CORE-17470
+ KdpDebugLogInit: Add resources cleanup in failure code paths.
Fix, in an NT-compatible manner, how (and when) the KD/KDBG BootPhase >=2
initialization steps are performed.
These are necessary for any functionality KDBG needs, that would depend
on the NT I/O Manager and the storage and filesystem stacks to be running.
This includes, creating the debug log file, and for KDBG, loading its
KDBinit initialization file.
As a result, file debug logging is fixed.
The old ReactOS-specific (NT-incompatible) callback we did in the middle
of IoInitSystem() is removed, in favor of a runtime mechanism that should
work on Windows as well.
The idea for this new mechanism is loosely inspired by the TDL4 rootkit,
see http://blog.w4kfu.com/public/tdl4_article/draft_tdl4article.html
but contrary to it, a specific hook is used instead, as well as the
technique of driver reinitialization:
https://web.archive.org/web/20211021050515/https://driverentry.com.br/en/blog/?p=261
Its rationale is as follows:
We want to be able to perform I/O-related initialization (starting a
logger thread for file log debugging, loading KDBinit file for KDBG,
etc.). A good place for this would be as early as possible, once the
I/O Manager has started the storage and the boot filesystem drivers.
Here is an overview of the initialization steps of the NT Kernel and
Executive:
----
KiSystemStartup(KeLoaderBlock)
if (Cpu == 0) KdInitSystem(0, KeLoaderBlock);
KiSwitchToBootStack() -> KiSystemStartupBootStack()
-> KiInitializeKernel() -> ExpInitializeExecutive(Cpu, KeLoaderBlock)
(NOTE: Any unexpected debugger break will call KdInitSystem(0, NULL); )
KdInitSystem(0, LoaderBlock) -> KdDebuggerInitialize0(LoaderBlock);
ExpInitializeExecutive(Cpu == 0): ExpInitializationPhase = 0;
HalInitSystem(0, KeLoaderBlock); <-- Sets HalInitPnpDriver callback.
...
PsInitSystem(LoaderBlock)
PsCreateSystemThread(Phase1Initialization)
Phase1Initialization(Discard): ExpInitializationPhase = 1;
HalInitSystem(1, KeLoaderBlock);
...
Early initialization of Ob, Ex, Ke.
KdInitSystem(1, KeLoaderBlock);
...
KdDebuggerInitialize1(LoaderBlock);
...
IoInitSystem(LoaderBlock);
...
----
As we can see, KdDebuggerInitialize1() is the last KD initialization
routine the kernel calls, and is called *before* the I/O Manager starts.
Thus, direct Nt/ZwCreateFile ... calls done there would fail. Also,
we want to do the I/O initialization as soon as possible. There does
not seem to be any exported way to be notified about the I/O manager
initialization steps... that is, unless we somehow become a driver and
insert ourselves in the flow!
Since we are not a regular driver, we need to invoke IoCreateDriver()
to create one. However, remember that we are currently running *before*
IoInitSystem(), the I/O subsystem is not initialized yet. Due to this,
calling IoCreateDriver(), much like any other IO functions, would lead
to a crash, because it calls
ObCreateObject(..., IoDriverObjectType, ...), and IoDriverObjectType
is non-initialized yet (it's NULL).
The chosen solution is to hook a "known" exported callback: namely, the
HalInitPnpDriver() callback (it initializes the "HAL Root Bus Driver").
It is set very early on by the HAL via the HalInitSystem(0, ...) call,
and is called early on by IoInitSystem() before any driver is loaded,
but after the I/O Manager has been minimally set up so that new drivers
can be created.
When the hook: KdpInitDriver() is called, we create our driver with
IoCreateDriver(), specifying its entrypoint KdpDriverEntry(), then
restore and call the original HalInitPnpDriver() callback.
Another possible unexplored alternative, could be to insert ourselves
in the KeLoaderBlock->LoadOrderListHead boot modules list, or in the
KeLoaderBlock->BootDriverListHead boot-driver list. (Note that while
we may be able to do this, because boot-drivers are resident in memory,
much like we are, we cannot insert ourselves in the system-driver list
however, since those drivers are expected to come from PE image files.)
Once the KdpDriverEntry() driver entrypoint is called, we register
KdpDriverReinit() for re-initialization with the I/O Manager, in order
to provide more initialization points. KdpDriverReinit() calls the KD
providers at BootPhase >= 2, and schedules further reinitializations
(at most 3 more) if any of the providers request so.
CORE-10749
The dmesg command is now available even if screen output is disabled.
Co-authored-by: Hermès Bélusca-Maïto <hermes.belusca-maito@reactos.org>
- KdbSymInit() in kdb_symbols.c only initializes symbols implementation
support.
- The rest of KdbInitialize gets moved into kdb_cli.c and initializes
the KDBG debugger itself.
- Move KdbDebugPrint to kdb_cli.c as well.
Access check is an expensive operation, that is, whenever an access to an
object is performed an access check has to be done to ensure the access
can be allowed to the calling thread who attempts to access such object.
Currently SepAnalyzeAcesFromDacl allocates a block of pool memory for
access check rights, nagging the Memory Manager like a desperate naughty
creep. So instead initialize the access rights as a simple variable in
SepAccessCheck and pass it out as an address to SepAnalyzeAcesFromDacl so
that the function will fill it up with access rights. This helps with
performance, avoiding wasting a few bits of memory just to hold these
access rights.
In addition to that, add a few asserts and fix the copyright header on
both se.h and accesschk.c, to reflect the Coding Style rules.
See this command's documentation:
https://docs.microsoft.com/en-us/windows-hardware/drivers/debugger/-dbgprint
and the section "DbgPrint buffer and the debugger"
https://docs.microsoft.com/en-us/windows-hardware/drivers/debugger/reading-and-filtering-debugging-messages#dbgprint-buffer-and-the-debugger
for more details.
- Loosely implement the function, based on our existing circular printout
buffers in kdio.c.
- Enable its usage in the KdpPrint() and KdpPrompt() functions.
Notice that this function will *only* capture the strings being sent **to**
the debugger, and not the strings the debugger itself produce. (This means
that we cannot use the KdPrintCircularBuffer as a replacement for our
KDBG dmesg one, for example...)
How to test:
Run ReactOS under WinDbg, and use the !dbgprint command to view the
buffer. You can also use the Memory Window, place yourself at the
address pointed by KdPrintCircularBuffer and KdPrintWritePointer, and
read its contents.
What you should observe:
Prior notice: The circular buffer in debug builds of ReactOS and Windows
is 0x8000 bytes large. In release builds, its size is down to 0x1000.
1- When you start e.g. the 2nd-stage GUI installation of ReactOS, going
past the initial "devices installation" and letting it stabilize on
the Welcome page, break into WinDbg and run the !dbgprint command. You
should notice that the end of its output is weirdly truncated, compared
to what has been actually emitted to the debug output. Comparing this
with the actual contents of the circular buffer (via Memory Window),
shows that the buffer contents is actually correct.
2- Copy all the text that has been output by the !dbgprint command and
paste it in an editor; count the number of all characters appearing +
newlines (only CR or LF), and observe that this number is "mysteriously"
equal to 16384 == 0x4000.
3- Continue running ReactOS installation for a little while, breaking back
back into WinDbg and looking at !dbgprint again. Its output seems to be
still stopping at the same place as before (but the actual buffer memory
contents shows otherwise). Continue running ROS installation, and break
into the debugger when ROS is about to restart. You should now observe
that the dbgprint buffer rolled over:
dd nt!KdPrintRolloverCount shows 1.
Carefully analysing the output of !dbgprint, however, you will notice
that it looks a bit garbage-y: the first part of the output is actually
truncated after 16384 characters, then you get a second part of the
buffer showing what ReactOS was printing while shutting down. Then
you get again what was shown at the top of the !dbgprint output.
(Of course, comparing with the actual contents of the circular buffer
in memory shows that its contents are fine...)
The reason of these strange observations, is because there is an intrinsic
bug in the !dbgprint command implementation (in kdexts.dll). Essentially,
it displays the contents of the circular buffer in two single dprintf()
calls: one for the "older" (bottom) part of the buffer:
[WritePointer, EndOfBuffer]
and one for the "newer" (upper) part of the buffer:
[CircularBuffer, WritePointer[ .
The first aspect of the bug (causing observation 3), is that those two
parts are not necessarily NULL-terminated strings (especially after
rollover), so for example, displaying the upper part of the buffer, will
potentially also display part of the buffer's bottom part.
The second aspect of the bug (explaining observations 1 and 2), is due
to the implementation of the dprintf() function (callback in dbgenv.dll).
There, it uses a fixed-sized buffer of size 0x4000 == 16384 characters.
Since the output of the circular buffer is not done by little chunks,
but by the two large parts, if any of those are larger than 0x4000 they
get truncated on display.
(This last observation is confirmed in a completely different context by
https://community.osr.com/discussion/112439/dprintf-s-max-string-length .)
- Remove KdbInit() macro and directly use KdbpCliInit() (since the place
where it was used was already within an #ifdef KDBG block).
- Declare KdpKdbgInit() only when KDBG is defined, move its definition
into kdio.c and remove the legacy wrappers/kdbg.c file.
And in KdbInitialize(), set KdpInitRoutine directly to the former,
instead of using the KdpKdbgInit indirection.
- Don't reset KdComPortInUse in KdpDebugLogInit().
- Minor refactorings: KdpSerialDebugPrint -> KdpSerialPrint and make it
static; argument name "Message" -> "String", "StringLength" -> "Length".
debug.c will serve as a centralized facility for security debugging routines and everything related to that. This file will be expanded with further debug functions for the Security subsystem if needed.
On current master, ReactOS faces these problems:
- ObCreateObject charges both paged and non paged pool a size of TOKEN structure, not the actual dynamic contents of WHAT IS inside a token. For paged pool charge the size is that of the dynamic area (primary group + default DACL if any). This is basically what DynamicCharged is for.
For the non paged pool charge, the actual charge is that of TOKEN structure upon creation. On duplication and filtering however, the paged pool charge size is that of the inherited dynamic charged space from an existing token whereas the non paged pool size is that of the calculated token body
length for the new duplicated/filtered token. On current master, we're literally cheating the kernel by charging the wrong amount of quota not taking into account the dynamic contents which they come from UM.
- Both DynamicCharged and DynamicAvailable are not fully handled (DynamicAvailable is pretty much poorly handled with some cases still to be taking into account). DynamicCharged is barely handled, like at all.
- As a result of these two points above, NtSetInformationToken doesn't check when the caller wants to set up a new default token DACL or primary group if the newly DACL or the said group exceeds the dynamic charged boundary. So what happens is that I'm going to act like a smug bastard fat politician and whack
the primary group and DACL of an token however I want to, because why in the hell not? In reality no, the kernel has to punish whoever attempts to do that, although we currently don't.
- The dynamic area (aka DynamicPart) only picks up the default DACL but not the primary group as well. Generally the dynamic part is composed of primary group and default DACL, if provided.
In addition to that, we aren't returning the dynamic charged and available area in token statistics. SepComputeAvailableDynamicSpace helper is here to accommodate that. Apparently Windows is calculating the dynamic available area rather than just querying the DynamicAvailable field directly from the token.
My theory regarding this is like the following: on Windows both TokenDefaultDacl and TokenPrimaryGroup classes are barely used by the system components during startup (LSASS provides both a DACL and primary group when calling NtCreateToken anyway). In fact DynamicAvailable is 0 during token creation, duplication and filtering when inspecting a token with WinDBG. So
if an application wants to query token statistics that application will face a dynamic available space of 0.
NtQueryInformationToken is by far the only system call in NT where ReturnLength simply cannot be optional. On Windows this parameter is always probed and an argument to NULL directly leads to an access violation exception.
This is due to the fact of how tokens work, as its information contents (token user, owner, primary group, et al) are dynamic and can vary throughout over time in memory.
What happens on current ReactOS master however is that ReturnLength is only probed if the parameter is not NULL. On a NULL case scenario the probing checks succeed and NtQueryInformationToken fails later. For this, just get rid of CompleteProbing
parameter and opt in for a bit mask flag based approach, with ICIF_FORCE_RETURN_LENGTH_PROBE being set on DefaultQueryInfoBufferCheck which NtQueryInformationToken calls it to do sanity checks.
In addition to that...
- Document the ICIF probe helpers
- Annotate the ICIF prope helpers with SAL
- With the riddance of CompleteProbing and adoption of flags based approach, add ICIF_PROBE_READ_WRITE and ICIF_PROBE_READ flags alongside with ICIF_FORCE_RETURN_LENGTH_PROBE
The current state of Security manager's code is kind of a mess. Mainly, there's code scattered around places where they shouldn't belong and token implementation (token.c) is already of a bloat in itself as it is. The file has over 6k lines and it's subject to grow exponentially with improvements, features, whatever that is.
With that being said, the token implementation code in the kernel will be split accordingly and rest of the code moved to appropriate places. The new layout will look as follows (excluding the already existing files):
- client.c (Client security implementation code)
- objtype.c (Object type list implementation code -- more code related to object types will be put here when I'm going to implement object type access checks in the future)
- subject.c (Subject security context support)
The token implementation in the kernel will be split in 4 distinct files as shown:
- token.c (Base token support routines)
- tokenlif.c (Life management of a token object -- that is Duplication, Creation and Filtering)
- tokencls.c (Token Query/Set Information Classes support)
- tokenadj.c (Token privileges/groups adjusting support)
In addition to that, tidy up the internal header and reorganize it as well.
This is needed to store FPU state information when saving or restoring the floating point state of a system due to a call to KeSaveFloatingPointState or KeRestoreFloatingPointState.
Implement the correct start-stop sequence for resource rebalancing
without the actual rebalancing. Also move IoInvalidateDeviceState
processing into the enumeration thread as it should be.
CORE-17519
Implement initial token debug code. For now debug information that is being tracked are: process image file name, process and thread client IDs and token creation method. More specific debug code can be added later only if needed.
As for the token creation method, this follows the same principle as on Windows where the creation method is defined by a value denoting the first letter of the said method of creation. That is, 0xC is for token creation, 0xD is for token duplication and 0xF is for token filtering. The debug field names are taken from Windows PDB symbols for WinDBG debug extension support purposes. The names must not be changed!
- Add a new cmboot.h header to isolate the boot-support definitions
shared with the NT/ReactOS bootloader.
- Move CmpFreeDriverList() to cmboot.c so that we can use it for
cleanup paths in the NT/ReactOS bootloader.
- CmpFindControlSet(): Directly build the control set name in UNICODE,
instead of doing an ANSI->UNICODE conversion.
- Directly assign the CurrentControlSet\Services constant string,
instead of going the route of init-empty-string + append-string.
This is possible since that string is not modified later.
- Remove ASSERT(FALSE), replacing them with correct failure handling.
- Add cleanup paths in CmpAddDriverToList().
- Simplify and fix CmpFreeDriverList(): it's the full DriverNode
that needs to be freed; not the LIST_ENTRY pointer.
- Add other validity checks:
* Registry value types and data sizes;
* For multi-strings, verify that they are NULL-terminated.
* For (multi-)strings, check whether they are NULL-terminated before
optionally removing their trailing NULL character from the count.
Check also whether they are of zero-length and take appropriate
action where necessary.
- Add CmpIsDriverInList() for future usage in CMBOOT compiled in
bootloader mode.
- Add SAL annotations and Doxygen documentation.
- Add debug traces.
- Formatting / code style fixes.
** TODO: Fix SafeBoot support **
- inbv.c now only contains the Inbv-specific API and nothing else.
- It will make easier for people to write their own boot themes & animations,
by just copying/adapting the bootanim.c file (and the resources).
- Add SAL annotations.
- All INBV progress bar functions (except for InbvIndicateProgress())
should not be INIT-only functions, since they can be (not yet in ROS)
used at later times -- namely, for feedback during hibernation.
In particular, the progress percentage specified to InbvUpdateProgressBar(),
or the progress feedback made with InbvIndicateProgress() calls, is
**relative** to the progress sub-range specified with a previous call to
InbvSetProgressBarSubset() (by default, the range is 0...100%).
This functionality is used e.g. when the number of progress steps is
unknown prior, for example when loading drivers: in this case progress
is made within a given percentage range.
This bug has always been with us since 2010.
-- Rewrite PspChargeProcessQuotaSpecifiedPool and PspReturnProcessQuotaSpecifiedPool private kernel routines, with the goal to implement the algorithms necessary to manage the fields of quota blocks (Usage, Return, Limit and Peak).
-- Invoke the Mm if quota limit raising or returning has to be done
-- When destroying a quota block, make sure that we're giving back all the rest of non-returned quotas to Memory Mm
-- Crash the system with QUOTA_UNDERFLOW if someone is returning way too much quota than it was previously charged
-- When a process exits, ensure that it doesn't hold up any charged quotas in QuotaUsage field of the process object, that way we're enforcing proper kernel consistency
-- Implement PsChargeSharedPoolQuota and PsChargeProcessPageFileQuota functions, used exclusively by the Object Manager. These routines are used to charge or return amount of quotas of a newly created object.
-- On PspInheritQuota, when assigning to process the default quota block if no parent process is given, we must increment the reference counts as we're using it
-- Handle the ProcessCount reference field, as it wasn't used
-- Annotate the functions with SAL
-- Document the code
=== REMARKS ===
Windows LogOn (Winlogon) is responsible for setting up a different quota block for all the processes within an interactive session, which is what we don't do. What we're currently doing instead is we're using the default block, PspDefaultQuotaBlock, for all the processes
across the system. The default block contains the default limits of -1 (which would imply no limits). By definition, the kernel won't ever return STATUS_QUOTA_EXCEEDED as we literally don't set up a definite limit for regular processes. This situation has to be tackled
in the future.
=== TODO FOR FUTURE ===
Most of the code in PspChargeProcessQuotaSpecifiedPool and PspReturnProcessQuotaSpecifiedPool private routines must be refactored in order to reduce the usage of the quota spin lock, possibly wrapping such code in a loop and whatnot.
CORE-17784
OBP_SYSTEM_PROCESS_QUOTA is a macro that'll be used as a way to assign a dummy quota block to system processes, as we mustn't do anything to those in case the Object Manager is charging or returning pool quotas.
Declare PsReturnSharedPoolQuota and PsChargeSharedPoolQuota prototypes and annotate the functions. Furthermore, add two definitions related to quota pool limits threshold -- PSP_NON_PAGED_POOL_QUOTA_THRESHOLD and PSP_PAGED_POOL_QUOTA_THRESHOLD. For further details, see the commit description of "[NTOS:MM] Add the pool quota prototypes and some definitions".
Declare the MmRaisePoolQuota and MmReturnPoolQuota prototypes in the header and add some definitions related to pool quotas, namely MmTotalNonPagedPoolQuota and MmTotalPagedPoolQuota. These variables are used internally by the kernel as sort of "containers" (for the lack of a better term)
which uphold the amount of quotas that the Process Manager is requesting the Memory Manager to raise or return the pool quota limit. In addition to that, add some definitions needed for both of these functions.
The definitions, MI_CHARGE_PAGED_POOL_QUOTA and MI_CHARGE_NON_PAGED_POOL_QUOTA respectively, bear some interesting aspect. Seemingly the 0x80000 and 0x10000 values (that would denote to 524288 and 65536 specifically) are used as quota "limits" or in other words, thresholds that the kernel
uses. So for example if one would want to raise the quota limit charge, MmRaisePoolQuota will raise it so based on this formula -- NewMaxQuota = CurrentQuota + LIMIT_VALUE. LIMIT_VALUE can be either MI_CHARGE_PAGED_POOL_QUOTA or MI_CHARGE_NON_PAGED_POOL_QUOTA, depending a per quota pool basis.
What's more interesting is that these values are pervasive in Process Manager even. This is when quotas are to be returned back and trim the limit of the quota block if needed, the kernel would either take the amount provided by the caller of quotas to return or the threshold (paged or not paged)
if the amount to return exceeds the said threshold in question.
We have a special file, tag.h, which serves as a place to store whatever kernel pool allocation tag yet we still have some tags sparse over the kernel code... So just re-group them in one unique place.
SIDs are variadic by nature which means their lengths can vary in a given amount of time and certain factors that allow for this happen. This also especially can lead to issues when capturing SIDs and attributes because SeCaptureSidAndAttributesArray might end up overwriting the buffer during the time it's been called.
Therefore when we're copying the SIDs, validate their lengths. In addition to that, update the documentation header accordingly and add some debug prints in code.
- Disentangle the usage of ObpAcquireDirectoryLockExclusive() when it's
used only for accessing a directory structure, or as part of a lookup
operation.
The Obp*DirectoryLock*() -- both shared and exclusive -- functions
are only for locking an OB directory, for reading or writing its
structure members.
When performing lookup operations (insertions/deletions of entries
within a directory), use a ObpAcquireLookupContextLock() function that
exclusively locks the directory and saves extra lock state, that can
be used by ObpReleaseLookupContextObject() for cleanup.
- Add documentation for these functions.
