would cause an
invalid memory access if the fix is undone.
PR-URL: https://github.com/libuv/libuv/pull/488
Reviewed-By: Ben Noordhuis <info@bnoordhuis.nl>
On Windows we create multiple pipe handles (system handles) which are
attached to pending accept requests. Each of these will take turns in
replacing the reference in handle->handle, so make sure we allow for
that **only** for pipe servers.
PR-URL: https://github.com/libuv/libuv/pull/488
Reviewed-By: Ben Noordhuis <info@bnoordhuis.nl>
