A hang was encountered, apparently triggered by commit 63b503916d,
changing tls_pathbufs from malloc'ed to HeapAlloc'ed memory. After
lengthy debugging it transpired that adding the heap handle to the
tls_pathbuf struct added 8 bytes to the cygtls area, thus moving
the "context" member by 8 bytes, too, so it was suddently unaligned.
Fix this for now by changing the alignment.
Fix this once and for all, by adding code to the gentls_offsets script
to check if the alignment of the "context" member is 16 bytes. If not,
print a matching error message, remove the just generated file, and exit
with error.
FIXME: It would be really nice to find a way to auomate the correct
alignment of the "context" member, but I don't see any way to use
alignment attributes to get what we need here.
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
while debugging a problem introduced in commit
63b503916d ("Cygwin: tls_pathbuf: Use Windows heap")
a hang in fork was encountered using the original implementation
of tls_pathbuf:
Using tmp_pathbuf inside the code block guarded by __malloc_trylock
may call malloc from tmp_pathbuf::w_get and thus trying to lock an
exclusive SRW lock recursively, which results in a deadlock.
Allocate a small SECURITY_ATTRIBUTES block on the stack rather than
allocating a 64K tmp_pathbuf. This avoids the potential malloc call.
Drop the __malloc_trylock call entirely. There must not be a malloc
call inside the frok::parent block guarded by __malloc_lock, and
just trying to lock is too dangerous inside fork while other threads
might actually chage the content of the heap. Additionally, add a
comment frowning on malloc usage inside tis code block.
Fixes: 44a79a6eca ("Cygwin: convert malloc lock to SRWLOCK")
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
Rather than using malloc/free for the buffers, we're now using
HeapAlloc/HeapFree on a HEAP_NO_SERIALIZE heap created for this
thread.
Advantages:
- Less contention. Our malloc/free doesn't scale well in
multithreaded scenarios
- Even faster heap allocation by using a non serialized heap.
- Internal, local, temporary data not cluttering the user heap.
- Internal, local, temporary data not copied over to child process
at fork().
Disadvantage:
- A forked process has to start allocating temporary buffers from
scratch. However, this should be alleviated by the fact that
buffer allocation usually reaches its peak very early in process
runtime, so the longer the proceess runs, the less buffers have
to allocated, and, only few processes don't exec after fork
anyway.
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
shmat may call shmget. shmget locks by itself as necessary,
so there's no reason to keep the lock active and recurse into
the lock. Use SRWLOCK and only lock as required.
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
So far we use a single muto to guard three different datastructures
inside class authz_ctx: the authz HANDLE, the user context HANDLE
and the context cache list. Split the single muto into three
independent SRWLOCKs and guard all datastrcutures as necessary to
avoid thread contention.
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
The old technique was from a time when we had to reduce stack pressure
by moving 64K buffers elsewhere. It was implemented using a static
global buffer, guarded by a muto. However, that adds a lock which may
unnecessarily serialize threads.
Use Windows heap buffers per invocation instead. HeapAlloc/HeapFree are
pretty fast, scale nicely in multithreaded scenarios and don't serialize
threads unnecessarily.
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
CloseHandle gets redefined to a macro calling an internal function
in debug.h when building with -DDEBUGGING, but profiler has no access
to that function.
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
This simple testcase:
locale_t st = newlocale(LC_ALL_MASK, "C", (locale_t)0);
locale_t st2 = newlocale(LC_CTYPE_MASK, "en_US.UTF-8", st);
is sufficient to reproduce a crash in _newlocale_r. After the first call
to newlocale, `st' points to __C_locale, which is const. When using `st'
as locale base in the second call, _newlocale_r tries to set pointers
inside base to NULL. This is bad if base is __C_locale, obviously.
Add a test to avoid trying to overwrite pointer values inside base if
base is __C_locale.
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
This way, the sem API is all in the same place, even if the
underlying semaphore class is still in thread.cc.
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
Rename CygwinCreateThread to create_posix_thread and move
from miscfuncs.cc to create_posix_thread.cc, inbcluding all
related functions. Analogue for the prototypes.
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
Currently it is possible for symlink_info::check to return -1 in case
we're searching for foo and find foo.lnk that is not a Cygwin symlink.
This contradicts the new meaning attached to a negative return value
in commit 19d59ce75d. Fix this by setting "res" to 0 at the beginning
of the main loop and not seting it to -1 later.
Also fix the commentary preceding the function definition to reflect
the current behavior.
Addresses: https://cygwin.com/pipermail/cygwin/2022-August/252030.html
provide entire internal and external pthread API from inside the
same file.
While I dislike to have another even larger file, this is basically
cleaning up the source and grouping the external API into useful
chunks. Splitting the file cleanly is tricky due to usage of inline
methods is_good_object and verifyable_object_isvalid.
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
So far, wmemset used the C implemantation from newlib. Let's use
the optimized assembler code instead.
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
This function is just bad. It really only works for ASCII
chars, everything else is broken after the conversion.
Introduce new helper function sys_mbstouni to replace
RtlCreateUnicodeStringFromAsciiz in hash_path_name.
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
mount_info::get_mounts_here used RtlCreateUnicodeStringFromAsciiz
which translates bytes into wide chars verbatim.
Create a new function sys_mbstouni_alloc which can be used from
mount_info::get_mounts_here to convert multibyte mount point
strings to UNICODE_STRINGS in a locale-aware way.
For symmetry, create a function mount_info::free_mounts_here,
so the knwoledge how to free the UNICODE_STRING buffers is
encapsulated in the same class.
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
This reverts commit d9e9c7b5a7. The latter added ".dll" to the
blessed_executable_suffixes array because on 32-bit Windows, the
GetBinaryType function would report that a 64-bit DLL is an
executable, contrary to the documentation of that function.
That anomaly does not exist on 64-bit Windows, so we can remove ".dll"
from the list. Reverting the commit does, however, change the
behavior of the rename(2) syscall in the following unlikely situation:
Suppose we have an executable foo.exe and we make the call
rename ("foo", "bar.dll");
Previously, foo.exe would be renamed to bar.dll. So bar.dll would
then be an executable without the .exe extension. The new behavior is
that foo.exe will be renamed to bar.dll.exe. [Exception: If there
already existed an executable (not a DLL!) with the name bar.dll, then
.exe will not be appended.]
The actual reason for these wrappers are lost in time, there's no
hint even in the pre-2000 ChangeLog files. Apparently they were
masking the prototypes or, alternatively, macros from newlib to
clash with the definitions in syscalls.cc.
They are not needed anymore, so just drop them.
This uncovered that the buffer pointer to pwrite is erronously
non-const. Fix this on the way out.
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
To avoid linear searches for error codes, autogenerate errmap as
simple array of errno values indexed by Windows error codes.
Restrict to the first 9000 Windows error codes, we don't care for
most of them anyway.
Define errmap in its own file errmap.h to clean up errno.cc.
Change geterrno_from_win_error accordingly.
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
Avoid linear searches for error codes by converting wsock_errmap
to a ordered array, indexed by WinSock error code (subtracted by
WSABASEERR.
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
This decouples host_errmap from the errmap_t definition which is
about to be changed in a followup patch.
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
Given that 64 bit Cygwin defines all file access types (off_t,
fpos_t, and derived types) as 64 bit anyway, there's no reason
left to rely on the stdio64 part of newlib. Use base functions
and base types.
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
- Previously, LoadLibraryA() is hooked for ConEmu cygwin connector.
With this patch, GetProcAddress() for "RequestTermConnector" is
hooked instead which is more essential for ConEmu cygwin connector.
- Recent commit "Cygwin: console: Add missing input_mutex guard."
has a problem that causes NULL pointer access if cygheap->ctty
is NULL. This patch fixes the issue.
gdb master recently learnt how to use GetThreadDescription() [1], so set
threadnames using SetThreadDescription() [available since Windows 10
1607] as well.
This is superior to using a special exception to indicate the thread
name to the debugger, because the thread name isn't missed if you don't
have a debugger attached at the time it's set.
It's not clear what the encoding of a thread name string is, we assume
UTF8 for the moment.
For the moment, continue to use the old method as well, for the benefit
of older gdb versions etc.
[1] https://sourceware.org/pipermail/gdb-patches/2022-April/187833.html
Back in early Cygwin development a function based access to the
environment was exported, the internal environ in Cygwin was called
__cygwin_environ and cur_environ() was used to access the environment
indirectly . The history of that necessity is not documented,
but kept in i686 for backward compatibility.
The x86_64 port eventually used __cygwin_environ directly and exported
it as DATA under the usual name environ.
We don't need the i686 workaround anymore, so just rename
__cygwin_environ to environ, drop the cur_environ() macro and
simply export environ under its own name.
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
- rename "is64bit" to "is_x86_64"
- Always set sym_prefix to empty string and add a FIXME
- speclib: rename uscore to sym_prefix and fix bug in string handling
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
- ConEmu cygwin connector conflicts with cons_master_thread since
it does not use read() to read console input. With this patch,
cons_master_thread is disabled in ConEmu cygwin connector.
Somehow this patch looks like it was pushed before having been
finished. Let's try again...
Fixes: e46f15c2d1 ("Cygwin: hookapi: drop handling i386 targets")
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
Only the assembler snippet is really x86_64-specific, so minimize the
conditional code block to this snippet.
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
Add a _REENT_CLEANUP() macro to encapsulate access to the
__cleanup member of struct reent. This will help to replace the
struct member with a thread-local storage object in a follow up
patch.
Add a _REENT_ERRNO() macro to encapsulate the access to the
_errno member of struct reent. This will help to replace the
structure member with a thread-local storage object in a follow
up patch.
Replace uses of __errno_r() with _REENT_ERRNO(). Keep __errno_r() macro for
potential users outside of Newlib.
Jon Turney
Corinna Vinschen
Ken Brown
Takashi Yano
Markus Mützel
Matt Joyce