Vulnerabilities > CVE-2016-7644 - Use After Free vulnerability in Apple Iphone OS
Attack vector
LOCAL Attack complexity
LOW Privileges required
NONE Confidentiality impact
HIGH Integrity impact
HIGH Availability impact
HIGH Summary
An issue was discovered in certain Apple products. iOS before 10.2 is affected. macOS before 10.12.2 is affected. watchOS before 3.1.3 is affected. The issue involves the "Kernel" component. It allows attackers to execute arbitrary code in a privileged context or cause a denial of service (use-after-free) via a crafted app.
Vulnerable Configurations
Common Weakness Enumeration (CWE)
Exploit-Db
| description | iOS 10.1.1 / macOS 10.12 16A323 XNU Kernel - set_dp_control_port Lack of Locking Use-After-Free. CVE-2016-7637,CVE-2016-7644,CVE-2016-7661. Local exploit for... |
| file | exploits/multiple/local/40931.txt |
| id | EDB-ID:40931 |
| last seen | 2016-12-16 |
| modified | 2016-12-16 |
| platform | multiple |
| port | |
| published | 2016-12-16 |
| reporter | Exploit-DB |
| source | https://www.exploit-db.com/download/40931/ |
| title | iOS 10.1.1 / macOS 10.12 16A323 XNU Kernel - set_dp_control_port Lack of Locking Use-After-Free |
| type | local |
Nessus
| NASL family | MacOS X Local Security Checks |
| NASL id | MACOS_10_12_2.NASL |
| description | The remote host is running a version of macOS that is 10.12.x prior to 10.12.2. It is, therefore, affected by multiple vulnerabilities in the following components : - apache_mod_php - AppleGraphicsPowerManagement - Assets - Audio - Bluetooth - CoreCapture - CoreFoundation - CoreGraphics - CoreMedia External Displays - CoreMedia Playback - CoreStorage - CoreText - curl - Directory Services - Disk Images - FontParser - Foundation - Grapher - ICU - ImageIO - Intel Graphics Driver - IOFireWireFamily - IOAcceleratorFamily - IOHIDFamily - IOKit - IOSurface - Kernel - kext tools - libarchive - LibreSSL - OpenLDAP - OpenPAM - OpenSSL - Power Management - Security - syslog - WiFi - xar Note that successful exploitation of the most serious issues can result in arbitrary code execution. Furthermore, CVE-2016-6304, CVE-2016-7596, and CVE-2016-7604 also affect Mac OS X versions 10.10.5 and 10.11.6. However, this plugin does not check those versions. |
| last seen | 2020-06-01 |
| modified | 2020-06-02 |
| plugin id | 95917 |
| published | 2016-12-16 |
| reporter | This script is Copyright (C) 2016-2020 and is owned by Tenable, Inc. or an Affiliate thereof. |
| source | https://www.tenable.com/plugins/nessus/95917 |
| title | macOS 10.12.x < 10.12.2 Multiple Vulnerabilities |
| code | |
Seebug
| bulletinFamily | exploit |
| description | set_dp_control_port is a MIG method on the host_priv_port so this bug is a root->kernel escalation. ``` kern_return_t set_dp_control_port( host_priv_t host_priv, ipc_port_t control_port) { if (host_priv == HOST_PRIV_NULL) return (KERN_INVALID_HOST); if (IP_VALID(dynamic_pager_control_port)) ipc_port_release_send(dynamic_pager_control_port); dynamic_pager_control_port = control_port; return KERN_SUCCESS; } ``` This should be an atomic operation; there's no locking so two threads can race to see the same value for dynamic_pager_control_port and release two references when the kernel only holds one. This PoC triggers the bug such that the first thread frees the port and the second uses it; a more sensible approach towards exploiting it would be to use this race to try to decrement the reference count of a port with two references to zero such that you end up with a dangling port pointer. Tested on MacOS 10.12 16A323 **附件:[dpcp.c ](https://bugs.chromium.org/p/project-zero/issues/attachment?aid=254836)** * ianbeer */ READ THIS FIRST: if you do not have an iPod touch 6g running 10.1.1 (14b100) or and iPad mini 2 running 10.1.1 (14b100) this project will not work out of the box(*)! You need to fix up a couple of offsets - see the section futher down "Adding support for more devices" ``` (*) more precisely, I only have those devices and have only tested it on them. (*) 1b4150 will probably also work, I haven't tested it. ``` Contents: 1 - Build Instructions 2 - Adding support for other devices 3 - Notes on the bugs and exploits *** (1) Build Instructions *** * download and install Xcode 8.1 or higher * download Jonathan Levin’s collection of arm64 iOS binaries: + Follow the link for "The 64-bit tgz pack" here: http://newosxbook.com/tools/iOSBinaries.html (you want iosbinpack64.tgz) + extract it into the iosbinpack64 directory which is already in the mach_portal source dir so that directly underneath iosbinpack64 you have the bin/, etc/, sbin/, usr/ directories When you expand the iosbinpack64 directory in the xcode folder view you should see those folders * open this .xcodeproj * if you don't have an apple id make one now at https://appleid.apple.com * if you don't have a developer signing certificate you can make a free one now in Xcode * in Xcode go Xcode->Preference->Accounts and click the '+' in the lower left hand corner and add your apple id * select your account then "View Details" and under signing identites click Create next to iOS Development * connect your iDevice and click "trust" in the pop up on it * wait for xcode to process symbol files for this device * in the box to the right of the play and stop buttons in the top left corner of the xcode window select your iDevice * in the left hand window pane select the mach_portal project and navigate to the General tab * in the signing window select your personal team * We now need to fix up a few things: * go to Build Settings -> Packaging and give your project a new, unique bundle identifier (eg change it from "com.example.mach_portal" to "com.ios.test.account.mach_portal" where ios.test.account is your apple id. (it doesn’t have to be your apple id, just a unique string)) * We also need to register a unique App Group: * In the capabilities view scroll down to the App Groups section, remove the existing App Group ("group.mach_portal") and add a new unique one (eg "group.ios.test.account.mach_portal") * open jailbreak.c and change the app_group variable to this new app group id. * on the iDevice go to settings -> General -> Device Management and select your apple ID and click trust * in xcode click view -> debug area -> activate console so you can see debugging output (there's no output on the iDevice screen at all, that's normal) * make sure your iDevice and host are connected to the same wifi network and that network allows client to client connections. Note down the iDevice's ip address. * click play to run the app on the iDevice. If it fails press and hold the power and home buttons to reset the device. If Xcode asks you to enable developer mode on this mac agree. * if it succeeds you should see: "shell listening on port 4141" printed to the debug consol * the kernel exploit is only around 50% reliable (this can certainly be improved, read the code and make it better!) it will fail more often if there is high system load - try leaving the device for a minute after rebooting it and connecting it to you mac before trying again * connect to that port with netcat: nc X.X.X.X 4141 where X.X.X.X is your iDevice’s ip address * you have a root shell :) There’s no controlling terminal so fancy curses gui stuff won't work unless you fix that * you can run any pseudo-signed thin ARM64 binaries - if you want the kernel task port it's host special port 4 * copy your custom testing tools to the iosbinpack64 directory and they'll be bundled with the .app so you can run them from the shell * you're running as an unsandboxed root user so you can talk to any iokit user clients/mach services * amfid is patched to allow any signatures/entitlements * When you’re done hold power and home to reset the device *** (2) Adding support for other devices *** * you have to do this manually, sorry! * download the ipsw for your device from https://www.theiphonewiki.com/wiki/Firmware The bugs are there in any version <= 10.1.1 but the further back you go the more offsets will be wrong so ideally stick to 10.1.1 (and for anything earlier that iOS 10 the kernel cache is encrypted so you'll have to do the rest yourself) * for >= iOS 10 unzip the ipsw and hexdump the kernel.release.* file like this: ``` $ hexdump -C kernelcache.release.n51 | head 00000000 30 83 b5 9b 0d 16 04 49 4d 34 50 16 04 6b 72 6e |0......IM4P..krn| 00000010 6c 16 1c 4b 65 72 6e 65 6c 43 61 63 68 65 42 75 |l..KernelCacheBu| 00000020 69 6c 64 65 72 2d 31 31 36 32 2e 32 30 2e 31 04 |ilder-1162.20.1.| 00000030 83 b5 9a de 63 6f 6d 70 6c 7a 73 73 83 13 7d ae |....complzss..}.| 00000040 01 64 80 00 00 b5 29 5e 00 00 00 01 00 00 00 00 |.d....)^........| 00000050 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 |................| * 000001b0 00 00 00 00 ff cf fa ed fe 0c 00 00 01 d5 00 f6 |................| 000001c0 f0 02 f6 f0 14 f6 f0 38 0e 9a f3 f1 20 f6 f1 00 |.......8.... ...| 000001d0 19 ff f1 f5 f0 5f 9f 5f 54 45 58 54 09 02 1c 03 |....._._TEXT....| ``` * note down the offset of the ff cf fa ed fe byte sequence (in this case it's 0x1b4) * compile lzssdec from http://nah6.com/~itsme/cvs-xdadevtools/iphone/tools/lzssdec.cpp * run a command like: lzssdec -o 0x1b4 < kernel.release.n51 > kernel.decompressed * open the decompressed kernelcache in a recent version of IDA Pro (with support for iOS kextcaches) * say yes when IDA asks to split by kext * let the auto-analysis run - depending on how fast your computer is this might take a while! (it takes my 2013 MBP about 30 minutes) * go view -> open subviews -> segments and find the __TEXT:HEADER segment, the start should be FFFFFFF007004000 if it isn't note this down as you'll need to work out a couple of offsets relative to this * go view -> open subviews -> names and find the kernproc data symbol. * subtract the __TEXT:HEADER value from that, this is the kernproc offset eg for iPhone 5S 10.1.1 kernproc is at FFFFFFF0075AE0E0 making the offset: 0x5AA0E0 * now the harder one! We need to find allproc which isn't exported so is harder to find: * go view -> open subviews -> strings and find the string "pgrp_add : pgrp is dead adding process" * hit 'x' on the autogenerated string symbol name; you should see this symbol referenced from two functions * open the smaller of those functions in the IDA graph view * this is pgrp_add in the XNU source * scroll to the bottom of the CFG, the final three nodes all reference the same global variable with code like this: ADRP X8, #qword_FFFFFFF0075A8128@PAGE LDR X9, [X8,#qword_FFFFFFF0075A8128@PAGEOFF] * that's the address of allproc - subtract the kernel base to get the offset, in this case it's: 0x5A4128 * open offset.c and add support for your device. You should only have to update those two variable (kernproc and allproc) The structure offsets should stay the same, at least for recent kernels. If you want to target a much older kernel you'll also have to work out all the structure offsets - this is much more fiddly. * 32-bit devices: All the offsets will be totally different and the code which manipulates the kernel data structures will also be completely wrong. There's no reason it wouldn't work but you'll have to fix the code to make it work *** fixing userspace stuff *** I also rely on a handful of offsets in amfid; you should be able to find those very easily if they're different on your target. See the code and alse the section "Patch amfid" below. *** (3) Notes on the bugs and exploits *** This project is called "mach_portal" - it's the result of a research project I did this year looking at mach ports. (All the bugs used involve mach ports :-) ) There are two main bugs plus one more which is only used to force a service to restart: CVE-2016-7637: Broken kernel mach port name uref handling on iOS/MacOS can lead to privileged port name replacement in other processes CVE-2016-7644: XNU kernel UaF due to lack of locking in set_dp_control_port CVE-2016-7661: MacOS/iOS arbitrary port replacement in powerd There is no untether (persistent codesigning bypass) but the exploit will temporarily disable codesigning while it runs so you can run unsigned binaries. The high level exploit flow is like this: I use CVE-2016-7637 to replace launchd's send right to com.apple.iohideventsystem with a send right to a port for which I hold the receive right. I use CVE-2016-7661 to crash the powerd daemon (which runs as root). It gets automatically restarted and as part of its startup it will lookup the com.apple.iohideventsystem mach service and send its own task port to that service. Since I hold the receive right for that port this means that powerd actually sends me its task port giving me complete control over it :-) I use powerd's task port to get the host_priv port which I use to trigger the kernel bug. The kernel bug is a lack of locking when releasing a reference on a port. I allocate a large number of mach ports then trigger the bug on around 20 of them which are likely to be allocated near each other in the kernel. I use no-more-senders notifications so I can deterministically know when I've managed to over-release a port so that I can actually give myself dangling port pointers at an exact point in time later. I free all these mach ports (leaving myself with ~20 dangling mach port pointers) and force a zone GC. I try to move the page pointed to by all the dangling port pointers into the kalloc.4096 zone and then I send myself a large number of mach message containing OOL ports with send rights to the host port. I set up these OOL port pages so that overlapping the dangling port's context pointers there's a pointer to the host port ipc_port and the dangling port's lock and is_guarded fields are replaced with NULL pointers. If that all worked I can call mach_port_get_context on each of the dangling ports and I should get back the address of the host port ipc_port. The kernel task port is allocated at around the same time as the host port and as such they both end up in the same kernel zone page. I work out the base of this page then call mach_port_set_context on all of the dangling ports passing each possible address of the kernel task port in turn. I then receive all the ports I sent to myself and if everything worked I've ended receiving a send right to the kernel task port :) Here's a more detailed writeup of the sandbox escape part of the exploit. You'll have to read the code for the kernel exploit, I haven't written a longer writeup for that yet. *** Sandbox escape *** When sending and receiving mach messages from userspace there are two important kernel objects; ipc_entry and ipc_object. ipc_entry's are the per-process handles or names which a process uses to refer to a particular ipc_object. ipc_object is the actual message queue (or kernel object) which the port refers to. ipc_entrys have a pointer to the ipc_object they are a handle for along with the ie_bits field which contains the urefs and capacility bits for this name/handle (whether this is a send right, receive right etc.) ``` struct ipc_entry { struct ipc_object *ie_object; ipc_entry_bits_t ie_bits; mach_port_index_t ie_index; union { mach_port_index_t next; /* next in freelist, or... */ ipc_table_index_t request; /* dead name request notify */ } index; }; #define IE_BITS_UREFS_MASK 0x0000ffff /* 16 bits of user-reference */ #define IE_BITS_UREFS(bits) ((bits) & IE_BITS_UREFS_MASK) ``` The low 16 bits of the ie_bits field are the user-reference (uref) count for this name. Each time a new right is received by a process, if it already had a name for that right the kernel will increment the urefs count. Userspace can also arbitrarily control this reference count via mach_port_mod_refs and mach_port_deallocate. When the reference count hits 0 the entry is free'd and the name can be re-used to name another right (this is actually only the case for send rights). ipc_right_copyout is called when a right will be copied into a space (for example by sending a port right in a mach message to another process.) Here's the code to handle the sending of a send right: ``` case MACH_MSG_TYPE_PORT_SEND: assert(port->ip_srights > 0); if (bits & MACH_PORT_TYPE_SEND) { mach_port_urefs_t urefs = IE_BITS_UREFS(bits); assert(port->ip_srights > 1); assert(urefs > 0); assert(urefs < MACH_PORT_UREFS_MAX); if (urefs+1 == MACH_PORT_UREFS_MAX) { if (overflow) { /* leave urefs pegged to maximum */ <---- (1) port->ip_srights--; ip_unlock(port); ip_release(port); return KERN_SUCCESS; } ip_unlock(port); return KERN_UREFS_OVERFLOW; } port->ip_srights--; ip_unlock(port); ip_release(port); ... entry->ie_bits = (bits | MACH_PORT_TYPE_SEND) + 1; <---- (2) ipc_entry_modified(space, name, entry); break; ``` If copying this right into this space would cause that right's name's urefs count in that space to hit 0xffff then (if overflow is true) we reach the code at (1) which claims in the comment that it will leave urefs pegged at maximum. This branch doesn't increase the urefs but still returns KERN_SUCCESS. Almost all callers pass overflow=true. The reason for this "pegging" was probably not to prevent the reference count from becoming incorrect but rather because at (2) if the urefs count wasn't capped the reference count would overflow the 16-bit bitfield into the capability bits. The issue is that the urefs count isn't "pegged" at all. I would expect "pegged" to mean that the urefs count will now stay at 0xfffe and cannot be decremented - leaking the name and associated ipc_object but avoiding the possibilty of a name being over-released. In fact all that the "peg" does is prevent the urefs count from exceeding 0xfffe; it doesn't prevent userspace from believing it has more urefs than that (by eg making the copyout's fail.) What does this actually mean? Let's consider the behaviour of mach_msg_server or dispatch_mig_server. They receive mach service messages in a loop and if the message they receieved didn't corrispond to the MIG schema they pass that received message to mach_msg_destroy. Here's the code where mach_msg_destroy destroys an ool_ports_descriptor_t: ``` case MACH_MSG_OOL_PORTS_DESCRIPTOR : { mach_port_t *ports; mach_msg_ool_ports_descriptor_t *dsc; mach_msg_type_number_t j; /* * Destroy port rights carried in the message */ dsc = &saddr->ool_ports; ports = (mach_port_t *) dsc->address; for (j = 0; j < dsc->count; j++, ports++) { mach_msg_destroy_port(*ports, dsc->disposition); // calls mach_port_deallocate } ... ``` This will call mach_port_deallocate for each ool_port name received. If we send such a service a mach message with eg 0x20000 copies of the same port right as ool ports the ipc_entry for that name will actually only have 0xfffe urefs. After 0xfffe calls to mach_port_deallocate the urefs will hit 0 and the kernel will free the ipc_entry and mark that name as free. From this point on the name can be re-used to name another right (for example by sending another message received on another thread) but the first thread will still call mach_port_deallocate 0x10002 times on that name. This leads to something like a use-after-deallocate of the mach port name - strictly a userspace bug (there's no kernel memory corruption etc here) but caused by a kernel bug. The challenge to exploiting this bug is getting the exact same port name reused in an interesting way. This requires us to dig in a bit to exacly what a port name is, how they're allocated and under what circumstances they'll be reused. Mach ports are stored in a flat array of ipc_entrys: ``` struct ipc_entry { struct ipc_object *ie_object; ipc_entry_bits_t ie_bits; mach_port_index_t ie_index; union { mach_port_index_t next; /* next in freelist, or... */ ipc_table_index_t request; /* dead name request notify */ } index; }; ``` mach port names are made up of two fields, the upper 24 bits are an index into the ipc_entrys table and the lower 8 bits are a generation number. Each time an entry in the ipc_entrys table is reused the generation number is incremented. There are 64 generations, so after an entry has been reallocated 64 times it will have the same generation number. The generation number is checked in ipc_entry_lookup: ``` if (index < space->is_table_size) { entry = &space->is_table[index]; if (IE_BITS_GEN(entry->ie_bits) != MACH_PORT_GEN(name) || IE_BITS_TYPE(entry->ie_bits) == MACH_PORT_TYPE_NONE) entry = IE_NULL; } ``` here entry is the ipc_entry struct in the kernel and name is the user-supplied mach port name. Entry allocation: The ipc_entry table maintains a simple LIFO free list for entries; if this list is free the table will be grown. The table is never shrunk. Reliably looping mach port names: To exploit this bug we need a primitive that allows us to loop a mach port's generation number around. After triggering the urefs bug to free the target mach port name in the target process we immediately send a message with N ool ports (with send rights) and no reply port. Since the target port was the most recently freed it will be at the head of the freelist and will be reused to name the first of the ool ports contained in the message (but with an incremented generation number.) Since this message is not expected by the service (in this case we send an invalid XPC request to launchd) it will get passed to mach_msg_destroy which will pass each of the ports to mach_port_deallocate freeing them in the order in which they appear in the message. Since the freed port was reused to name the first ool port it will be the first to be freed. This will push the name N entries down the freelist. We then send another 62 of these looper messages but with 2N ool ports. This has the effect of looping the generation number of the target port around while leaving it in approximately the middle of the freelist. The next time the target entry in the table is allocated it will have exactly the same mach port name as the original target right we triggered the urefs bug on. For this iOS exploit I target the send right to com.apple.iohideventsystem which launchd has, and which I can lookup from inside the container sandbox I look up the iohideventsystem service in launchd then use the urefs bug to free launchd's send right and use the looper messages to spin the generation number round. I then register a large number of dummy services with launchd so that one of them reuses the same mach port name as launchd thinks the iohideventsystem service has. (We can't register global mach services from inside the container sandbox but we can register App Group-restricted services, which work just the same for our purposes. This is why the exploit needs the App Groups capability.) Now when any process looks up com.apple.iohideventsystem launchd will actually send them a send right to one of my dummy services :) I add all those dummy services to a portset and use that recieve right and the legitimate iohideventsystem send right I still have to MITM all these new connections to iohideventsystem. As mentioned earlier clients of iohideventsystem send it their task ports, so all I have to do is crash a process which runs as root and is a client of iohideventsystem. When it restarts it will send it's task port to me :-) *** Powerd crasher *** To crash powerd I use CVE-2016-7661: powerd checks in with launchd to get a server port and then wraps that in a CFPort: ``` pmServerMachPort = _SC_CFMachPortCreateWithPort( "PowerManagement", serverPort, mig_server_callback, &context); ``` It also asks to receive dead name notifications for other ports on that same server port: ``` mach_port_request_notification( mach_task_self(), // task notify_port_in, // port that will die MACH_NOTIFY_DEAD_NAME, // msgid 1, // make-send count CFMachPortGetPort(pmServerMachPort), // notify port MACH_MSG_TYPE_MAKE_SEND_ONCE, // notifyPoly &oldNotify); // previous mig_server_callback is called off of the mach port run loop source to handle new messages on pmServerMachPort: static void mig_server_callback(CFMachPortRef port, void *msg, CFIndex size, void *info) { mig_reply_error_t * bufRequest = msg; mig_reply_error_t * bufReply = CFAllocatorAllocate( NULL, _powermanagement_subsystem.maxsize, 0); mach_msg_return_t mr; int options; __MACH_PORT_DEBUG(true, "mig_server_callback", serverPort); /* we have a request message */ (void) pm_mig_demux(&bufRequest->Head, &bufReply->Head); This passes the raw message to pm_mig_demux: static boolean_t pm_mig_demux( mach_msg_header_t * request, mach_msg_header_t * reply) { mach_dead_name_notification_t *deadRequest = (mach_dead_name_notification_t *)request; boolean_t processed = FALSE; processed = powermanagement_server(request, reply); if (processed) return true; if (MACH_NOTIFY_DEAD_NAME == request->msgh_id) { __MACH_PORT_DEBUG(true, "pm_mig_demux: Dead name port should have 1+ send right(s)", deadRequest->not_port); PMConnectionHandleDeadName(deadRequest->not_port); __MACH_PORT_DEBUG(true, "pm_mig_demux: Deallocating dead name port", deadRequest->not_port); mach_port_deallocate(mach_task_self(), deadRequest->not_port); reply->msgh_bits = 0; reply->msgh_remote_port = MACH_PORT_NULL; return TRUE; } ``` This passes the message to the MIG-generated code for the powermanagement subsystem, if that fails (because the msgh_id doesn't match the subsystem for example) then this compares the message's msgh_id field to MACH_NOTIFY_DEAD_NAME. deadRequest is the message cast to a mach_dead_name_notification_t which is defined like this in mach/notify.h: ``` typedef struct { mach_msg_header_t not_header; NDR_record_t NDR; mach_port_name_t not_port;/* MACH_MSG_TYPE_PORT_NAME */ mach_msg_format_0_trailer_t trailer; } mach_dead_name_notification_t; ``` This is a simple message, not a complex one. not_port is just a completely controlled integer which in this case will get passed directly to mach_port_deallocate. The powerd code expects that only the kernel will send a MACH_NOTIFY_DEAD_NAME message but actually anyone can send this and force the privileged process to drop a reference on a controlled mach port name :) Multiplexing these two things (notifications and a mach service) onto the same port isn't possible to do safely as the kernel doesn't prevent user->user spoofing of notification messages - usually this wouldn't be a problem as attackers shouldn't have access to the notification port. You could probably do quite interesting things with this bug but in this case I just want to crash the service. I do that by spoofing no-more-senders notifications for powerd's task port. Once powerd's send right to its own task port has been freed pretty much everything breaks - in this case I send a copy_powersources_info message, the receving code doesn't check the return value of a call to mach_vm_allocate which fails because the task's task port is wrong and leads to the use of an uninitialized pointer. *** Kernel Bug **** See above for a short writeup of the kernel bug exploit. I will try to write a long-form writeup soon, but the code should be kind of clear. *** Post-exploitation **** I've taken a slightly different approach post-exploitation. Everything is data-only, I don't make any patches to r/o kernel memory. This means things should also work on the iPhone 7 but I don't have one to test :( There are a number of downsides to taking this approach though: * technically a lot of these things I do are racy, but in pratice it works perfectly well enough for a research platform * some things become quite fiddly which are simple with a TEXT patch This is also a research project for me; there are almost certainly far more downsides that I'm not aware of. iOS is complex, undocumented place and I don't really know what I'm doing! The flow works like this: Walk the process list and find the following tasks: amfid mach_portal containermanagerd launchd Disable the sandbox: sb_evaluate has a short-circuit success path if the process has the kern_cred credentials; neither the plaform policy nor the process's sandbox profile will be evaluated. We can use the kernel memory access to give the mach_portal process the kernel's credentials and we're no longer sandboxed. Fix launchd: The sandbox escape made a mess in launchd so I fix up launchd's send right to iohideventsystem to point back to the correct port. I then restart powerd because otherwise we hit a watchdog timeout. Patch amfid: In order to run unsigned binaries and have somethign like a proper shell environment we need to convince amfid to allow binaries with invalid signatures. Previous efforts in this area have replaced amfids import of MISValidateSignature to a function which would always return 0 (success) but amfid now calls MISValidateSignatureAndCopyInfo which takes an out pointer to a CFDictionary which is expected to contain the correct CDHash so just replacing the import won't work. I instead set myself as amfid's exception handler and point the MISValidateSignatureAndCopyInfo to an invalid address. This means that amfid will crash whenever it validates a signature, and since we're the exception handler we get a message on the exception port with the crashing thread state. I read the path to the file to be validated from amfid's address space, compute the CDHash SHA1 myself and write that into the reply message which amfid will send back to the kernel then resume execution of amfid so it can send the reply. Unsandbox containermangerd: Since I haven't had time to investigate LvVM yet I don't remount the rootfs r/w which means that all the binaries we run are from the user partition. This means that we can't prevent the kernel from requesting that containermanagerd allocate a container for them. I did test out doing a similar patch for containermanagerd as I did for amfid which parsed the sb_packbuff requests from the kernel and fixed them up so that containermanagerd didn't get upset but it seemed easier to just unsandbox it so it can make the directories it wants. This decision should be revisited, it's not ideal! Make sure all child processes are also unsandboxed: Since the sandbox defeat involves cheating by using the kern_cred we need a way to make sure all our child processes also have the kern_cred. This is kind of a hack but it works fine for my purposes. You should really revisit this if you want to improve on this code! I allocate a new mach port and set that as my bootstrap port and spin up a thread which mitm's between that port and a real send right to launchd. I request an audit trailer with each message which allows me to get the sender of the message and thus be notified when a new child starts. I then use the kernel memory access to find that pid's proc structure and give it and all its threads the kernel creds. A constructor in libxpc will make a synchronous request to the bootstrap port during dyld initialization before any application code actually runs so this works well enough to allow all our children to run unsandboxed Set kernel task port as host special port: I also set the kernel task port as host special port 4 so you can easily get at it without having to rewrite the exploit code. Shell: I chmod everything in the iosbinpack64 directory to be executable then run bash on a bind shell on port 4141. This isn't ideal but is enough to run test tools and explore the system, talk to all the userclients, devices, mach services, sysctls etc that you want to. **附件：[mach_portal_redist.zip ](https://bugs.chromium.org/p/project-zero/issues/attachment?aid=263891}** [Through the mach portal.pdf ](https://bugs.chromium.org/p/project-zero/issues/attachment?aid=280146) |
| id | SSV:92960 |
| last seen | 2017-11-19 |
| modified | 2017-04-17 |
| published | 2017-04-17 |
| reporter | Root |
| title | XNU kernel UaF due to lack of locking in set_dp_control_port (CVE-2016-7644) |