This became a requirement, following a previous CMake update.
Otherwise the built EXEs would not export their intended functions.
Addendum to commit d8e92b5a (see PR #1335)
CORE-15406
[LOADCONFIG] The GCC rant had nothing to do with GCC, but everything with CMake...
[SPOOLSV] Add the missing exported stubs.
We get file matches when their attributes match and we can obtain
a valid 8.3 file name. This last condition is of upmost importance.
We cannot reliably use GetShortPathName() on the filename returned by
FindFirstFile()/FindNextFile() because the result is uncertain (one
either needs to build a full path before calling the function just to
get the final short path form, or change the current working directory
to the one where the search is being made, etc.)
The Find*file() functions return anyway the short file. name (in the
cAlternateFileName member) if the file does have such 8.3 filename and
its real name (in cFileName) is longer. Otherwise, cFileName could
already be *THE* short filename because it has the correct format for
it. Check this latter case with RtlIsNameLegalDOS8Dot3() and use it if
so. Otherwise this means the file has a long filename that cannot be
converted to short filename format (because e.g. it is in a volume where
short filenames are unavailable), and we skip such files: they wouldn't
be accessible from DOS anyways.
- Doxygen-document demFileFindFirst(), demFileFindNext() and associated
flags and structures. Update their annotations.
This fixes TurboC 2.x installation.
This fixes the age-long annoyance that wherever you ran a DOS program
with NTVDM, its cmos.ram file would be created in the same current
directory the DOS program was being started.
This created at least two problems:
- useless pollution of directories with cmos.ram files;
- for installers, e.g. of Turbo C 1.x that enumerate the install files
from their directories to be copied/extracted somewhere, the cmos.ram
file could then be enumerated along and cause the installation to
interrupt and/or fail.
Now the cmos.ram file is created in the same directory NTVDM resides
(usually in System32...).
Also, debug-print out only loading errors instead of successes as well.
The server-side CsrSrvIdentifyAlertableThread and CsrSrvSetPriorityClass
functions are completely removed in Win2k3+, and are since stubbed by
CsrSrvUnusedFunction instead. They however were present up to Windows XP,
albeit with an extremely minimal implementation.
The corresponding client-side CsrIdentifyAlertableThread and CsrSetPriorityClass
now become just stubs that either trivially succeed or fail, respectively.
See https://www.geoffchappell.com/studies/windows/win32/csrsrv/api/srvinit/apidispatch.htm
for more information.
- Fix typo "al*T*ertable" --> "alertable".
- Remove ROS-specific CSRSS_IDENTIFY_ALERTABLE_THREAD that was
deprecated since ages (at least before 2005)!
With introduction of 74e527b, a330b56 and subsequently 2791ecd ReactOS can now properly set IOPL (I/O privilege level) for user mode trusted processes. With that said, enable the compilation of raise hard error function call code back.
ProcessUserModeIOPL is strictly implemented for 32-bit architecture so it doesn't make any sense for CSRSS to gather user mode I/O privilege when that won't work anyway.
CORE-17545
subsystems/win32/csrsrv/api.c:63:9: warning: variable 'ServerDll' is used uninitialized whenever '||' condition is true [-Wsometimes-uninitialized]
if ((ServerId >= CSR_SERVER_DLL_MAX) ||
^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
subsystems/win32/csrsrv/api.c:67:76: note: uninitialized use occurs here
DPRINT1("CSRSS: %lx is invalid ServerDllIndex (%08x)\n", ServerId, ServerDll);
^~~~~~~~~
Addendum to commit 0a392b18.
The actual problem that existed all along was that the buffers being
validated with CsrValidateMessageBuffer() were not the correct ones!
What had to be checked is the string buffer **INSIDE** the UNICODE_STRING
structures! Indeed, it is these buffers that we are allocating on client side,
see https://github.com/reactos/reactos/blob/9b421af1/dll/win32/kernel32/client/dosdev.c#L324-L336
Dedicated to Pierre Schweitzer.
Since indeeed the memcpy/move type of functions usually don't operate on
pointers to 'volatile' data, don't make ClientCaptureBuffer volatile.
Instead when retrieving the values of its 'Size' and 'PointerCount'
members (see the SEH block), use the so-called technique of "volatile
glasses" [1], where the pointer will be explicitly casted to "volatile TYPE*"
where needed.
[1]: https://docs.microsoft.com/archive/blogs/itgoestoeleven/why-your-user-mode-pointer-captures-are-probably-broken
- Improve capture buffer validation in CsrCaptureArguments(), by
implementing the checks done by Windows 2003 (NT 5.2) described
in section "Server-Side Validation and Capture" of the article
https://www.geoffchappell.com/studies/windows/win32/csrsrv/api/apireqst/capture_header.htm
- In CsrReleaseCapturedArguments(), protect the data copy back into
the client buffer within a SEH block.
Rename inaccurate names 'LocalCaptureBuffer' and 'RemoteCaptureBuffer'
into 'ClientCaptureBuffer' and 'ServerCaptureBuffer' respectively.
(Recall: CSRSRV is the 'Server', and any app doing LPC calls to it is
the 'Client'.)
Also, trigger the less fatal breakpoints only if CSRSS/CSRSRV is being
debugged (the 'BeingDebugged' flag is set in the current PEB). This will
avoid any unhandled breakpoint exceptions when testing/fuzzing running
debug builds of ReactOS without any debugger attached.