Xen Security Modules : XSM-FLASK
XSM Documentation
Overview
This is intended to be a hands-on description of Xen Security Modules (XSM) as well as their structure and operation. This document has been compiled based on docs/misc/xsm-flask.txt, mailing list discussions, Xen source code, and other support material contained within the Xen source-tree.
A reader of this document should be able to: 1) Build and install Xen with XSM enabled 2) Create and load a security policy 3) Tag virtual machines with security labels managed by the security policy 4) Utilize a simple application making use of privileges permitted by the policy
What is XSM
Xen offers a security framework called Xen Security Modules, or XSM, to enable an administrator or developer to exert fine-grained control over a Xen domain and its capabilities. Specifically, XSM makes it possible to define permissible interactions between domains, the hypervisor itself, and related resources such as memory and devices.
Via XSM, Xen implements a type of mandatory access control via a security architecture called FLASK using a module of the same name. An abbreviation for Flux Advanced Security Kernel, FLASK has its origins in several trusted operating system research projects, managed jointly between the United States’ National Security Agency, the Secure Computing Corporation, and the University of Utah. The specific intent of FLASK is to separate security enforcement from security policy, better isolating logical components of security systems.
As applied to operating system design, FLASK brings about security enhancements that improve auditing of actions and more granular control thereof. Today, it is perhaps best known through its expression in Security Enhanced Linux (SELinux).
In the same sense that SELinux is a collection of Linux security enhancements, auditing and policy mechanisms, so also is the FLASK Xen Security Module to Xen itself. Similarly, each rely upon a set of administration tools to craft security policies and guide their enforcement.
Below are some illustrative examples of what the FLASK Xen Security Module can make possible:
- Prevent two domains from communicating via event channels or grants - Grant a set of privileged capabilities to otherwise unprivileged domains - Control which domains can use device passthrough (and which devices) - Restrict or audit operations performed by privileged domains - Prevent a privileged domain from arbitrarily mapping pages from other domains - Isolation hypervisor support components (such as qemu or xenstore) from each other, and provide assurance that each may only perform its designated actions
Linux and Xen version requirements for use of XSM
● Xen version 4.3 and newer: Specifically, commit http://xenbits.xen.org/gitweb/?p=xen.git;a=commit;h=996cf47b8081fc6b2efd4fa9a69f6b77028719e6 and later. Earlier versions lack full XSM support (but may still be usable in a restricted sense, depending upon the use case). Prior to this commit, only the sysctl hypercalls and one domctl hypercall XEN_DOMCTL_getdomaininfo were truly usable by non-privileged domains. ● Linux version 3.8 or newer, or alternatively, the adaptation of the following patch to an earlier version (known to apply cleanly to 3.7.1): https://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/commit/?id=30d4b180e20c081f435143f8bc211c66a930608a
Enabling XSM in Xen
XSM is a compile-time option for Xen, and by default it is a disabled feature. To enable the framework, Xen must be compiled with XSM and FLASK explicitly enabled.
To do so, edit Config.mk or the .config file and set XSM_ENABLE and FLASK_ENABLE to "y"
NOTE: this change requires a make clean, rebuild, reinstall and reboot.
Compiling the sample FLASK policy
During typical Xen compilation, compilation of the FLASK sample policy is not performed due to the necessary dependency upon the SElinux compiler, otherwise known as ‘checkpolicy.’ This tool must be installed before proceeding:
Installing checkpolicy ● apt-get install checkpolicy ● yum install checkpolicy ● or similar
Next, to build the sample policy itself: ● cd xen-source-tree/ ● make -C tools/flask/policy
The demonstrtion policy is generated from statements defined in files within the tools/flask/policy path.. After compilation, the resultant policy will be named xenpolicy.24 and reside within the directory tools/flask/policy.
Quoting from its description, this policy illustrates “most of the features of FLASK that can be used without dom0 disaggregation. The main types for domUs are:
- domU_t is a domain that can communicate with any other domU_t - isolated_domU_t can only communicate with dom0 - prot_domU_t is a domain type whose creation can be disabled with a boolean - nomigrate_t is a domain that must be created via the nomigrate_t_building type, and whose memory cannot be read by dom0 once created
HVM domains with stubdomain device models use two types (one per domain):
- domHVM_t is an HVM domain that uses a stubdomain device model - dm_dom_t is the device model for a domain with type domHVM_t”
Booting Xen with FLASK
When Xen is compiled with XSM, the following grub command-line options can be used to set FLASK properties. Using Debian and Ubuntu as examples, these options can be placed in /etc/default/grub within the variable GRUB_CMDLINE_XEN_DEFAULT. Please consult distribution documentation for the recommended method. When package management scripts regenerate the grub configuration (e.g., during a kernel upgrade or at the request of a sysadmin), a system configured in this manner will be guaranteed to properly embed XSM parameters within the grub configuration. For testing purposes, manually editing the grub command line at boot will also suffice.
For example: ● GRUB_CMDLINE_XEN_DEFAULT="flask_enforcing=1”
Xen XSM command line parameters ● flask_enforcing ○ The default value for flask_enforcing is '0'. This parameter causes the platform to boot in permissive mode which means that the policy is loaded but not enforced. This mode is often helpful for developing new systems and policies as the policy violations are reported on the xen console and may be viewed in dom0 through 'xl dmesg'.
To boot the platform into enforcing mode, which means that the policy is loaded and enforced, append 'flask_enforcing=1' on the grub line. ○ This parameter may also be changed through the flask hypercall. ○ This parameter may also be changed through the ‘xl setenforce’ command
● flask_enabled
○ The default value for flask_enabled is '1'. This parameter causes the platform to enable the FLASK security module under the XSM framework. The parameter may be enabled/disabled only once per boot. If the parameter is set to '0', only a reboot can re-enable flask. When flask_enabled is '0' the DUMMY module is enforced; when this module name is encountered in XSM-related documentation and mailing list discussions, it should be interpreted as a “pass through” default XSM policy which matches behavior of a system without XSM enabled at compile-time.
○ This parameter may also be changed through the flask hypercall.
○ This parameter may also be queried through the ‘xl getenforce’ command
○ This parameter may be altered only once per boot.
The compiled policy must be installed into a grub-accessible location (such as /boot where kernels are also stored), as it is loaded during the Xen boot process. Otherwise, FLASK will not be enabled at runtime. It is recommended by the authors of XSM to add an extra subdirectory, such as /boot/flask, to use as a repository for XSM policies.
The policy has to be added to grub as a module after the Xen kernel multiboot line, for example as:
module /boot/flask/xenpolicy.24
To automatically generate this line when running grub-update, the Xen grub-mkconfig helper script has to be modified in /etc/grub.d/20_linux_xen. This helper scripts are different depending on your grub version (and distribution).
Once Xen is booted with the FLASK policy, the policy (or a new one) can be reloaded using the “xl loadpolicy” command.
Creating domains with security labels
Once FLASK is loaded, domains need to be created using a security label in their configuration. Without this line, the domains will be classified “unlabeled”.
seclabel='system_u:system_r:domU_t'
The label’s three sections are: user, role and type. These will be discussed further in this document.
Monitoring XSM logs XSM:Flask will emit avc: denied messages when a permission is denied by the policy, just like SELinux, and can be viewed with xl dmesg:
xl dmesg | grep avc
An example output looks like:
(XEN) avc: denied { setparam } for domid=0 scontext=system_u:system_r:dom0_t tcontext=system_u:system_r:domU_t tclass=hvm
FLASK
The sample FLASK policy can be found in tools/flask/policy/policy/modules/xen. The two files found here are xen.te, the policy description, and xen.if, a collection of macros to be used in the policy description.
Users, roles, types and attributes
Flask provides multiple levels of categories to abstract away from hypercalls.
On the most basic level, a “type” can be defined to specify the hypercalls that type is allowed to execute and how it is allowed to execute them. When XSM is enforcing the policy, only those hypercalls will be executable that are explicitly allowed in the policy.
The sample XSM policy defines several types, for example: xen_t, dom0_t and domU_t. A new type can be defined with the “type” keyword:
type new_type_t;
Each type can be assigned into an attribute category, to easily reference each type having the attribute. The sample policy defines several attributes, for example: xen_type, domain_type, resource_type, etc.
A type can be assigned one or multiple attributes, for example
type xen_t, xen_type, mls_priv;
assigns the xen_t type the attributes xen_type and mls_priv.
Defining which hypercalls are allowed for each type is done using the “allow” keyword. Hypercalls are grouped into security classes (defined in xen/xsm/flask/policy/access_vectors) which will be discussed later. For example the following
allow dom0_t security_t:security check_context;
enables the dom0_t type to execute the check_context hypercall in the security class targeting a security_t type. The allow line can be generalized as follows:
allow <source type> <target type>:<security class> <hypercall>;
Multiple hypercalls belonging to the same security class can be defined by wrapping the hypercalls into {}, for example:
allow dom0_t dom0_t:resource { add remove };
Using the attributes we can quickly allow each type with that attribute access to hypercalls:
# Allow all domains to use (unprivileged parts of) the tmem hypercall allow domain_type xen_t:xen tmem_op;
Roles are the second level above types, where we can define sets of types belonging to a role. While attributes are only accessible internally inside the policy (as a shortcut), a role is part of the security label. A role defines which types belong to that role, therefore it can prevent us from erroneously assigning a type to a VM when it’s role doesn’t have that type.
The sample FLASK policy defines two roles: system_r and vm_r. A role definition is as follows:
role system_r;
role system_r types { xen_type domain_type };
The first line declares a new role while the second line assigns the types belonging to that role (using attributes). When we want to include all types with an attribute except some, we can exclude those types by placing “-” in front of them, like
role vm_r;
role vm_r types { domain_type -dom0_t };
In the above, the vm_r role is defined as all types with the domain_type attribute, except the dom0_t type.
On the highest level we can define users. Users are not defined in the .te (like roles and types), rather they are defined in tools/flask/policy/policy/users. This allows us to define a generic set of users across multiple security modules. By default the FLASK policy will create three users: system_u, customer_1 and customer_2.
A user can be assigned multiple roles but by default the system_u user is assigned the system_r role, and the customer_* users are assigned the vm_r role.
A new user can be defined in the users file as
user <user name> roles { <role1> <role2> };
Constraints
Constraints are defined in the file tools/flask/policy/policy/constraints. Constraints are a safety net to specifically disallow undesired operations that the policy may allow otherwise. The sample Flask policy defines two constraints to prevent event channels and grants between different customers. A constraints syntax is as follows:
constrain <security class> { <hypercall> } ( expression );
For example:
constrain grant { map_read map_write copy } (
u1 == system_u or
u2 == system_u or
u1 == u2
);
The constraint defines that the hypercalls belonging to the security class are only allowed when the expression is true. The expression can contain any user, role and type defined in the policy in the following format:
- expression : ( expression )
- | not expression
- | expression and expression
- | expression or expression
- | u1 op u2
- | r1 role_op r2
- | t1 op t2
- | u1 op names
- | u2 op names
- | r1 op names
- | r2 op names
- | t1 op names
- | t2 op names
- op : == | !=
- role_op : == | != | eq | dom | domby | incomp
- names : name | { name_list }
- name_list : name | name_list name
Macros
Macros are used heavily across the sample Flask policy. For each security module a macro file can be defined as a .if file. The sample macro file (tools/flask/policy/policy/modules/xen/xen.if) contains several commonly used macros, for example for declaring a new domain type and it’s associated sub-types and to setup it’s common allow rules. Macros can call other macros. The macro .if file is required to exist for the policy to compile, even if no macros are defined for a module.
Adding new security modules New security modules can be easily added by editing tools/flask/policy/policy/modules.conf and adding the new modules name as such:
<module_name> = on
The module has to defined in the folder tools/flask/policy/policy/modules/<module_name> with two files:
<module_name>.te <module_name>.if
When multiple security modules are present the modules can’t contain duplicate role or type declarations.
The sample policy is recommended to be left turned on as it contains several definitions required to boot Xen.
Security classes
Security classes are defined in xen/xsm/flask/policy/access_vectors and categorize each hypercall into one of the classes. Each class can hold a maximum of 32 members.
class xen: Class xen consists of operations dealing with the hypervisor itself. Unless otherwise specified, the source is the domain executing the hypercall, and the target is the xen initial sid (type xen_t).
Description Hypercall XENPF_settime settime XEN_SYSCTL_tbuf_op tbufcontrol CONSOLEIO_read, XEN_SYSCTL_readconsole readconsole XEN_SYSCTL_readconsole with clear=1 clearconsole XEN_SYSCTL_perfc_op perfcontrol XENPF_add_memtype mtrr_add XENPF_del_memtype mtrr_del XENPF_read_memtype mtrr_read XENPF_microcode_update microcode XEN_SYSCTL_physinfo, XEN_SYSCTL_topologyinfo, XEN_SYSCTL_numainfo physinfo XENPF_platform_quirk quirk CONSOLEIO_write writeconsole PHYSDEVOP_apic_read, PHYSDEVOP_alloc_irq_vector readapic PHYSDEVOP_apic_write writeapic Most XENOPROF_* privprofile XENOPROF_{init,enable_virq,disable_virq,get_buffer} nonprivprofile kexec hypercall kexec XENPF_firmware_info, XENPF_efi_runtime_call firmware XENPF_enter_acpi_sleep sleep XENPF_change_freq frequency XENPF_getidletime getidle XEN_SYSCTL_debug_keys debug XEN_SYSCTL_getcpuinfo, XENPF_get_cpu_version, XENPF_get_cpuinfo getcpuinfo XEN_SYSCTL_availheap heap XEN_SYSCTL_get_pmstat, XEN_SYSCTL_pm_op, XENPF_set_processor_pminfo, XENPF_core_parking pm_op mca hypercall mca_op XEN_SYSCTL_lockprof_op lockprof XEN_SYSCTL_cpupool_op cpupool_op tmem hypercall (any access) tmem_op TMEM_CONTROL command of tmem hypercall tmem_control XEN_SYSCTL_scheduler_op with XEN_DOMCTL_SCHEDOP_getinfo, XEN_SYSCTL_sched_id getscheduler XEN_SYSCTL_scheduler_op with XEN_DOMCTL_SCHEDOP_putinfo setscheduler
class domain and class domain2:
Classes domain and domain2 consist of operations that a domain performs on
another domain or on itself. Unless otherwise specified, the source is the
domain executing the hypercall, and the target is the domain being operated on
(which may result in a _self or _target type). Transitions in class domain are used to produce the _self and _target types.
class domain:
Description Hypercall XEN_DOMCTL_setvcpucontext setvcpucontext XEN_DOMCTL_pausedomain pause XEN_DOMCTL_unpausedomain unpause XEN_DOMCTL_resumedomain resume XEN_DOMCTL_createdomain create checked in FLASK_RELABEL_DOMAIN for any relabel operation:
source = the old label of the domain target = the new label of the domain
see also the domain2 relabel{from,to,self} permissions transition XEN_DOMCTL_max_vcpus max_vcpus XEN_DOMCTL_destroydomain destroy XEN_DOMCTL_setvcpuaffinity setvcpuaffinity XEN_DOMCTL_getvcpuaffinity getvcpuaffinity XEN_DOMCTL_scheduler_op with XEN_DOMCTL_SCHEDOP_getinfo getscheduler XEN_DOMCTL_getdomaininfo, XEN_SYSCTL_getdomaininfolist getdomaininfo XEN_DOMCTL_getvcpuinfo getvcpuinfo XEN_DOMCTL_getvcpucontext getvcpucontext XEN_DOMCTL_max_mem setdomainmaxmem XEN_DOMCTL_setdomainhandle setdomainhandle XEN_DOMCTL_setdebugging setdebugging XEN_DOMCTL_hypercall_init hypercall XEN_DOMCTL_settimeoffset settime checked in XEN_DOMCTL_set_target:
source = the new device model domain target = the new target domain
see also the domain2 make_priv_for and set_as_target checks set_target SCHEDOP_remote_shutdown shutdown XEN_DOMCTL_set{,_machine}_address_size setaddrsize XEN_DOMCTL_get{,_machine}_address_size getaddrsize XEN_DOMCTL_sendtrigger trigger XEN_DOMCTL_get_ext_vcpucontext getextvcpucontext XEN_DOMCTL_set_ext_vcpucontext setextvcpucontext XEN_DOMCTL_getvcpuextstate getvcpuextstate XEN_DOMCTL_setvcpuextstate setvcpuextstate XENMEM_get_pod_target getpodtarget XENMEM_set_pod_target setpodtarget XEN_DOMCTL_subscribe, XEN_DOMCTL_disable_migrate, XEN_DOMCTL_suppress_spurious_page_faults set_misc_info XEN_DOMCTL_set_virq_handler set_virq_handler
class domain2 (continuation of class domain):
checked in FLASK_RELABEL_DOMAIN with non-DOMID_SELF:
source = the domain making the hypercall target = the old label of the domain being relabeled relabelfrom
checked in FLASK_RELABEL_DOMAIN with non-DOMID_SELF:
source = the domain making the hypercall target = the new label of the domain being relabeled relabelto
checked in FLASK_RELABEL_DOMAIN, only with DOMID_SELF:
source = the old label of the domain target = the new label of the domain
see also domain__transition relabelself checked in XEN_DOMCTL_set_target:
source = the domain making the hypercall target = the new device model domain
make_priv_for checked in XEN_DOMCTL_set_target:
source = the domain making the hypercall target = the new target domain set_as_target
XEN_DOMCTL_set_cpuid set_cpuid XEN_DOMCTL_gettscinfo gettsc XEN_DOMCTL_settscinfo settsc XEN_DOMCTL_scheduler_op with XEN_DOMCTL_SCHEDOP_putinfo setscheduler
class hvm: Similar to class domain, but primarily contains domctls related to HVM domains.
XEN_DOMCTL_sethvmcontext sethvmc XEN_DOMCTL_gethvmcontext, XEN_DOMCTL_gethvmcontext_partial gethvmc HVMOP_set_param setparam HVMOP_get_param getparam HVMOP_set_pci_intx_level (also needs hvmctl) pcilevel HVMOP_set_isa_irq_level irqlevel HVMOP_set_pci_link_route pciroute, bind_irq XEN_DOMCTL_pin_mem_cacheattr cacheattr HVMOP_track_dirty_vram trackdirtyvram HVMOP_modified_memory, HVMOP_get_mem_type, HVMOP_set_mem_type, HVMOP_set_mem_access, HVMOP_get_mem_access, HVMOP_pagetable_dying, HVMOP_inject_trap hvmctl XEN_DOMCTL_set_access_required mem_event XEN_DOMCTL_mem_sharing_op and XENMEM_sharing_op_{share,add_physmap} with:
source = the domain making the hypercall target = domain whose memory is being shared mem_sharing
XEN_DOMCTL_audit_p2m audit_p2m HVMOP_inject_msi send_irq checked in XENMEM_sharing_op_{share,add_physmap} with:
source = domain whose memory is being shared target = client domain share_mem
class event: Class event describes event channels. Interdomain event channels have their own security label which is computed using a type transition between the source and target domains. Each endpoint has its own label, and the permission checks must pass on both endpoints for an event channel to be established.
when creating an interdomain event channel endpoint:
source = event channel label target = remote domain the event channel binds to. This may be a _self or _target label if the endpoints are related as such.
This permission is checked when creating an unbound event channel and when the interdomain event channel is established. bind EVTCHNOP_send:
source = domain sending the event target = event channel label send EVTCHNOP_status; same as _send status
when creating an interdomain event channel endpoint:
source = the domain creating the channel (which might not be an endpoint) target = event channel label create
EVTCHNOP_reset:
source = domain making the hypercall target = domain whose event channels are being reset reset
class grant:
Class grant describes pages shared by grant mappings. Pages use the security label of their owning domain.
GNTTABOP_map_grant_ref with any access map_read GNTTABOP_map_grant_ref with write access map_write GNTTABOP_unmap_grant_ref unmap GNTTABOP_transfer transfer GNTTABOP_setup_table, GNTTABOP_get_status_frames (target is commonly _self) setup GNTTABOP_copy copy GNTTABOP_query_size, GNTTABOP_get_version query
class mmu: Class mmu describes pages of memory not accessed using grants. Permissions are checked using the domain ID used to access the page - the most common case is a domain's own ID (the _self label). Using DOMID_IO in the map command to restrict the mapping to IO memory will result in the target being domio_t, and migration uses read-only mappings with a target of DOMID_XEN (domxen_t).
checked when using mmu_update to map a page readably
source = domain making the hypercall (which might not own the page table) target = domain whose pages are being mapped map_read
checked when using mmu_update to map a page writably
- source = domain making the hypercall
- target = domain whose pages are being mapped
map_write XEN_DOMCTL_getpageframeinfo* pageinfo XEN_DOMCTL_getmemlist pagelist XENMEM_{increase,decrease}_reservation, XENMEM_populate_physmap adjust XENMEM_{current,maximum}_reservation, XENMEM_maximum_gpfn stat mmu_update MMU_MACHPHYS_UPDATE updatemp XENMEM_add_to_physmap, XENMEM_remove_from_physmap physmap MMUEXT_PIN_L*_TABLE pinpage XENMEM_machine_memory_map (with target xen_t), XENMEM_set_memory_map (with domain target) memorymap checked when using mmu_update to update the page tables of another domain
source = domain making the hypercall target = domain whose page tables are being modified
remote_remap the mmuext_op hypercall acting on the target domain mmuext_op XENMEM_exchange:
source = domain making the hypercall target = domain whose pages are being exchanged
exchange
class shadow: Control of the paging_domctl split by subop
Description Operation XEN_DOMCTL_SHADOW_OP_OFF disable enable, get/set allocation enable enable, read, and clean log logdirty
class resource: Class resource is used to describe the resources used in hardware device passthrough. Resources include: hardware IRQs, MMIO regions, x86 I/O ports, and PCI devices; see docs/misc/xsm-flask.txt for how to label them.
Access to the legacy PCI configuration space on x86 via port 0xCF8/CFC requires IS_PRIV, even with FLASK. Writes to the BARs are checked as "setup", while other reads/writes are "use"; the target is the PCI device whose configuration space is being modified. Accesses to the MMIO-based PCI express configuration space described by the ACPI MCFG table are controlled as MMIO accesses, and cannot special-case BAR writes.
The {add,remove}_{irq,ioport,iomem,device} permissions use:
source = domain making the hypercall target = resource's security label
Description Operation
checked when adding a resource to a domain:
source = domain making the hypercall target = domain which will have access to the resource add
checked when removing a resource from a domain:
source = domain making the hypercall target = domain which will no longer have access to the resource remove
checked when adding a resource to a domain:
source = domain which will have access to the resource target = resource's security label
also checked when using some core Xen devices (target xen_t) use PHYSDEVOP_map_pirq and ioapic writes for dom0, when acting on real IRQs
For GSI interrupts, the IRQ's label is indexed by the IRQ number For MSI interrupts, the label of the PCI device is used add_irq
PHYSDEVOP_unmap_pirq (same as map, and only for real IRQs) remove_irq XEN_DOMCTL_ioport_permission, XEN_DOMCTL_ioport_mapping add_ioport, remove_ioport XEN_DOMCTL_iomem_permission, XEN_DOMCTL_memory_mapping add_iomem, remove_iomem XEN_DOMCTL_get_device_group, XEN_DOMCTL_test_assign_device:
source = domain making the hypercall target = PCI device being queried
stat_device XEN_DOMCTL_assign_device add_device XEN_DOMCTL_deassign_device remove_device checked for PCI hot and cold-plug hypercalls, with target as the PCI device checked for CPU and memory hotplug with xen_t as the target plug checked for PCI hot-unplug hypercalls, with target as the PCI device checked for CPU offlining with xen_t as the target unplug checked for PHYSDEVOP_restore_msi* (target PCI device) checked for PHYSDEVOP_setup_gsi (target IRQ) checked for PHYSDEVOP_pci_mmcfg_reserved (target xen_t) setup
class security: Class security describes the FLASK security server itself; these operations are accessed using the xsm_op hypercall. The source is the domain invoking the hypercall, and the target is security_t.
Any domain with access to load_policy or setenforce must be trusted, since it can bypass the rest of the security policy.
Description Operation
use the security server to compute an access check compute_av
use the security server to compute a type transition compute_create
use the security server to compute member selection compute_member
sid <-> context string conversions check_context
allow loading a new XSM/FLASK policy load_policy
use the security server to compute an object relabel compute_relabel
use the security server to list the SIDs reachable by a given user compute_user
allow switching between enforcing and permissive mode setenforce
allow changing policy booleans setbool
allow changing security server configuration parmeters setsecparam
add ocontext label definitions for resources add_ocontext
remove ocontext label definitions for resources del_ocontext
Some of these examples require Dom0 Disaggregation to be useful, since the domain build process requires the ability to write to the new domain's memory.
More information:
