| Age | Commit message (Collapse) | Author | Files | Lines |
|---|
|
This works for aarch64 hosts when the target is aarch64, x86_64, or i686,
with only a small #if block in one function that needs to care.
Signed-off-by: Jeremy Drake <cygwin@jdrake.com>
|
|
This makes find_fast_cwd_pointer more resiliant in the face of changes
to the generated code in ntdll.
Signed-off-by: Jeremy Drake <cygwin@jdrake.com>
|
|
This ifdefs out the large table of opcode strings (and the function that
references it) since we're only interested in walking machine code, not
generating disassembly, and makes a couple of other tables "const" so
that they end up in .rdata instead of .data.
Signed-off-by: Jeremy Drake <cygwin@jdrake.com>
|
|
This does not include the source files responsible for generating AT&T-
or Intel-syntax assembly output, or upstream's Makefile.{am,in}.
Signed-off-by: Jeremy Drake <cygwin@jdrake.com>
|
|
This is in preparation for rewriting it using udis86, and adding an
implementation for aarch64 hosts.
Signed-off-by: Jeremy Drake <cygwin@jdrake.com>
|
|
The functions defined by libacl, but not in POSIX.1e draft 17,
are accidentally documented as POSIX functions. Mark them
as libacl functions instead.
Fixes: 004e95cf3594 ("Cygwin: winsup/doc/posix.xml: add missing POSIX.1e functions to docs")
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
|
|
Add the following functions:
acl_get_fd_np
acl_get_link_np
acl_get_perm_np
acl_is_trivial_np
acl_set_fd_np
acl_set_link_np
acl_strip_np
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
|
|
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
|
|
If native DLLs are dlopen'ed, we have to keep track of the load
count, so dlclose doesn't misbehave. We also have to keep track
for the sake of a forked child so the handle returned by dlopen
is still valid in the child.
To accomplish that, add a new type DLL_NATIVE.
Fixes: 82b31085f6ad ("Cygwin: dlfcn: avoid ENOENT on dlcose after dlopen(cygwin1.dll)")
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
|
|
Native (non-cygwin) ninja creates pipe with size == 0, and starts
cygwin process with that pipe. This causes infinite loop in the
fhandler_fifo_pipe::raw_write(). Ideally, the pipe implementation
in cygwin could work even with pipe size == 0, however, it seems
impossible due to:
(1) select() does not work for that pipe because PeekNamedPipe()
always returns 0. Read side is ready to read only when the
write side is about to write, but there is no way to know that.
(2) The cause of the problem:
https://cygwin.com/pipermail/cygwin/2025-January/257143.html
cannot be avoidable. To avoid CancelIo() problem, the patch
https://cygwin.com/pipermail/cygwin-patches/2025q1/013451.html
restricts the data size less than the current pipe space.
However, if pipe size is zero this is impossible.
This patch adds just a workaround for native ninja that avoid
infinite loop in raw_write().
Addresses: https://github.com/msys2/msys2-runtime/issues/270
Reported-by: Christoph Reiter <reiter.christoph@gmail.com>
Co-authored-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Reviewed-by: Corinna Vinschen <corinna@vinschen.de>
Signed-off-by: Takashi Yano <takashi.yano@nifty.ne.jp>
|
|
Previously, opening /dev/ptmx fails without setting errno when it
is opened too many times. With this patch, return EMFILE in that
situation.
Addresses: https://cygwin.com/pipermail/cygwin/2025-March/257786.html
Fixes: 09738c30627c ("Cygwin: pty: setup new pty on opening the master, not in constructor")
Reported-by: Christian Franke <christian.franke@t-online.de>
Signed-off-by: Takashi Yano <takashi.yano@nifty.ne.jp>
|
|
So far, exception::handle returned prematurely if the value of
EXCEPTION_RECORD::ExceptionFlags was non-0.
Starting with Windows 11 we're getting into trouble with that,
if the exception is software generated, for instance by calling
"throw" in a C++ application.
In that case EXCEPTION_SOFTWARE_ORIGINATE (0x80) is set in
EXCEPTION_RECORD::ExceptionFlags.
Change the condition for exiting prematurely to do this only if
any other flag except EXCEPTION_SOFTWARE_ORIGINATE is set in
EXCEPTION_RECORD::ExceptionFlags.
Fixes: Silent change in Windows exception handling
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
|
|
Signed-off-by: Christian Franke <christian.franke@t-online.de>
|
|
While accounts from the BUILTIN, NT AUTHORITY, and NT SERVICE domains
can be owner of a file, they are always treated as group entries if they
show up as additional entrys in a Windows ACL. Consequentially, it
shouldn't be possible to add or remove them as USER entry, for instance,
via setfacl.
Add a check to disallow BUILTIN, NT AUTHORITY, and NT SERVICE accounts
as USER entries in a POSIX ACL.
Fixes: bc444e5aa4ca ("Reapply POSIX ACL changes.")
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
|
|
from scratch
When creating default ACL entries from scratch, set_posix_access
accidentally sets the type to GROUP_OBJ and OTHER_OBJ, rather than
DEF_GROUP_OBJ/DEF_OTHER_OBJ. This shouldn't be noticable most of
time since creating from scratch is seldom necessary.
Fixes: bc444e5aa4ca ("Reapply POSIX ACL changes.")
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
|
|
It's more helpful when reading the source.
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
|
|
Refcounting is performed even if the DLL is loaded at least once
with RTLD_NODELETE. In that case we must not decrement the count
on dlclose, even if FreeLibrary will never unload the DLL.
Use INT_MAX as dll count if RTLD_NODELETE has been used at least once.
Skip any further refcounting and make sure the DLL is still pinned
in memory after fork.
Fixes: 33297d810d90 ("Cygwin: dlfcn: Fix reference counting")
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
|
|
When dlopen'ing the Cygwin DLL, refcounting the number of dlopen
calls fails because dll_list::find returns NULL if the address
resolves to the Cygwin DLL.
Fix this by adding a bool parameter to dll_list::find which allows
find to return the dll entry for the Cygwin DLL for the sake of
dlopen/dlclose/fork.
Fixes: 33297d810d90 ("Cygwin: dlfcn: Fix reference counting")
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
|
|
Introduce a cpu_relax.h header to provide architecture-specific
processor idling instructions for the`cancel3` and `cancel5` tests.
|
|
GCC 15 defaults to `-std=gnu23`, causing build failures in `testsuite`
due to outdated C function declarations. This patch updates the function
declarations to align with modern standards.
|
|
After the commit 0210c77311ae, the context passed to signal handler
cannot be accessed from the signal handler that uses alternate stack.
This is because the context locally copied is on the stack that is
different area from the signal handler uses. With this patch, copy
the context to alternate signal stack area to avoid this situation.
Addresses: https://cygwin.com/pipermail/cygwin/2025-March/257714.html
Fixes: 0210c77311ae ("Cygwin: signal: Use context locally copied in call_signal_handler()")
Reported-by: Bruno Haible <bruno@clisp.org>
Reviewed-by: Corinna Vischen <corinna@vinschen.de>
Signed-off-by: Takashi Yano <takashi.yano@nifty.ne.jp>
|
|
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
|
|
...for completeness. Introduced with Windows Server 2022.
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
|
|
This reverts commit cbe7543cdfdb7f3d270214877d4a4c3e78710bd3.
|
|
x86_64 ABI requires the direction flag in CPU flags register cleared.
https://learn.microsoft.com/en-us/cpp/build/x64-software-conventions
However, currently that flag is not maintained in signal handler.
Therefore, if the signal handler is called when that flag is set, it
destroys the data and may crash if rep instruction is used in the
signal handler. With this patch, the direction flag is cleared in
sigdelayed() by adding cld instruction.
Addresses: https://cygwin.com/pipermail/cygwin/2025-March/257704.html
Fixes: 1fd5e000ace5 ("import winsup-2000-02-17 snapshot")
Reported-by: Christian Franke <Christian.Franke@t-online.de>
Reviewed-by: Corinna Vischen <corinna@vinschen.de>
Signed-off-by: Takashi Yano <takashi.yano@nifty.ne.jp>
|
|
Previously, tty::restore sets the console mode to a predetermined
console mode that is widely common for many non-cygwin console apps.
So, if a non-cygwin app that is started from cygwin process changes
the console mode and executes cygwin sub-process, the console mode
is changed to the predetermined mode rather than being restored the
original mode that is set by the non-cygwin app.
With this patch, the console mode is stored when a cygwin process is
started from non-cygwin app, then tty::restore restores the previous
console mode that is used by the previous non-cygwin app when the
cygwin app exits.
Addresses: https://github.com/msys2/msys2-runtime/issues/268
Fixes: 3312f2d21f13 ("Cygwin: console: Redesign mode set strategy on close().")
Reported-by: Eu Pin Tien, Jeremy Drake <cygwin@jdrake.com>
Signed-off-by: Takashi Yano <takashi.yano@nifty.ne.jp>
|
|
Document including cygwin/acl.h functions when using Solaris ACL functions.
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
|
|
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
|
|
The function sched_setpolicy(2) does not exist.
Fixes: 757424f74400 ("Cygwin: doc: document sched_setpolicy(2) and priority mapping")
Signed-off-by: Christian Franke <christian.franke@t-online.de>
|
|
Fixes: b9ed33dec0cb ("Cygwin: readdir: drop support for NT4/Win2K shares and Samba < 3.0.22")
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
|
|
Due to deficient testing, the current code doesn't return a valid result
to users of sched_getaffinity(). The updated code carries the determined
procmask through to the generation of result cpu mask.
Recognize Windows' limitation that if the process is multi-group (i.e.,
has threads in multiple cpu groups) there is no visibility to which
processors in other groups are being used. One could remedy this by
looping through all the process' threads, but that could be expensive
so is left for future contemplation. In addition, we'd have to maintain
our own copy of each thread's current group and mask in internal overhead.
(By the way, multi-group processes are only possible on Windows systems
with more than 64 hardware threads.)
Reported-by: Christian Franke <Christian.Franke@t-online.de>
Addresses: https://cygwin.com/pipermail/cygwin/2025-March/257616.html
Signed-off-by: Mark Geisert <mark@maxrnd.com>
Fixes: 641ecb07533e ("Cygwin: Implement sched_[gs]etaffinity()")
|
|
Currently when starting a process from bash via a native symlink, argv[0] is
set to the realpath of the executable and not to the link name. This patch fixes
it so the path of the symlink is seen instead.
The cause is a path conversion in perhaps_suffix which follows native
symlinks. Hence the fix this patch uses is to add PC_SYM_NOFOLLOW_REP when
calling path_conv::check to prevent that.
Fixes: be371651146c ("* path.cc (path_conv::check): Don't follow reparse point symlinks if PC_SYM_NOFOLLOW_REP flag is set.")
Signed-off-by: Chris Denton <chris@chrisdenton.dev>
|
|
Keep the functions for backward compatibility, but don't export
them from libcygwin.a for new applications. Drop the workaround
from commit ad29c33229e5 ("Cygwin: stdlib.h: revert declaring
arc4random_stir() as function").
Fixes: ad29c33229e5 ("Cygwin: stdlib.h: revert declaring arc4random_stir() as function")
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
|
|
After the commit 1d97f8aa4385, context_copy in call_signal_handle()
is not used anymore. This patch removes it.
Fixes: 1d97f8aa4385 ("Cygwin: signals: don't evaluate SA_SIGINFO context after handler returns")
Reviewed-by: Corinna Vinschen <corinna@vinschen.de>
Signed-off-by: Takashi Yano <takashi.yano@nifty.ne.jp>
|
|
If the signal handler is called from inside of another signal handler,
_cygtls::context may be destroyed by call_signal_handler() newly called.
To avoid this situation, this patch used context locally copied in
call_signal_handler().
Addresses: https://cygwin.com/pipermail/cygwin-patches/2025q1/013483.html
Fixes: 9043956ce859 ("Only construct ucontext for SA_SIGINFO signal handlers")
Reported-by: Takashi Yano <takashi.yano@nifty.ne.jp>
Reviewed-by: Corinna Vinschen <corinna@vinschen.de>
Signed-off-by: Takashi Yano <takashi.yano@nifty.ne.jp>
|
|
If __SIGFLUSHFAST is sent while the signal queue or the pipe is full,
it causes deadlock. The mechanism is as follows.
When sending __SIGFLUSHFAST, sig_send() waits for response (wakeup
event) from wait_sig(). Therefore, the main thread does not process
signals while waiting the wakeup event. However, if main thread
does not process signals, signals in the queue are not dequeued.
As a result, the sigpacket for __SIGFLUSHFAST stays in the signal
pipe and not being processed. Thus the wakeup event will never be
triggered.
This did not occur with old signal queue design, because the queue
never becomes full. This patch use alternative way for flushing
the queue if the queue and the pipe is full, i.e., sending a event
to trigger flushing in wait_sig(). Do the same for __SIGFLUSH as well.
Addresses: https://cygwin.com/pipermail/cygwin-patches/2025q1/013461.html
Fixes: 8ac22f8f16a8 ("Cygwin: signal: Redesign signal queue handling")
Reported-by: Christian Franke <Christian.Franke@t-online.de>
Reviewed-by: Corinna Vinschen <corinna@vinschen.de>
Signed-off-by: Takashi Yano <takashi.yano@nifty.ne.jp>
|
|
On Linux, SIGSTOP while the process is stopping is ignored. With
this patch, SIGSTOP is processed as a success with no effect while
the process is already stopping.
Fixes: 8f8eeb70ba73 ("* exceptions.cc (sigpacket::process): ...")
Reviewed-by: Corinna Vinschen <corinna@vinschen.de>
Signed-off-by: Takashi Yano <takashi.yano@nifty.ne.jp>
|
|
Currently, sigpacket::process() clears some signals from the queue.
For instance, SIGSTOP clears SIGCONT and SIGCONT clears SIGSTOP, etc.
These might be needed by previous queue design, however, new queue
design does not require that. On the contrary, that causes signal
lost. With this patch, these sig_clear() calls are removed.
Addresses: https://cygwin.com/pipermail/cygwin/2025-March/257582.html
Fixes: 1fd5e000ace5 ("import winsup-2000-02-17 snapshot")
Reported-by: Christian Franke <Christian.Franke@t-online.de>
Reviewed-by: Corinna Vinschen <corinna@vinschen.de>
Signed-off-by: Takashi Yano <takashi.yano@nifty.ne.jp>
|
|
The previous implementation of the signal queue behaves as:
1) Signals in the queue are processed in a disordered manner.
2) If the same signal is already in the queue, new signal is discarded.
Strictly speaking, these behaviours do not violate POSIX. However,
these could be a cause of unexpected behaviour in some software. In
Linux, some important signals such as SIGSTOP/SIGCONT do not seem to
behave like that.
This patch prevents SIGKILL, SIGSTOP, SIGCONT, and SIGRT* from that
issue. Moreover, if SA_SIGINFO is set in sa_flags, the signal is
treated almost as the same.
Addresses: https://cygwin.com/pipermail/cygwin/2025-March/257582.html
Fixes: 7ac6173643b1 ("(pending_signals): New class.")
Reported by: Christian Franke <Christian.Franke@t-online.de>
Reviewed-by: Corinna Vinschen <corinna@vinschen.de>, Christian Franke <Christian.Franke@t-online.de>
Signed-off-by: Takashi Yano <takashi.yano@nifty.ne.jp>
|
|
Commit a942476236b5 ("Cygwin: sigdelayed: pop return address from
signal stack earlier") failed to take two facts into account:
- _cygtls::call_signal_handler() potentially needs the return address
as well, and
- the signal handler may be interrupted by another signal.
Revert the change in sigdelayed() and handle the signal stack manipulation
in _cygtls::call_signal_handler() instead.
Given we're poping the latest addresses from the signal stack early,
there's no need for a big signal stack anymore. Reduce the size of the
stack to 4 entries, plus one dummy entry. Move _cygtls::pop() from
assembler to C++ code and make sure that stackptr neither underflows nor
overflows the signal stack.
Fixes: a942476236b5 ("Cygwin: sigdelayed: pop return address from signal stack earlier")
Co-authored-by: Corinna Vinschen <corinna@vinschen.de>
Signed-off-by: Takashi Yano <takashi.yano@nifty.ne.jp>
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
|
|
The idea to define arc4random_stir() as macro to avoid autoconf
problems in commit a8891c932192 ("Cygwin: stdlib.h: resurrect
arc4random_stir() definition") worked fine for OpenSSH, but not
for libbsd. During the build process the macro shadows the
definition of arc4random_stir() and results in an error message
like this:
error: macro "arc4random_stir" passed 1 arguments, but takes just 0
Give up and just declare arc4random_stir() in a compatible way with
libbsd again.
Fixes: a8891c932192 ("Cygwin: stdlib.h: resurrect arc4random_stir() definition")
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
|
|
When a signal handler is called via sigdelayed, sigdelayed pops
the original address to return to from the signal stack. It only
does so after the signal handler has been called and returned.
If the signal handler never returns, for instance because it calls
longjmp or setcontext, the address stays on the signal stack and
the stack loses one slot. Given the stack has a depth of 256 entries,
no signal handler will be called after 256 signal handlers long
jumping out.
Pop the return address from the signal stack prior to calling the
signal handler and store it in a callee-saved register. This avoids
filling up the signal stack in a scenario like this.
Addresses: https://cygwin.com/pipermail/cygwin/2025-March/257648.html
Fixes: 61522196c715 ("* Merge in cygwin-64bit-branch.")
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
|
|
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
|
|
After return from a signal handler created with the SA_SIGINFO flag,
Cygwin checks if the handler changed the context. If so, Cygwin
erroneously uses the context in a call to setcontext().
Drop this special behaviour. Neither Linux man pages, nor POSIX
documentation treat the context pointer given to the handler as
anything other than a read-only area.
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
|
|
This reverts commit c5f9eed1c045ad05e968c9a9b8fb22b1d46cd8ea.
This was the wrong thing to do because it leads to signal handling
not working after swapcontext.
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
|
|
RtlCaptureContext always returnes a CONTEXT_FULL type context anyway.
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
|
|
GetProcessGroupAffinity() requires PROCESS_QUERY_LIMITED_INFORMATION.
Fixes: 641ecb07533e ("Cygwin: Implement sched_[gs]etaffinity()")
Signed-off-by: Christian Franke <christian.franke@t-online.de>
|
|
The commit 3c1308ed890e adds a guard to stop signal handling on exit()
in call_signal_handler(). However, the signal that is already queued
but does not use signal handler may be going to process even with that
patch.
This patch add one more guard at the begining of sigpacket::process()
to avoid that situation.
Fixes: 3c1308ed890e ("Cygwin: signal: Fix a problem that process hangs on exit")
Reviewed-by: Corinna Vinschen <corinna@vinschen.de>
Signed-off-by: Takashi Yano <takashi.yano@nifty.ne.jp>
|
|
When transferring huge file using scp, the "lost connection" error
sometimes happen. This is due to fhandler_pipe_fifo::raw_write()
accidentally sends data that is not reported in th return value when
interrupted by a signal. The cause of the problem is that CancelIo()
responds success even if NtWriteFile() already sends the data.
The following testcase using plain Win32 APIs reproduces the issue.
The output will be something like:
W: 8589934592
R: 9280061440
Much more data was received than the sender thought it had sent.
#include <windows.h>
#include <stdio.h>
DWORD WINAPI read_thread(LPVOID arg)
{
char buf[32768];
ssize_t sum = 0;
HANDLE pr = (HANDLE) arg;
while (1) {
DWORD nb;
BOOL res = ReadFile(pr, buf, sizeof(buf), &nb, NULL);
if (!res) break;
sum += nb;
}
printf("R: %lld\n", sum);
}
int main()
{
const char pipename[] = "\\\\.\\pipe\\testpipe";
char buf[65536*4] = {0,};
OVERLAPPED ov = {0,};
ssize_t sum = 0;
HANDLE pw = CreateNamedPipe(pipename,
PIPE_ACCESS_OUTBOUND | FILE_FLAG_OVERLAPPED,
PIPE_TYPE_BYTE, 1, 65536, 0, 0, NULL);
HANDLE pr = CreateFile(pipename, GENERIC_READ, 0, NULL, OPEN_EXISTING,
FILE_ATTRIBUTE_NORMAL, NULL);
HANDLE thr = CreateThread(NULL, 0, read_thread, pr, 0, NULL);
ov.hEvent = CreateEvent(NULL, FALSE, FALSE, NULL);
int cnt = 0;
while (sum < 8192LL*1024*1024) {
DWORD nb;
BOOL res = WriteFile(pw, buf, sizeof(buf), &nb, &ov);
if (res) sum += nb;
else if (GetLastError() == ERROR_IO_PENDING) {
if ((cnt & 3) == 0) CancelIo (pw);
res = GetOverlappedResult(pw, &ov, &nb, TRUE);
if (!res && GetLastError() != ERROR_OPERATION_ABORTED) {
printf("%08x\n", GetLastError());
break;
}
sum += nb;
}
cnt++;
}
printf("W: %lld\n", sum);
CloseHandle(pw);
CloseHandle(ov.hEvenet);
WaitForSingleObject(thr, INFINITE);
CloseHandle(pr);
return 0;
}
It seems that CancelIo() works as expected when the NtWriteFile()
is stucked on the pipe full. If the writing is on-going, CancelIo()
reports cancelled, but the data may already be written.
To avoid this issue, this patch waits for completion of writing a
while before issuing CancelIo(), and if it has not completed yet,
call CancelIo().
The wait is done using existing cygwait(select_sem, 10, cw_cancel).
If the semaphore select_sem is released in the read side, the write
operation will be on-going. Otherwise, the write operation is still
stucking, so CancelIo() can safely cancel the I/O.
Addresses: https://cygwin.com/pipermail/cygwin/2025-January/257143.html
Fixes: 4003e3dfa1b9 ("Cygwin: pipes: always terminate async IO in blocking mode")
Reported-by: Jay M Martin <jaymmartin_buy@cox.net>
Reviewed-by: Corinna Vinschen <corinna@vinschen.de>
Signed-off-by: Takashi Yano <takashi.yano@nifty.ne.jp>
|
|
An autoconf test for arc4random_stir() succeeds on Cygwin, because
we export the symbol for backward compatibility. However, when
building with -Werror afterwards, the definition of arc4random_stir()
is missing in the stdlib.h header, so there's a chance to encounter
an error like this with newer GCC:
sshd.c:1113:25: error: implicit declaration of function ‘arc4random_stir’; did you mean ‘arc4random_buf’? [-Wimplicit-function-declaration]
1113 | arc4random_stir();
| ^~~~~~~~~~~~~~~
| arc4random_buf
This occurs for instance when building OpenSSH 9.9p2 with GCC 15.0.
Define arc4random_stir() as an empty macro in cygwin/stdlib.h
to avoid the above compiler warning.
Do not define arc4random_addrandom() yet, in the hope that this function
call is really, really not used anymore.
Fixes: 1cca343e4714 ("Drop redundant arc4random prototypes from cygwin/stdlib.h")
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
|
