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>
PR 29515 points out our documentation builds are broken, let's just move
over to the new non-recursive builds.
Signed-off-by: Palmer Dabbelt <palmer@rivosinc.com>
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 patch makes syscalls for SH architecture respecting the global option
"--disable-newlib-supplied-syscalls". This is useful when a bare-metal
toolchain is needed.
Signed-off-by: Yilin Sun <imi415@imi.moe>
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.]
Corinna Vinschen
Palmer Dabbelt
Jeff Johnston
Yilin Sun via Newlib
Ken Brown
Takashi Yano