Rather than waiting for signalfd_select_wait in a thread, which is racy,
create a global event "my_pendingsigs_evt" which is set and reset by
wait_sig depending only on the fact if blocked signals are pending or not.
This in turn allows to WFMO on this event in select as soon as signalfds
are present in the read descriptor set. Select's peek and verify
will then check if one of the present signalfds is affected.
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
Borrow a trick from strace to lessen occurrences of "??? => ???" in ldd
output. Specifically, if the module name isn't found in the usual place
in the mapped image, use the file handle we have to look up the name.
On sigwaitinfo or reading from a signalfd, signal processing sets up
signal handling via sigdelayed even if the handler address is NULL.
This doesn't have any impact on sigwaitinfo scenarios (or at least, I
wasn't able to come up with a reproducer) but it breaks signalfd
scenarios, where eventually a call to call_signal_handler from
sigdelayed will try to call the NULL function.
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
- If screen is alternated on console, cursor position is not restored
correctly in the case of xterm compatible mode is enabled. For example,
the shell prompt is shown at incorrect position after using vim.
This patch fixes this problem.
- The horizontal tab positions are broken after resizing console window.
This seems to be a bug of xterm compatible mode of windows console.
This workaround fixes this problem.
The default implementation of the fenv.h methods return
-EOPNOTSUP. Some of these have implementations appropriate
for soft-float.
The intention of the new fenv.h is that it be portable
and that architectures provide their own implementation
of sys/fenv.h.
This avoids collisions of shmat maps with Windows own datastructures
when allocating top-down.
This patch moves the mmap_allocator class definition into its
own files and just uses it from mmap and shmat.
Adding default winvar 'PATH=C:\cygwin64\binZ' to an environment that is
already allocated for 'SYSTEMROOT=ZWINDIR=Z', we need to count that
trailing (Z)ero as well. Otherwise we trigger this assertion failure:
$ /bin/env -i SYSTEMROOT= WINDIR= /bin/env
assertion "(s - envblock) <= tl" failed: file "/home/corinna/src/cygwin/cygwin-3.0.7/cygwin-3.0.7-1.x86_64/src/newlib-cygwin/winsup/cygwin/environ.cc", line 1302, function: char** build_env(const char* const*, WCHAR*&, int&, bool, HANDLE)
Aborted (core dumped)
When the exec family of functions is called for a script-like
file, the av::setup function handles the exec[vl]p case as
well. The execve case for files not starting with a she-bang
is handled first by returning ENOEXEC. Only after that, the
file's executability is checked.
This leads to the problem that ENOEXEC is returned for non-executable
files as well. A calling shell interprets this as a file it should try
to run as script. This is not desired for non-executable files.
Fix this problem by checking the file for executability first. Only
after that, follow the other potential code paths.
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
This patch supplies an implementation of the CPU_SET(3) processor
affinity macros as documented on the relevant Linux man page.
There is a mostly superset implementation of cpusets under newlib's
libc/sys/RTEMS/include/sys that has Linux and FreeBSD compatibility
and is built on top of FreeBSD bitsets. This Cygwin implementation
and the RTEMS one could be combined if desired at some future point.
Do not attach to the child before it was successfully initialized, or we
would need more sophisticated cleanup on child initialization failure,
like suppressing SIGCHILD delivery with multiple threads ("waitproc")
involved.
Improves "Cygwin: fork: Remember child not before success.",
commit f03ea8e1c5, which leads to fork
problems if cygserver is running:
https://cygwin.com/ml/cygwin-patches/2019-q2/msg00155.html
During fork, the child process requires the process table to be
initialized for fixup_shms_after_fork, while still allowing subsequent
dlls.load_after_fork to fail silently (for when the "forkable" hardlinks
are not created yet).
pinfo::remember not performing reattach anymore requires explicit
pinfo::reattach now where appropriate.
Prepares to improve "Cygwin: fork: Remember child not before success."
commit f03ea8e1c5, which leads to fork
problems if cygserver is running:
https://cygwin.com/ml/cygwin-patches/2019-q2/msg00155.html
Introducing an independent Cygwin PID introduced a regression:
The expectation is that the myself pinfo pointer always points to a
specific address right in front of the loaded Cygwin DLL.
However, the independent Cygwin PID changes broke this. To create
myself at the right address requires to call init with h0 set to
INVALID_HANDLE_VALUE or an existing address:
void
pinfo::init (pid_t n, DWORD flag, HANDLE h0)
{
[...]
if (!h0 || myself.h)
[...]
else
{
shloc = SH_MYSELF;
if (h0 == INVALID_HANDLE_VALUE) <-- !!!
h0 = NULL;
}
The aforementioned commits changed that so h0 was always NULL, this way
creating myself at an arbitrary address.
This patch makes sure to set the handle to INVALID_HANDLE_VALUE again
when creating a new process, so init knows that myself has to be created
in the right spot. While at it, fix a potential uninitialized handle
value in child_info_spawn::handle_spawn.
Fixes: b5e1003722 ("Cygwin: processes: use dedicated Cygwin PID rather than Windows PID")
Fixes: 88605243a1 ("Cygwin: fix child getting another pid after spawnve")
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
_pinfo::set_ctty sets myself's pgid to the ctty pgid if the process has
been started from a non-Cygwin process. This isn't the right thing to
do when started from GDB. GDB starts the application via standard
Windows means, not via Cygwin fork/exec, so it's treated as being
a non-Cygwin parent.
But we want the app running in it's own process group. So skip this
step when running under a debugger
Signed-off-by: Corinna Vinschen <corinna-cygwin@cygwin.com>
Return -1 with EINVAL if pgid < 0.
Previously tcsetpgrp() would blindly go ahead and set the pgid of the
controlling terminal to a negative value, causing later calls to
various functions to fail.
For example, gdb has code like the following:
tcsetpgrp (0, getpgid (inf->pid));
If getpgid (inf->pid) fails (returns -1), then this code would set the
pgid of fd 0 to -1, so that some later calls to getpgid() would also
return -1. This caused the problem reported here:
https://cygwin.com/ml/cygwin/2019-07/msg00166.html.
So far negative values were denoting files, positive values
denoting directories. We should prefer a less error prone
method. Redefine virtual_ftype_t to contain only positive
values and replace checks for negativ or positive values with
inline functions virt_ftype_isfile() and virt_ftype_isdir().
Drop outdcated comments referring to numerical virtual_ftype_t
values.
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
commit 2607639992 "Improve error handling in /proc/[pid]/
virtual files." changed the return value of the /proc/PID
formatting functions to return -1 instead of 0 in the error
case to allow a filesize of 0.
The patch neglected to change this behaviour for /proc/PID/fd
content. This patch fixes that.
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
Change path_conv::is_lnk_special() so that it returns false on socket
files.
is_lnk_special() is called by rename2() in order to deal with special
files (FIFOs and symlinks, for example) whose Win32 names usually have
a ".lnk" suffix. Socket files do not fall into this category, and
this change prevents ".lnk" from being appended erroneously when such
files are renamed.
Remove a now redundant !pc.issocket() from fhandler_disk_file::link().
It is used only once, and the name is supposed to suggest "device that
is not based on the filesystem". This intended meaning is clearer if
we just replace is_auto_device() by its definition at the place where
it's used.
The sigpending mechanism failed to check if the pending signal was a
process-wide signal, or a signal for the curent thread. Fix that by
adding a matching conditional to wait_sig's __SIGPENDING code.
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
So far sig_send's return type is int. The problem with this is
that sig_send returns a sigset_t on __SIGPENDING, and sigset_t
is defined as long type. So the function only returns the lower
32 bit of sigset_t, which is fine on 32 bit, but casts away the
pending RT signals on 64 bit.
Fix this by changing the return type of sig_send to sigset_t, so
as not to narrow down the sigset when returning from handling
__SIGPENDING. Make sure to cast correctly in all invocations
of sig_send.
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
Have sched_getaffinity() interface like glibc's, and provide an
undocumented internal interface __sched_getaffinity_sys() like the Linux
kernel's sched_getaffinity() for benefit of taskset(1).
Add a function timerfd_tracker::enter_critical_section_cancelable,
which is like enter_critical_section but honors a cancel event. Call
this when a timer expires while the timerfd thread is in its inner
loop. This avoids a deadlock if timerfd_tracker::dtor has entered its
critical section and is trying to cancel the thread. See
https://cygwin.com/ml/cygwin/2019-06/msg00096.html.
This patch set implements the Linux syscalls sched_getaffinity,
sched_setaffinity, pthread_getaffinity_np, and pthread_setaffinity_np.
Linux has a straightforward view of the cpu sets used in affinity masks.
They are simply long (1024-bit) bit masks. This code emulates that view
while internally dealing with Windows' distribution of available CPUs among
processor groups.
Make it a manual reset event. It's only used once to allow a reader
to open, and there's no reason to ever reset it. Defensively set it
when a client connection is recorded, even though it should be set by
the writer that connected.
Call NtReadFile directly instead of calling fhandler_base::raw_read.
In addition to being simpler, this gives us access to the return value
from NtReadFile.
Add a method fifo_client_handler::pipe_state that queries Windows for
the state of a pipe instance. Use this to help terminate the
listen_client thread cleanly.
If the last client handler is useless, delete it instead of declaring
it invalid.
fhandler_fifo::close could be called from a signal handler or another
thread at a time when another function is holding the fifo_client
lock. This could prevent the listen_client thread from acting on the
thread termination event. Avoid a deadlock by calling
fifo_client_unlock at the beginning of fhandler_fifo::close.
libX11 provides <X11/Xlocale.h>. The build of libX11 itself adds
include/X11 to the compiler's include path. This results in a name
collision with /usr/include/xlocale.h on case-insensitive filesystems.
Commit 90e35b1eb3 renamed sys/_locale.h to xlocale.h in March 2017 under
the assumption that we should provide the locale_t type in the same file
as on Linux, FreeBSD, and Darwin.
A few weeks later (June 2017), glibc removed the xlocale.h file in favor
of bits/types/locale_t.h, which shouldn't be included directly anyway.
For reference and the reasoning, see
https://sourceware.org/git/?p=glibc.git;a=commit;h=f0be25b6336d
Given the above, revert 90e35b1eb3 and
fix additional usage of xlocale.h.
32 bit Cygwin performs a POSIX-compatible mapping after EOF which
is not supported in this form on Windows. The 64 bit Windows
kernel never supported the AT_ROUND_TO_PAGE mapping flag, so we
couldn't page-aligned map the space right after the file's EOF.
So mapping beyond EOF was disabled in 64 bit Windows and WOW64.
However, if mmap works, a matching munmap should work as well,
*and* it should not accidentally unmap unrelated memory.
Therefore we enable mapping beyond EOF on 64 bit as well. Since
that mapping is always 64K aligned, the are between the last file
page and the next 64K allocation boundary will be unallocated.
There's no way around that.
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
When using NtCreateFile when creating a directory that already exists,
it will correctly return 'STATUS_OBJECT_NAME_COLLISION'.
However using this function to create a directory (and all its parents)
a normal use would be to start with mkdir(‘/cygdrive/c’) which translates
to ‘C:\’ for which it'll instead return ‘STATUS_ACCESS_DENIED’.
An increasing number of tools are being included in Windows which have the
same names as those included in Cygwin packages. Indicating which one is
first in PATH can be helpful in diagnosing behavioural discrepencies
between them.
Also, fix the alphabetization of ssh.
64 bit Windows started out with a 44 bit address space due to a
restriction of the AMD64 CPUs at the time. Starting with Windows
8.1, these CPUs are not supported anymore and Windows switched to
the full 48 bit address space supported by AMD64.
Cygwin didn't follow suit yet so mmaps are still restricted to
the lower 44 bit address space. Fix that by using a system-specific
upper address for mmap allocations, 44 bit up to Windows 8, 48 bit
starting with Windows 8.1.
While at it, move the heap by another 8 Gigs to leave some space
for a potential extension of DLL address space, and restrict the
mmap lower address so the heap can grow to 32 Gigs before colliding
with mmaps.
Do not remember the child before it was successfully initialized, or we
would need more sophisticated cleanup on child initialization failure,
like cleaning up the process table and suppressing SIGCHILD delivery
with multiple threads ("waitproc") involved. Compared to that, the
potential slowdown due to an extra yield () call should be negligible.
Pause the child process after performing fork fixups even if there were
no dynamically loaded dlls with extra data/bss transfers to wait for.
This allows the parent process to cancel the current fork call even if
the child process was successfully initialized already.
This is a preparation for when the parent does remember the child no
earlier than after successful child initialization.
Querying the ntlength and existence of the /var/run/cygfork directory in
the very first Cygwin process should not use nt_max_path_buf, as that
one is used by dll_list::alloc already.
NtQueryVirtualMemory for MemorySectionName may return some old path even
if the process was just started, for when some directory in between was
renamed - maybe because the NT file cache is hot for the old path still.
This was seen during gcc bootstrap, returning a MemorySectionName of
".../gcc/xgcc.exe" even if started as ".../prev-gcc/xgcc.exe", where the
directory rename from "gcc" to "prev-gcc" was done the moment before.
As we stat the module's real file right after loading now, there is no
point in using NtQueryVirtualMemory with MemorySectionName any more, and
we can use what GetModuleFileName returned instead.
-Add vendor identification
-Support in get_cpu_cache
Background:
Chengdu Haiguang IC Design Co., Ltd (Hygon) is a Joint Venture
between AMD and Haiguang Information Technology Co.,Ltd., aims at
providing high performance x86 processor for China server market.
Its first generation processor codename is Dhyana, which
originates from AMD technology and shares most of the
architecture with AMD's family 17h, but with different CPU Vendor
ID("HygonGenuine")/Family series number(Family 18h).
Related Hygon kernel patch can be found on:
http://lkml.kernel.org/r/5ce86123a7b9dad925ac583d88d2f921040e859b.1538583282.git.puwen@hygon.cn
Signed-off-by: Jinke Fan <fanjinke@hygon.cn>
There's no longer a need to consider the connect_evt after fork/exec.
After stopping the listen_client thread, all client handlers should be
in the fc_connected or fc_invalid states, so their connect_evt members
won't be used again.
Also remove code in fhandler_fifo::dup that just repeats things
already done in stop_listen_client.
Add a method fhandler_fifo::check_listen_client_thread that checks
whether the thread is running. Use it in raw_read instead of just
testing the handle listen_client_thr.
On exit from the listen_client thread, make sure there's no pending
FSCTL_PIPE_LISTEN request. Otherwise we might get a client connection
after restarting the thread, and we won't have a handle for
communicating with that client.
Remove the retry loop in the case of STATUS_PIPE_LISTENING; that case
shouldn't occur.
Remove the now-unused fc_connecting value from
fifo_client_connect_state.
It's now up to the caller to pass a handle to open_pipe and, if
desired, to call set_handle on return.
This will be useful for a future commit, in which we will open a
client connection without setting an io_handle.
When opening a duplexer, open a client connection to the first client
handler. Previously we gave the duplexer a bogus write handle, which
was just a duplicate of the first client handler's handle. This meant
that we had a pipe server with no clients connected, and all I/O
attempts failed with STATUS_PIPE_LISTENING.
Extend the last fcntl change to duplexers.
Remove a now unused fifo_client_handler constructor, as well as the
long unusued method fifo_client_handler::connect.
Don't create the pipe in duplex mode; the server handle will only be
used for reading.
NtQueryVirtualMemory for MemorySectionName does not reliable return the
changed dll file name when another process does move the file around,
and we may end up creating forkable hardlinks to wrong dll files.
So query the file id when loading the dll rather than before fork.
Make stat_real_file_once a method of struct dll, to be more flexible on
where to use. Also, debug print memory section name queried for a dll.
This is a preparation to query the file id when loading the dll.
The blocking mode of the Windows pipe underlying a writer is set to
match that of the writer itself when the latter is opened. Define
fhandler_fifo::fcntl to keep the pipe and the writer in sync if the
blocking mode is changed via fcntl.
Use 'waitret' instead of 'ret' for the return value of cygwait, since
there is already a different 'ret' variable in use. The previous
double use of 'ret' was legal because of scoping rules, but possibly
confusing.
Commit debdfb8b, "Cygwin: AF_UNIX: Rework internal methods to return
HANDLEs", removed a set_handle call. As a result, the attempt to
exchange socket info fails with "Invalid Handle".
Fix this by moving the send_sock_info call into a new 'xchg_sock_info'
method, which is then invoked by the callers of open_pipe if
necessary, after they have called set_handle.
Also, call recv_peer_info in addition to send_sock_info when
exchanging socket information.
Remove the last argument of open_pipe, which is no longer needed.
Adjust all callers.
Martin Storsjö
Corinna Vinschen
Mark Geisert
Takashi Yano
Joel Sherrill
Lavrentiev, Anton
Michael Haubenwallner
Corinna Vinschen
Ken Brown
Ben Wijen
Yaakov Selkowitz
Jinke Fan