December 2022 - RFC - lists.sel4.systems

[JIRA] (RFC-11) MCS: set fault and timeout handler parameters while configuring TCBs
by Gerwin Klein (Jira) 16 Dec '22

16 Dec '22

1 0

[JIRA] (RFC-9) Add new capability for seL4 SMC Forwarding
by Gerwin Klein (Jira) 16 Dec '22

16 Dec '22

1 0

[JIRA] (RFC-13) MCS: Improve constraints on grant via reply
by Gerwin Klein (Jira) 16 Dec '22

16 Dec '22

1 0

[JIRA] (RFC-13) MCS: Improve constraints on grant via reply
by Gerwin Klein (Jira) 16 Dec '22

16 Dec '22

1 0

[JIRA] (RFC-9) Add new capability for seL4 SMC Forwarding
by Robbie VanVossen (Jira) 02 Dec '22

02 Dec '22

Robbie VanVossen ( https://sel4.atlassian.net/secure/ViewProfile.jspa?accountId=557058%3Ab7eb5… ) *updated* an issue RFCs ( https://sel4.atlassian.net/browse/RFC?atlOrigin=eyJpIjoiMjdiNDNlZTMwNDdlNDQ… ) / RFC ( https://sel4.atlassian.net/browse/RFC-9?atlOrigin=eyJpIjoiMjdiNDNlZTMwNDdlN… ) RFC-9 ( https://sel4.atlassian.net/browse/RFC-9?atlOrigin=eyJpIjoiMjdiNDNlZTMwNDdlN… ) Add new capability for seL4 SMC Forwarding ( https://sel4.atlassian.net/browse/RFC-9?atlOrigin=eyJpIjoiMjdiNDNlZTMwNDdlN… ) Change By: Robbie VanVossen ( https://sel4.atlassian.net/secure/ViewProfile.jspa?accountId=557058%3Ab7eb5… ) h1. Summary This RFC proposes to add a new , badged capability which certain threads can invoke so seL4 can make SMC calls on ARM platforms in EL2 (virtualized mode) on the thread’s behalf. h1. Motivation Currently, seL4 running in EL2 traps all SMC calls. Then , if setup, seL4 decides what calls a handler to do deal with these calls the call. Some SMC calls are emulated, such as PSCI for the camkes-vm , but most are just ignored. However, there may be some SMC calls that actually need to make it to the Supervisor, and this configuration does not allow that to happen. This RFC proposes a solution to this problem by giving threads creating a new badged , badged capability which can be given to specific threads to allows them to forward configured SMC calls to seL4 so that the microkernel will make the actual SMC call on the thread's behalf. h1. Guide-level explanation This change will allow you to forward to the microkernel, emulate, or block SMC calls based on platform specific configuration. When an SMC call is made by a VM, the microkernel intercepts the call, which it does for all SMC calls made at EL1. It then calls {{handle_smc()}} for the corresponding VMM thread. That function will decode the specific SMC call and decide whether to block it, emulate it, or forward the call back into seL4. There are examples already for multiple platforms on emulation and blocking. The zynqmp platform will have an example of forwarding the SMC with the new API call, {{seL4_ARM_SMC_Call()}}. h2. New concepts h3. seL4_ARM_SMC_Call() {{seL4_ARM_SMC_Call()}} is the new ObjectInvocation API call and can be called in two ways: # A VM attempts to make an SMC call which traps into the microkernel. The microkernel informs the VMM’s handler and the VMM decides that the SMC call really needs to be made. The VMM then calls {{seL4_ARM_SMC_Call()}} with the SMC capability and all the arguments that the VM provided. # Directly by constructing the SMC call from the thread with the SMC capability. This means that a trap isn’t needed. Our expected use case has been using the first approach in the camkes-vm for VMM handlers on the behalf of its VM. I am not sure if there is a use case for the second approach, but it is certainly possible. In either case, the function still requires the caller to set all 8 registers, even if the specific SMC call uses less arguments. h3. seL4_ARM_SMC {{seL4_ARM_SMC}} is the new capability. It needs to be given to any threads that you expect to call {{seL4_ARM_SMC_Call()}}, otherwise it will be denied. We expect that in most use cases, it will be given to each VMM thread. The capability is badged and the badge represents the SMC Function Identifier. Therefore, the specific SMC calls that each Thread can make is configurable and checked by seL4. h2. For users We plan to implement this configuration as an example configuration of camkes-vm for the zynqmp platform. Due to WCET implication of this approach, it will not be the default for all platforms. It should only be enabled when actually needed for a specific platform or application. Without this change, Linux is unable to run as a guest on camkes-vm for the zynqmp platform. Here is an example for how to enable SMC calls for some VMs in the {{devices.camkes}} file implemented for a platform in {{camkes-vm-apps}} {code:none}vm0.allow_smc = true; vm0.allowed_smc_functions = [ 0xC2000001, // PM_GET_API_VERSION 0xC2000017 // PM_FPGA_GET_STATUS ]; vm1.allow_smc = true; vm1.allowed_smc_functions = [ 0xC2000001, // PM_GET_API_VERSION 0xC200000D, // PM_REQUEST_NODE 0xC200000E // PM_RELEASE_NODE ]; vm2.allow_smc = false; vm2.allow_smc_functions = [];{code} As you can see in this example, you can configure each VM to allow or not allow any disallow SMC calls. For each VM that has SMC calls enabled, you then specify which SMC function calls they can make. These are integers that are defined per platform. As you can see, VMs can share the same calls functions and they can have different calls functions as well. h1. Reference-level explanation This implementation will meet the SMC Calling Convention specification from ARM ([ARM DEN 0028E v1.4|https://developer.arm.com/documentation/den0028/latest?_ga=2.116565828…] for 32-bit SMC calls. 64-bit SMC calls are partially supported, but full support could be added later. h2. seL4_ARM_SMC_Call() {code:c}static int seL4_ARM_SMC_Call ( seL4_ARM_SMC _service, seL4_ARM_SMCContext * smc_args, seL4_ARM_SMCContext * smc_response ){code} {noformat}Parameters [in] _service Capability to allow threads to interact with SMC calls. [in] smc_args The structure that has the provided arguments. [out] smc_response The structure to capture the responses. Returns{noformat} h2. struct seL4_ARM_SMCContext {code:c}typedef struct seL4_ARM_SMCContext_ { /* register arguments */ seL4_Word x0, x1, x2, x3, x4, x5, x6, x7; } seL4_ARM_SMCContext;{code} h1. Drawbacks There are a couple of drawbacks to allowing SMC calls to actually go through. Both relate to allowing a supervisor to do anything in the system past initialization time. One drawback is that the worst case execution time (WCET) is no longer guaranteed for the system. Each SMC call can take an arbitrarily long amount of time. This means a thread can essentially lock up all of the cores for an unspecified amount of time. The core the SMC call is made on is obviously blocked because the supervisor will not return to seL4 until it completes its function. The other cores will become locked as soon as seL4 runs and tries to grab the Big Kernel Lock (BKL). One potential fix for the multicore issue is to release the BKL before making the SMC call. However, that is out of the scope of this RFC as it has a lot of implications. The current approach will be to have this config option off by default and add documentation about how enabling it affects WCET. Another drawback is the effect this has on assurance. Since the Supervisor is a higher privilege level than seL4, it can make changes to the system of which seL4 is unaware. This would likely add on to the assumptions made if the proof ever included this configuration setting. The proofs would have to assume that the SMC calls that are made will not invalidate the proof. This is on the User to look at what each allowed SMC call does and check that it won’t interfere with the proof. h1. Rationale and alternatives This design is the best because it gives users flexibility by allowing them to make some SMC calls while still being able to emulate other SMC calls. It is the preferred method, because the microkernel shouldn't need to be updated with each new platform, VMM configuration, and secure application implementation. It can all be configured in user space. h2. Alternative approaches SMC filtering could be done in the microkernel so that developers are less likely to configure the system incorrectly. However, there are some cases where the filtering is dependent on specific threads and the application level (Secure world communication) instead of just platform dependent. On the first point, instead of creating a cap for making the SMC call, you could have a cap for each numbered SMC call so that you could configure what SMC call each thread can make using the root thread. However, that is quite a bit of work as that will differ for each ARM platform and adds another step when porting the microkernel to new ARM platforms. That could also solve the second point, but then users would likely need to modify the microkernel to support their specific secure world application, which is something that should really be avoided. Another part of the implementation that could be done differently is using the VCPU capability with an access right instead of creating a new type of capability. The kinds of access rights are already determined as read, write, grant, and grantreply. A write access right would probably make the most sense, but it still seems obtuse that a write access on a VCPU capability would allow the invoker to make an SMC call. Besides that, there are two other issues I see with this approach: # I can foresee a use-case where you have a health monitor component that isn’t associated with a VCPU but needs to be able to forward an SMC call to actually interact with EL3 or secure applications. # If the VCPU capability ever needs read/write access rights that make more sense with the type of object it is, for example, limiting which threads can actually read or write that VCPUs registers. For those reasons we believe a new capability makes the most sense. h2. Impact of not implementing this Not implementing this will limit the ARM platforms we can use for the camkes-vm and/or the guests that can run on those platforms. It will also limit support for applications that run in secure side of the trustzone that need to communicate with the nonsecure side (Where sel4, VMMs, and the VMs run). h1. Prior art Xen has a similar implementation for the Zynqmp where arbitration is done to allow/disallow/emulate SMC calls from guest VMs. The arbitration is quite specific and is automatically generated based on resource allocation for each VM. Xen does not have VMM pieces that run in Userspace. Basically all of the management runs in EL2. Therefore, all of the arbitration is done in Xen itself. This approach does not make sense for seL4 for the reasons stated above in the [#rationale|#rationale] section. However, this does show a need for this feature in hypervisors on ARMv8 platforms. If necessary, we could reference the Xen arbitration code to help improve the default arbitration for the zynqmp in our implementation. h1. Unresolved questions # How do we address the WCET issue? (Doesn’t need to be solved by this RFC because we are documenting the effects) ( https://sel4.atlassian.net/browse/RFC-9#add-comment?atlOrigin=eyJpIjoiMjdiN… ) Add Comment ( https://sel4.atlassian.net/browse/RFC-9#add-comment?atlOrigin=eyJpIjoiMjdiN… ) Get Jira notifications on your phone! Download the Jira Cloud app for Android ( https://play.google.com/store/apps/details?id=com.atlassian.android.jira.co… ) or iOS ( https://itunes.apple.com/app/apple-store/id1006972087?pt=696495&ct=EmailNot… ) This message was sent by Atlassian Jira (v1001.0.0-SNAPSHOT#100211- sha1:6a2b617 )

1 0

[JIRA] (RFC-9) Add new capability for seL4 SMC Forwarding
by Robbie VanVossen (Jira) 02 Dec '22

02 Dec '22

Robbie VanVossen ( https://sel4.atlassian.net/secure/ViewProfile.jspa?accountId=557058%3Ab7eb5… ) *updated* an issue RFCs ( https://sel4.atlassian.net/browse/RFC?atlOrigin=eyJpIjoiMDA0MjRmMzBkMmVjNDQ… ) / RFC ( https://sel4.atlassian.net/browse/RFC-9?atlOrigin=eyJpIjoiMDA0MjRmMzBkMmVjN… ) RFC-9 ( https://sel4.atlassian.net/browse/RFC-9?atlOrigin=eyJpIjoiMDA0MjRmMzBkMmVjN… ) Add new capability for seL4 SMC Forwarding ( https://sel4.atlassian.net/browse/RFC-9?atlOrigin=eyJpIjoiMDA0MjRmMzBkMmVjN… ) Change By: Robbie VanVossen ( https://sel4.atlassian.net/secure/ViewProfile.jspa?accountId=557058%3Ab7eb5… ) h1. Summary This RFC proposes to add a new capability which certain threads can invoke so seL4 can make SMC calls on ARM platforms in EL2 (virtualized mode) on the thread’s behalf. h1. Motivation Currently, seL4 running in EL2 traps all SMC calls. Then seL4 decides what to do with these calls. Some SMC calls are emulated, such as PSCI, but most are just ignored. However, there may be some SMC calls that actually need to make it to the Supervisor, and this configuration does not allow that to happen. This RFC proposes a solution to this problem by giving threads a new badged, capability which allows them to forward SMC calls to seL4 so that the microkernel will make the actual SMC call on the thread's behalf. h1. Guide-level explanation This change will allow you to forward to the microkernel, emulate, or block SMC calls based on platform specific configuration. When an SMC call is made by a VM, the microkernel intercepts the call, which it does for all SMC calls made at EL1. It then calls {{handle_smc()}} for the corresponding VMM thread. That function will decode the specific SMC call and decide whether to block it, emulate it, or forward the call back into seL4. There are examples already for multiple platforms on emulation and blocking. The zynqmp platform will have an example of forwarding the SMC with the new API call, {{seL4_ARM_SMC_Call()}}. h2. New concepts h3. seL4_ARM_SMC_Call() {{seL4_ARM_SMC_Call()}} is the new ObjectInvocation API call and can be called in two ways: # A VM attempts to make an SMC call which traps into the microkernel. The microkernel informs the VMM’s handler and the VMM decides that the SMC call really needs to be made. The VMM then calls {{seL4_ARM_SMC_Call()}} with the SMC capability and all the arguments that the VM provided. # Directly by constructing the SMC call from the thread with the SMC capability. This means that a trap isn’t needed. Our expected use case has been using the first approach in the camkes-vm for VMM handlers on the behalf of its VM. I am not sure if there is a use case for the second approach, but it is certainly possible. In either case, the function still requires the caller to set all 8 registers, even if the specific SMC call uses less arguments. h3. seL4_ARM_SMC {{seL4_ARM_SMC}} is the new capability. It needs to be given to any threads that you expect to call {{seL4_ARM_SMC_Call()}}, otherwise it will be denied. We expect that in most use cases, it will be given to each VMM thread. The capability is badged and the badge represents the SMC Function Identifier. Therefore, the specific SMC calls that each Thread can make is configurable and checked by seL4. h2. For users We plan to implement this configuration as an example configuration of camkes-vm for the zynqmp platform. Due to WCET implication of this approach, it will not be the default for all platforms. It should only be enabled when actually needed for a specific platform or application. Without this change, Linux is unable to run as a guest on camkes-vm for the zynqmp platform. Here is an example for how to enable SMC calls for some VMs in the {{devices.camkes}} file implemented for a platform in {{camkes-vm-apps}} {code:none}vm0.allow_smc = true; vm0.allowed_smc_functions = [ 0xC2000001, // PM_GET_API_VERSION 0xC2000017 // PM_FPGA_GET_STATUS ]; vm1.allow_smc = true; vm1.allowed_smc_functions = [ 0xC2000001, // PM_GET_API_VERSION 0xC200000D, // PM_REQUEST_NODE 0xC200000E // PM_RELEASE_NODE ]; vm2.allow_smc = false; vm2.allow_smc_functions = [];{code} As you can see in this example, you can configure each VM to allow or not allow any SMC calls. For each VM that has SMC calls enabled, you then specify which SMC function calls they can make. These are integers that are defined per platform. As you can see, VMs can share the same calls and they can have different calls as well. h1. Reference-level explanation This implementation will meet the SMC Calling Convention specification from ARM ([ARM DEN 0028E v1.4|https://developer.arm.com/documentation/den0028/latest?_ga=2.116565828…] for 32-bit SMC calls. 64-bit SMC calls are partially supported, but full support could be added later. h2. seL4_ARM_SMC_Call() {code:c}static int seL4_ARM_SMC_Call ( seL4_ARM_SMC _service, seL4_ARM_SMCContext * smc_args, seL4_ARM_SMCContext * smc_response ){code} {noformat}Parameters [in] _service Capability to allow threads to interact with SMC calls. [in] smc_args The structure that has the provided arguments. [out] smc_response The structure to capture the responses. Returns{noformat} h2. struct seL4_ARM_SMCContext {code:c}typedef struct seL4_ARM_SMCContext_ { /* register arguments */ seL4_Word x0, x1, x2, x3, x4, x5, x6, x7; } seL4_ARM_SMCContext;{code} h1. Drawbacks The main drawback of this approach is that you There are a couple of drawbacks to allowing SMC calls to actually go through. Both relate to allowing a supervisor to do anything in the system past initialization time. This means One drawback is that the worst case execution time is no longer guaranteed for the system. Each SMC call can take an arbitrarily long amount of time. This means a thread can essentially lock up all of the cores for an unspecified amount of time. The core the SMC call is made on is obviously blocked because the supervisor will not return to seL4 until it completes its function. The other cores will become locked as soon as seL4 runs and tries to grab the Big Kernel Lock (BKL). One potential fix for the multicore issue is to release the BKL before making the SMC call. However, that is out of the scope of this RFC as it has a lot of implications. The current approach will be to have this config option off by default and add documentation about how enabling it affects WCET. Another drawback is the effect this has on assurance. Since the Supervisor is a higher privilege level than seL4, it can make changes to the system of which seL4 is unaware. This would likely add on to the assumptions made if the proof ever included this configuration setting. The proofs would have to assume that the SMC calls that are made will not invalidate the proof. This is on the User to look at what each allowed SMC call does and check that it won’t interfere with the proof. h1. Rationale and alternatives This design is the best because it gives users flexibility by allowing them to make some SMC calls while still being able to emulate other SMC calls. It is the preferred method, because the microkernel shouldn't need to be updated with each new platform, VMM configuration, and secure application implementation. It can all be configured in user space. h2. Alternative approaches SMC filtering could be done in the microkernel so that developers are less likely to configure the system incorrectly. However, there are some cases where the filtering is dependent on specific threads and the application level (Secure world communication) instead of just platform dependent. On the first point, instead of creating a cap for making the SMC call, you could have a cap for each numbered SMC call so that you could configure what SMC call each thread can make using the root thread. However, that is quite a bit of work as that will differ for each ARM platform and adds another step when porting the microkernel to new ARM platforms. That could also solve the second point, but then users would likely need to modify the microkernel to support their specific secure world application, which is something that should really be avoided. Another part of the implementation that could be done differently is using the VCPU capability with an access right instead of creating a new type of capability. The kinds of access rights are already determined as read, write, grant, and grantreply. A write access right would probably make the most sense, but it still seems obtuse that a write access on a VCPU capability would allow the invoker to make an SMC call. Besides that, there are two other issues I see with this approach: # I can foresee a use-case where you have a health monitor component that isn’t associated with a VCPU but needs to be able to forward an SMC call to actually interact with EL3 or secure applications. # If the VCPU capability ever needs read/write access rights that make more sense with the type of object it is, for example, limiting which threads can actually read or write that VCPUs registers. For those reasons we believe a new capability makes the most sense. h2. Impact of not implementing this Not implementing this will limit the ARM platforms we can use for the camkes-vm and/or the guests that can run on those platforms. It will also limit support for applications that run in secure side of the trustzone that need to communicate with the nonsecure side (Where sel4, VMMs, and the VMs run). h1. Prior art Xen has a similar implementation for the Zynqmp where arbitration is done to allow/disallow/emulate SMC calls from guest VMs. The arbitration is quite specific and is automatically generated based on resource allocation for each VM. Xen does not have VMM pieces that run in Userspace. Basically all of the management runs in EL2. Therefore, all of the arbitration is done in Xen itself. This approach does not make sense for seL4 for the reasons stated above in the [#rationale|#rationale] section. However, this does show a need for this feature in hypervisors on ARMv8 platforms. If necessary, we could reference the Xen arbitration code to help improve the default arbitration for the zynqmp in our implementation. h1. Unresolved questions # Should this implementation be configurable? ## I would argue it makes the most sense to have it automatically enabled if you are using the VMM on ARM and {{KernelAllowSMCCalls}} is disabled. Then your default whitelist in the SMC handler, for most platforms, can be to allow no SMC calls through, which would mean the default configuration would essentially be the same as the second configuration that is defined in the motivation section. However, I am wondering if there is a case where you still just want to disable it completely? # Should the microkernel have an overriding blacklist? ## How and who decides what do we address the blacklist is WCET issue ? ## Is it configurable? ## I don' (Doesn’ t think need to be solved by this question blocks the RFC as because we are documenting the blacklist could be implemented easily enough after the proposed implementation. effects) ( https://sel4.atlassian.net/browse/RFC-9#add-comment?atlOrigin=eyJpIjoiMDA0M… ) Add Comment ( https://sel4.atlassian.net/browse/RFC-9#add-comment?atlOrigin=eyJpIjoiMDA0M… ) Get Jira notifications on your phone! Download the Jira Cloud app for Android ( https://play.google.com/store/apps/details?id=com.atlassian.android.jira.co… ) or iOS ( https://itunes.apple.com/app/apple-store/id1006972087?pt=696495&ct=EmailNot… ) This message was sent by Atlassian Jira (v1001.0.0-SNAPSHOT#100211- sha1:6a2b617 )

1 0

[JIRA] (RFC-9) Add new capability for seL4 SMC Forwarding
by Robbie VanVossen (Jira) 02 Dec '22

02 Dec '22

Robbie VanVossen ( https://sel4.atlassian.net/secure/ViewProfile.jspa?accountId=557058%3Ab7eb5… ) *updated* an issue RFCs ( https://sel4.atlassian.net/browse/RFC?atlOrigin=eyJpIjoiNzQwZjI2NzE0M2Y2NDh… ) / RFC ( https://sel4.atlassian.net/browse/RFC-9?atlOrigin=eyJpIjoiNzQwZjI2NzE0M2Y2N… ) RFC-9 ( https://sel4.atlassian.net/browse/RFC-9?atlOrigin=eyJpIjoiNzQwZjI2NzE0M2Y2N… ) Add new capability for seL4 SMC Forwarding ( https://sel4.atlassian.net/browse/RFC-9?atlOrigin=eyJpIjoiNzQwZjI2NzE0M2Y2N… ) Change By: Robbie VanVossen ( https://sel4.atlassian.net/secure/ViewProfile.jspa?accountId=557058%3Ab7eb5… ) h1. Summary This RFC proposes to add a new capability which certain threads can invoke so seL4 can make SMC calls on ARM platforms in EL2 (virtualized mode) on the thread’s behalf. h1. Motivation Currently, seL4 running in EL2 traps all SMC calls. Then seL4 decides what to do with these calls. Some SMC calls are emulated, such as PSCI, but most are just ignored. However, there may be some SMC calls that actually need to make it to the Supervisor, and this configuration does not allow that to happen. This RFC proposes a solution to this problem by giving threads a new badged, capability which allows them to forward SMC calls to seL4 so that the microkernel will make the actual SMC call on the thread's behalf. h1. Guide-level explanation This change will allow you to forward to the microkernel, emulate, or block SMC calls based on platform specific configuration. When an SMC call is made by a VM, the microkernel intercepts the call, which it does for all SMC calls made at EL1. It then calls {{handle_smc()}} for the corresponding VMM thread. That function will decode the specific SMC call and decide whether to block it, emulate it, or forward the call back into seL4. There are examples already for multiple platforms on emulation and blocking. The zynqmp platform will have an example of forwarding the SMC with the new API call, {{seL4_ARM_SMC_Call()}}. h2. New concepts h3. seL4_ARM_SMC_Call() {{seL4_ARM_SMC_Call()}} is the new ObjectInvocation API call and can be called in two ways: # A VM attempts to make an SMC call which traps into the microkernel. The microkernel informs the VMM’s handler and the VMM decides that the SMC call really needs to be made. The VMM then calls {{seL4_ARM_SMC_Call()}} with the SMC capability and all the arguments that the VM provided. # Directly by constructing the SMC call from the thread with the SMC capability. This means that a trap isn’t needed. Our expected use case has been using the first approach in the camkes-vm for VMM handlers on the behalf of its VM. I am not sure if there is a use case for the second approach, but it is certainly possible. In either case, the function still requires the caller to set all 8 registers, even if the specific SMC call uses less arguments. h3. seL4_ARM_SMC {{seL4_ARM_SMC}} is the new capability. It needs to be given to any threads that you expect to call {{seL4_ARM_SMC_Call()}}, otherwise it will be denied. We expect that in most use cases, it will be given to each VMM thread. The capability is badged and the badge represents the SMC Function Identifier. Therefore, the specific SMC calls that each Thread can make is configurable and checked by seL4. h2. For users We plan to implement this configuration as on for an example configuration of camkes-vm for the zynqmp platform. Due to WCET implication of this approach, it will not be the default for all platforms. It should only be enabled when actually needed for a specific platform or application. Without this change, Linux is unable to run as a guest on camkes-vm for the zynqmp platform. Here is an example for how to enable SMC calls for some VMs in the {{devices.camkes}} file implemented for a platform in {{camkes-vm-apps}} {code:none}vm0.allow_smc = true; vm0.allowed_smc_functions = [ 0xC2000001, // PM_GET_API_VERSION 0xC2000017 // PM_FPGA_GET_STATUS ]; vm1.allow_smc = true; vm1.allowed_smc_functions = [ 0xC2000001, // PM_GET_API_VERSION 0xC200000D, // PM_REQUEST_NODE 0xC200000E // PM_RELEASE_NODE ]; vm2.allow_smc = false; vm2.allow_smc_functions = [];{code} As you can see in this example, you can configure each VM to allow or not allow any SMC calls. For each VM that has SMC calls enabled, you then specify which SMC function calls they can make. These are integers that are defined per platform. As you can see, VMs can share the same calls and they can have different calls as well. h1. Reference-level explanation This implementation will meet the SMC Calling Convention specification from ARM ([ARM DEN 0028E v1.4|https://developer.arm.com/documentation/den0028/latest?_ga=2.116565828…] for 32-bit SMC calls. 64-bit SMC calls are partially supported, but full support could be added later. h2. seL4_ARM_SMC_Call() {code:c}static int seL4_ARM_SMC_Call ( seL4_ARM_SMC _service, seL4_ARM_SMCContext * smc_args, seL4_ARM_SMCContext * smc_response ){code} {noformat}Parameters [in] _service Capability to allow threads to interact with SMC calls. [in] smc_args The structure that has the provided arguments. [out] smc_response The structure to capture the responses. Returns{noformat} h2. struct seL4_ARM_SMCContext {code:c}typedef struct seL4_ARM_SMCContext_ { /* register arguments */ seL4_Word x0, x1, x2, x3, x4, x5, x6, x7; } seL4_ARM_SMCContext;{code} h1. Drawbacks The main drawback of this approach is that you are allowing a supervisor in the system past initialization. This means that the worst case execution time is no longer guaranteed for the system. Each SMC call can take an arbitrarily long amount of time. This means a thread can essentially lock up all of the cores for an unspecified amount of time. The core the SMC call is made on is obviously blocked because the supervisor will not return to seL4 until it completes its function. The other cores will become locked as soon as seL4 runs and tries to grab the Big Kernel Lock (BKL). One potential fix for the multicore issue is to release the BKL before making the SMC call. However, that is out of the scope of this RFC as it has a lot of implications. The current approach will be to have this config option off by default and add documentation about how enabling it affects WCET. h1. Rationale and alternatives This design is the best because it gives users flexibility by allowing them to make some SMC calls while still being able to emulate other SMC calls. It is the preferred method, because the microkernel shouldn't need to be updated with each new platform, VMM configuration, and secure application implementation. It can all be configured in user space. h2. Alternative approaches SMC filtering could be done in the microkernel so that developers are less likely to configure the system incorrectly. However, there are some cases where the filtering is dependent on specific threads and the application level (Secure world communication) instead of just platform dependent. On the first point, instead of creating a cap for making the SMC call, you could have a cap for each numbered SMC call so that you could configure what SMC call each thread can make using the root thread. However, that is quite a bit of work as that will differ for each ARM platform and adds another step when porting the microkernel to new ARM platforms. That could also solve the second point, but then users would likely need to modify the microkernel to support their specific secure world application, which is something that should really be avoided. Another part of the implementation that could be done differently is using the VCPU capability with an access right instead of creating a new type of capability. The kinds of access rights are already determined as read, write, grant, and grantreply. A write access right would probably make the most sense, but it still seems obtuse that a write access on a VCPU capability would allow the invoker to make an SMC call. Besides that, there are two other issues I see with this approach: # I can foresee a use-case where you have a health monitor component that isn’t associated with a VCPU but needs to be able to forward an SMC call to actually interact with EL3 or secure applications. # If the VCPU capability ever needs read/write access rights that make more sense with the type of object it is, for example, limiting which threads can actually read or write that VCPUs registers. For those reasons we believe a new capability makes the most sense. h2. Impact of not implementing this Not implementing this will limit the ARM platforms we can use for the camkes-vm and/or the guests that can run on those platforms. It will also limit support for applications that run in secure side of the trustzone that need to communicate with the nonsecure side (Where sel4, VMMs, and the VMs run). h1. Prior art Xen has a similar implementation for the Zynqmp where arbitration is done to allow/disallow/emulate SMC calls from guest VMs. The arbitration is quite specific and is automatically generated based on resource allocation for each VM. Xen does not have VMM pieces that run in Userspace. Basically all of the management runs in EL2. Therefore, all of the arbitration is done in Xen itself. This approach does not make sense for seL4 for the reasons stated above in the [#rationale|#rationale] section. However, this does show a need for this feature in hypervisors on ARMv8 platforms. If necessary, we could reference the Xen arbitration code to help improve the default arbitration for the zynqmp in our implementation. h1. Unresolved questions # Should this implementation be configurable? ## I would argue it makes the most sense to have it automatically enabled if you are using the VMM on ARM and {{KernelAllowSMCCalls}} is disabled. Then your default whitelist in the SMC handler, for most platforms, can be to allow no SMC calls through, which would mean the default configuration would essentially be the same as the second configuration that is defined in the motivation section. However, I am wondering if there is a case where you still just want to disable it completely? # Should the microkernel have an overriding blacklist? ## How and who decides what the blacklist is? ## Is it configurable? ## I don't think this question blocks the RFC as the blacklist could be implemented easily enough after the proposed implementation. ( https://sel4.atlassian.net/browse/RFC-9#add-comment?atlOrigin=eyJpIjoiNzQwZ… ) Add Comment ( https://sel4.atlassian.net/browse/RFC-9#add-comment?atlOrigin=eyJpIjoiNzQwZ… ) Get Jira notifications on your phone! Download the Jira Cloud app for Android ( https://play.google.com/store/apps/details?id=com.atlassian.android.jira.co… ) or iOS ( https://itunes.apple.com/app/apple-store/id1006972087?pt=696495&ct=EmailNot… ) This message was sent by Atlassian Jira (v1001.0.0-SNAPSHOT#100211- sha1:6a2b617 )

1 0

[JIRA] (RFC-9) Add new capability for seL4 SMC Forwarding
by Robbie VanVossen (Jira) 02 Dec '22

02 Dec '22

Robbie VanVossen ( https://sel4.atlassian.net/secure/ViewProfile.jspa?accountId=557058%3Ab7eb5… ) *updated* an issue RFCs ( https://sel4.atlassian.net/browse/RFC?atlOrigin=eyJpIjoiOTY3ZDUzM2E0NDNlNDJ… ) / RFC ( https://sel4.atlassian.net/browse/RFC-9?atlOrigin=eyJpIjoiOTY3ZDUzM2E0NDNlN… ) RFC-9 ( https://sel4.atlassian.net/browse/RFC-9?atlOrigin=eyJpIjoiOTY3ZDUzM2E0NDNlN… ) Add new capability for seL4 SMC Forwarding ( https://sel4.atlassian.net/browse/RFC-9?atlOrigin=eyJpIjoiOTY3ZDUzM2E0NDNlN… ) Change By: Robbie VanVossen ( https://sel4.atlassian.net/secure/ViewProfile.jspa?accountId=557058%3Ab7eb5… ) h1. Summary This RFC proposes to add a new capability which certain threads can invoke so seL4 can make SMC calls on ARM platforms in EL2 (virtualized mode) on the thread’s behalf. h1. Motivation Currently, seL4 running in EL2 traps all SMC calls. Then seL4 decides what to do with these calls. Some SMC calls are emulated, such as PSCI, but most are just ignored. However, there may be some SMC calls that actually need to make it to the Supervisor, and this configuration does not allow that to happen. This RFC proposes a solution to this problem by giving threads a new badged, capability which allows them to forward SMC calls to seL4 so that the microkernel will make the actual SMC call on the thread's behalf. h1. Guide-level explanation This change will allow you to forward to the microkernel, emulate, or block SMC calls based on platform specific configuration. When an SMC call is made by a VM, the microkernel intercepts the call, which it does for all SMC calls made at EL1. It then calls {{handle_smc()}} for the corresponding VMM thread. That function will decode the specific SMC call and decide whether to block it, emulate it, or forward the call back into seL4. There are examples already for multiple platforms on emulation and blocking. The zynqmp platform will have an example of forwarding the SMC with the new API call, {{seL4_ARM_SMC_Call()}}. h2. New concepts h3. seL4_ARM_SMC_Call() {{seL4_ARM_SMC_Call()}} is the new ObjectInvocation API call and can be called in two ways: # A VM attempts to make an SMC call which traps into the microkernel. The microkernel informs the VMM’s handler and the VMM decides that the SMC call really needs to be made. The VMM then calls {{seL4_ARM_SMC_Call()}} with the SMC capability and all the arguments that the VM provided. # Directly by constructing the SMC call from the thread with the SMC capability. This means that a trap isn’t needed. Our expected use case has been using the first approach in the camkes-vm for VMM handlers on the behalf of its VM. I am not sure if there is a use case for the second approach, but it is certainly possible. In either case, the function still requires the caller to set all 8 registers, even if the specific SMC call uses less arguments. h3. seL4_ARM_SMC {{seL4_ARM_SMC}} is the new capability. It needs to be given to any threads that you expect to call {{seL4_ARM_SMC_Call()}}, otherwise it will be denied. We expect that in most use cases, it will be given to each VMM thread. The capability is badged and the badge represents the SMC Function Identifier. Therefore, the specific SMC calls that each Thread can make is configurable and checked by seL4. h2. For users We plan to implement this configuration as the default on for users an example configuration of camkes-vm for the zynqmp platform. Due to WCET implication of this approach, it will not be the default for all platforms. It should only be enabled when actually needed for a specific platform or application. Without this change, Linux is unable to run as a guest on camkes-vm for the zynqmp platform. If a user enables SMC forwarding and needs Here is an example for how to update enable SMC calls for some VMs in the configuration {{devices.camkes}} file implemented for their a platform , they can make changes to in {{ projects camkes-vm-apps}} {code:none}vm0.allow_smc = true; vm0.allowed_smc_functions = [ 0xC2000001, / seL4_projects_libs / libsel4vmmplatsupport PM_GET_API_VERSION 0xC2000017 / src / plat PM_FPGA_GET_STATUS ]; vm1.allow_smc = true; vm1.allowed_smc_functions = [ 0xC2000001, / $KernelPlatform / devices PM_GET_API_VERSION 0xC200000D, / smc / PM_REQUEST_NODE 0xC200000E // PM_RELEASE_NODE ]; vm2. c allow_smc = false; vm2.allow_smc_functions = [];{code } } As you can see in this example, you can configure each VM to allow or not allow any SMC calls. For each VM that has SMC calls enabled, you then specify which SMC function calls they can make. These are integers that are defined per platform. As you can see, VMs can share the same calls and they can have different calls as well. h1. Reference-level explanation This implementation will meet the SMC Calling Convention specification from ARM ([ARM DEN 0028E v1.4|https://developer.arm.com/documentation/den0028/latest?_ga=2.116565828…] for 32-bit SMC calls. 64-bit SMC calls are partially supported, but full support could be added later. h2. seL4_ARM_SMC_Call() {code:c}static int seL4_ARM_SMC_Call ( seL4_ARM_SMC _service, seL4_ARM_SMCContext * smc_args, seL4_ARM_SMCContext * smc_response ){code} {noformat}Parameters [in] _service Capability to allow threads to interact with SMC calls. [in] smc_args The structure that has the provided arguments. [out] smc_response The structure to capture the responses. Returns{noformat} h2. struct seL4_ARM_SMCContext {code:c}typedef struct seL4_ARM_SMCContext_ { /* register arguments */ seL4_Word x0, x1, x2, x3, x4, x5, x6, x7; } seL4_ARM_SMCContext;{code} h1. Drawbacks The main drawback of this approach is that you are allowing a supervisor in the system past initialization. This means that the worst case execution time is no longer guaranteed for the system. Each SMC call can take an arbitrarily long amount of time. This means a thread can essentially lock up all of the cores for an unspecified amount of time. The core the SMC call is made on is obviously blocked because the supervisor will not return to seL4 until it completes its function. The other cores will become locked as soon as seL4 runs and tries to grab the Big Kernel Lock (BKL). One potential fix for the multicore issue is to release the BKL before making the SMC call. However, that is out of the scope of this RFC as it has a lot of implications. The current approach will be to have this config option off by default and add documentation about how enabling it affects WCET. h1. Rationale and alternatives This design is the best because it gives users flexibility by allowing them to make some SMC calls while still being able to emulate other SMC calls. It is the preferred method, because the microkernel shouldn't need to be updated with each new platform, VMM configuration, and secure application implementation. It can all be configured in user space. h2. Alternative approaches SMC filtering could be done in the microkernel so that developers are less likely to configure the system incorrectly. However, there are some cases where the filtering is dependent on specific threads and the application level (Secure world communication) instead of just platform dependent. On the first point, instead of creating a cap for making the SMC call, you could have a cap for each numbered SMC call so that you could configure what SMC call each thread can make using the root thread. However, that is quite a bit of work as that will differ for each ARM platform and adds another step when porting the microkernel to new ARM platforms. That could also solve the second point, but then users would likely need to modify the microkernel to support their specific secure world application, which is something that should really be avoided. Another part of the implementation that could be done differently is using the VCPU capability with an access right instead of creating a new type of capability. The kinds of access rights are already determined as read, write, grant, and grantreply. A write access right would probably make the most sense, but it still seems obtuse that a write access on a VCPU capability would allow the invoker to make an SMC call. Besides that, there are two other issues I see with this approach: # I can foresee a use-case where you have a health monitor component that isn’t associated with a VCPU but needs to be able to forward an SMC call to actually interact with EL3 or secure applications. # If the VCPU capability ever needs read/write access rights that make more sense with the type of object it is, for example, limiting which threads can actually read or write that VCPUs registers. For those reasons we believe a new capability makes the most sense. h2. Impact of not implementing this Not implementing this will limit the ARM platforms we can use for the camkes-vm and/or the guests that can run on those platforms. It will also limit support for applications that run in secure side of the trustzone that need to communicate with the nonsecure side (Where sel4, VMMs, and the VMs run). h1. Prior art Xen has a similar implementation for the Zynqmp where arbitration is done to allow/disallow/emulate SMC calls from guest VMs. The arbitration is quite specific and is automatically generated based on resource allocation for each VM. Xen does not have VMM pieces that run in Userspace. Basically all of the management runs in EL2. Therefore, all of the arbitration is done in Xen itself. This approach does not make sense for seL4 for the reasons stated above in the [#rationale|#rationale] section. However, this does show a need for this feature in hypervisors on ARMv8 platforms. If necessary, we could reference the Xen arbitration code to help improve the default arbitration for the zynqmp in our implementation. h1. Unresolved questions # Should this implementation be configurable? ## I would argue it makes the most sense to have it automatically enabled if you are using the VMM on ARM and {{KernelAllowSMCCalls}} is disabled. Then your default whitelist in the SMC handler, for most platforms, can be to allow no SMC calls through, which would mean the default configuration would essentially be the same as the second configuration that is defined in the motivation section. However, I am wondering if there is a case where you still just want to disable it completely? # Should the microkernel have an overriding blacklist? ## How and who decides what the blacklist is? ## Is it configurable? ## I don't think this question blocks the RFC as the blacklist could be implemented easily enough after the proposed implementation. ( https://sel4.atlassian.net/browse/RFC-9#add-comment?atlOrigin=eyJpIjoiOTY3Z… ) Add Comment ( https://sel4.atlassian.net/browse/RFC-9#add-comment?atlOrigin=eyJpIjoiOTY3Z… ) Get Jira notifications on your phone! Download the Jira Cloud app for Android ( https://play.google.com/store/apps/details?id=com.atlassian.android.jira.co… ) or iOS ( https://itunes.apple.com/app/apple-store/id1006972087?pt=696495&ct=EmailNot… ) This message was sent by Atlassian Jira (v1001.0.0-SNAPSHOT#100211- sha1:6a2b617 )

1 0

[JIRA] (RFC-9) Add new capability for seL4 SMC Forwarding
by Robbie VanVossen (Jira) 02 Dec '22

02 Dec '22

Robbie VanVossen ( https://sel4.atlassian.net/secure/ViewProfile.jspa?accountId=557058%3Ab7eb5… ) *updated* an issue RFCs ( https://sel4.atlassian.net/browse/RFC?atlOrigin=eyJpIjoiZjk5MTU0ODY3ODdhNDJ… ) / RFC ( https://sel4.atlassian.net/browse/RFC-9?atlOrigin=eyJpIjoiZjk5MTU0ODY3ODdhN… ) RFC-9 ( https://sel4.atlassian.net/browse/RFC-9?atlOrigin=eyJpIjoiZjk5MTU0ODY3ODdhN… ) Add new capability for seL4 SMC Forwarding ( https://sel4.atlassian.net/browse/RFC-9?atlOrigin=eyJpIjoiZjk5MTU0ODY3ODdhN… ) Change By: Robbie VanVossen ( https://sel4.atlassian.net/secure/ViewProfile.jspa?accountId=557058%3Ab7eb5… ) h1. Summary This RFC proposes to add a new capability which certain threads can invoke so seL4 can make SMC calls on ARM platforms in EL2 (virtualized mode) on the thread’s behalf. h1. Motivation There are currently two ways to handle SMC calls when using Currently, seL4 as a hypervisor on ARM. # You can configure the system so that any thread running at EL1/0 can directly make in EL2 traps all SMC calls. (Configure by enabling {{KernelAllowSMCCalls}} for the microkernel) # You can configure the system so that SMC calls are trapped into Then seL4. This allows seL4 to decide decides what to do with the these calls. In the current implementation Both approaches leave something to be desired. The first option has various security implications and is essentially unprovable as it would allow any thread to make secure calls that could put the system in a bad state. The second option is more secure and allows you to emulate some Some SMC calls are emulated , such as PSCI, however but most are just ignored. However , there may be some SMC calls that do actually need to be made on behalf of a thread make it to the Supervisor , and this configuration does not allow that to happen. We want This RFC proposes a solution to essentially combine the two approaches this problem by configuring the system for the second approach, and then allowing giving threads that have a new badged, capability which allows them to forward SMC calls to seL4 so that seL4 the microkernel will make the actual SMC call on the thread's behalf. h1. Guide-level explanation This change will allow you to forward to the microkernel, emulate, or block SMC calls based on platform specific configuration. When an SMC call is made by a VM, the microkernel intercepts the call , which it does for all SMC calls made at EL1. It then calls {{handle_smc()}} for the corresponding VMM thread. That function will decode the specific SMC call and decide whether to block it, emulate it, or forward the call back into seL4. There are examples already for multiple platforms on emulation and blocking. The zynqmp platform will have an example of forwarding the SMC with the new API call, {{ seL4_ARM_SMC_Forward seL4_ARM_SMC_Call ()}}. h2. New concepts h3. seL4_ARM_SMC_Forward seL4_ARM_SMC_Call () {{ seL4_ARM_SMC_Forward seL4_ARM_SMC_Call ()}} is the new ObjectInvocation API call and can be called in two ways: # A VM attempts to make an SMC call which traps into the microkernel. The microkernel informs the VMM’s handler and the VMM decides that the SMC call really needs to be made. The VMM then calls {{ seL4_ARM_SMC_Forward seL4_ARM_SMC_Call ()}} with the SMC capability and all the arguments that the VM provided. # Directly by constructing the SMC call from the thread with the SMC capability. This means that a trap isn’t needed. Our expected use case has been using the first approach in the camkes-vm for VMM handlers on the behalf of its VM. I am not sure if there is a use case for the second approach, but it is certainly possible. In either case, the function still requires the caller to set all 8 registers, even if the specific SMC call uses less arguments. h3. seL4_ARM_SMC {{seL4_ARM_SMC}} is the new capability. It needs to be given to any threads that you expect to call {{ seL4_ARM_SMC_Forward seL4_ARM_SMC_Call ()}}, otherwise it will be denied. In We expect that in most use cases, it will be given to each VMM thread. The capability is badged and the badge represents the SMC Function Identifier. Therefore, the specific SMC calls that each Thread can make is configurable and checked by seL4. h2. For users We plan to implement this configuration as the default for users of camkes-vm for ARM platforms the zynqmp platform. However Due to WCET implication of this approach , it should be quite seamless. We will implement a not be the default configuration for each platform that denies forwarding on any SMC calls so that functionality will remain the same all platforms. The exception will It should only be enabled when actually needed for the zynqmp platform, which up until now has not been a supported specific platform for the camkes-vm or application. This was actually due to the fact that Without this platform required {{KernelAllowSMCCalls}} to be enabled to get change, Linux working is unable to run as a VM. This change will fix that issue so that guest on camkes-vm for the zynqmp will be a valid platform. This platform will have a different default, which allows specific SMC calls to be forwarded on so that Linux guests will boot correctly. If a user enables SMC forwarding and needs to update the configuration for their platform, they can make changes to {{projects/seL4_projects_libs/libsel4vmmplatsupport/src/plat/$KernelPlatform/devices/smc.c}}. h1. Reference-level explanation This implementation will meet the SMC Calling Convention specification from ARM ([ARM DEN 0028E v1.4|https://developer.arm.com/documentation/den0028/latest?_ga=2.116565828…] for 32-bit SMC calls. 64-bit SMC calls are partially supported, but full support could be added later. h2. seL4_ARM_SMC_Forward seL4_ARM_SMC_Call () {code:c} static int seL4_ARM_SMC_Forward seL4_ARM_SMC_Call ( seL4_ARM_SMC _service, seL4_ARM_SMCContext * smc_args, seL4_ARM_SMCContext * smc_response ) {code} {noformat} Parameters [in] _service Capability to allow threads to interact with SMC calls. [in] smc_args The structure that has the provided arguments. [out] smc_response The structure to capture the responses. Returns {noformat} h2. struct seL4_ARM_SMCContext {code:c} typedef struct seL4_ARM_SMCContext_ { /* register arguments */ seL4_Word x0, x1, x2, x3, x4, x5, x6, x7; } seL4_ARM_SMCContext; {code} h1. Drawbacks The main drawback of this approach is that you are allowing the user to make the decision on which SMC calls should go through or not. This will require a fairly advanced configuration and allows supervisor in the user to open themselves up to system -level vulnerabilities past initialization. However, I think This means that the need worst case execution time is no longer guaranteed for the feature outweighs the risk system. Also, to help mitigate the risk in the camkes-vm, I think we Each SMC call can setup take an arbitrarily long amount of time. This means a default configuration that allows as few SMC calls as possible thread can essentially lock up all of the cores for most platforms an unspecified amount of time. We have done some testing with Linux The core the SMC call is made on is obviously blocked because the zynqmp platform supervisor will not return to seL4 until it completes its function. The other cores will become locked as soon as seL4 runs and have identified tries to grab the minimum SMC calls we need Big Kernel Lock (BKL). One potential fix for the VMM's multicore issue is to support just to boot release the VMs BKL before making the SMC call. So However, that platform configuration will be more permissive but will meet is out of the needs scope of this RFC as it has a Linux guest VM lot of implications. One other way The current approach will be to mitigate have this issue is identifying some SMC calls that should never be made config option off by a thread default and add documentation about how enabling it affects WCET. Then create a platform-dependent blacklist in the microkernel that would disallow certain SMC calls even from threads that have the SMC capability. h1. Rationale and alternatives {anchor:rationale} This design is the best because it gives users flexibility by allowing them to make some SMC calls while still being able to emulate other SMC calls. It is the preferred method, because the microkernel shouldn't need to be updated with each new platform, VMM configuration, and secure application implementation. It can all be configured in user space. h2. Alternative approaches SMC filtering could be done in the microkernel so that developers are less likely to configure the system incorrectly. However, there are some cases where the filtering is dependent on specific threads and the application level (Secure world communication) instead of just platform dependent. On the first point, instead of creating a cap for making the SMC call, you could have a cap for each numbered SMC call so that you could configure what SMC call each thread can make using the root thread. However, that is quite a bit of work as that will differ for each ARM platform and adds another step when porting the microkernel to new ARM platforms. That could also solve the second point, but then users would likely need to modify the microkernel to support their specific secure world application, which is something that should really be avoided. Another part of the implementation that could be done differently is using the VCPU capability with an access right instead of creating a new type of capability. The kinds of access rights are already determined as read, write, grant, and grantreply. A write access right would probably make the most sense, but it still seems obtuse that a write access on a VCPU capability would allow the invoker to make an SMC call. Besides that, there are two other issues I see with this approach: # I can foresee a use-case where you have a health monitor component that isn’t associated with a VCPU but needs to be able to forward an SMC call to actually interact with EL3 or secure applications. # If the VCPU capability ever needs read/write access rights that make more sense with the type of object it is, for example, limiting which threads can actually read or write that VCPUs registers. For those reasons we believe a new capability makes the most sense. h2. Impact of not implementing this Not implementing this will limit the ARM platforms we can use for the camkes-vm and/or the guests that can run on those platforms. It will also limit support for applications that run in secure side of the trustzone that need to communicate with the nonsecure side (Where sel4, VMMs, and the VMs run). h1. Prior art Xen has a similar implementation for the Zynqmp where arbitration is done to allow/disallow/emulate SMC calls from guest VMs. The arbitration is quite specific and is automatically generated based on resource allocation for each VM. Xen does not have VMM pieces that run in Userspace. Basically all of the management runs in EL2. Therefore, all of the arbitration is done in Xen itself. This approach does not make sense for seL4 for the reasons stated above in the [#rationale |#rationale ] section. However, this does show a need for this feature in hypervisors on ARMv8 platforms. If necessary, we could reference the Xen arbitration code to help improve the default arbitration for the zynqmp in our implementation. h1. Unresolved questions # Should this implementation be configurable? ## I would argue it makes the most sense to have it automatically enabled if you are using the VMM on ARM and {{KernelAllowSMCCalls}} is disabled. Then your default whitelist in the SMC handler, for most platforms, can be to allow no SMC calls through, which would mean the default configuration would essentially be the same as the second configuration that is defined in the motivation section. However, I am wondering if there is a case where you still just want to disable it completely? # Should the microkernel have an overriding blacklist? ## How and who decides what the blacklist is? ## Is it configurable? ## I don't think this question blocks the RFC as the blacklist could be implemented easily enough after the proposed implementation. ( https://sel4.atlassian.net/browse/RFC-9#add-comment?atlOrigin=eyJpIjoiZjk5M… ) Add Comment ( https://sel4.atlassian.net/browse/RFC-9#add-comment?atlOrigin=eyJpIjoiZjk5M… ) Get Jira notifications on your phone! Download the Jira Cloud app for Android ( https://play.google.com/store/apps/details?id=com.atlassian.android.jira.co… ) or iOS ( https://itunes.apple.com/app/apple-store/id1006972087?pt=696495&ct=EmailNot… ) This message was sent by Atlassian Jira (v1001.0.0-SNAPSHOT#100211- sha1:6a2b617 )

1 0