KVM SGX virtualization support
Subsystems (6)
cover KVM SGX virtualization support
--- Disclaimer ---
These patches were originally written by Sean Christopherson while at Intel. Now that Sean has left Intel, I (Kai) have taken over getting them upstream. This series needs more review before it can be merged. It is being posted publicly and under RFC so Sean and others can review it. Maintainers are safe ignoring it for now.
------------------
Hi all,
This series adds KVM SGX virtualization support. The first 12 patches starting with x86/sgx or x86/cpu.. are necessary changes to x86 and SGX core/driver to support KVM SGX virtualization, while the rest are patches to KVM subsystem.
Please help to review this series. Also I'd like to hear what is the proper way to merge this series, since it contains change to both x86/SGX and KVM subsystem. Any feedback is highly appreciated. And please let me know if I forgot to CC anyone, or anyone wants to be removed from CC. Thanks in advance!
This series is based against latest tip tree's x86/sgx branch. You can also get the code from tip branch of kvm-sgx repo on github:
https://github.com/intel/kvm-sgx.git tip
It also requires Qemu changes to create VM with SGX support. You can find Qemu repo here:
https://github.com/intel/qemu-sgx.git next
Please refer to README.md of above qemu-sgx repo for detail on how to create guest with SGX support. At meantime, for your quick reference you can use below command to create SGX guest:
#qemu-system-x86_64 -smp 4 -m 2G -drive file=<your_vm_image>,if=virtio \ -cpu host,+sgx_provisionkey \ -sgx-epc id=epc1,memdev=mem1 \ -object memory-backend-epc,id=mem1,size=64M,prealloc
Please note that the SGX relevant part is:
-cpu host,+sgx_provisionkey \ -sgx-epc id=epc1,memdev=mem1 \ -object memory-backend-epc,id=mem1,size=64M,prealloc
And you can change other parameters of your qemu command based on your needs.
=========
KVM SGX virtualization Overview
- Virtual EPC
"Virtual EPC" is the EPC section exposed by KVM to guest so SGX software in guest can discover it and use it to create SGX enclaves. KVM exposes SGX to guest via CPUID, and exposes one or more "virtual EPC" sections for guest. The size of "virtual EPC" is passed as Qemu parameter when creating the guest, and the base address is calcualted internally according to guest's configuration.
To support virtual EPC, add a new misc device /dev/sgx_virt_epc to SGX core/driver to allow userspace (Qemu) to allocate "raw" EPC, and use it as "virtual EPC" for guest. Obviously, unlike EPC allocated for host SGX driver, virtual EPC allocated via /dev/sgx_virt_epc doesn't have enclave associated, and how virtual EPC is used by guest is compeletely controlled by guest's SGX software.
Implement the "raw" EPC allocation in the x86 core-SGX subsystem via /dev/sgx_virt_epc rather than in KVM. Doing so has two major advantages:
- Does not require changes to KVM's uAPI, e.g. EPC gets handled as just another memory backend for guests.
- EPC management is wholly contained in the SGX subsystem, e.g. SGX does not have to export any symbols, changes to reclaim flows don't need to be routed through KVM, SGX's dirty laundry doesn't have to get aired out for the world to see, and so on and so forth.
The virtual EPC allocated to guests is currently not reclaimable, due to reclaiming EPC from KVM guests is not currently supported. Due to the complications of handling reclaim conflicts between guest and host, KVM EPC oversubscription, which allows total virtual EPC size greater than physical EPC by being able to reclaiming guests' EPC, is significantly more complex than basic support for SGX virtualization.
- Support SGX virtualization without SGX Launch Control unlocked mode
Although SGX driver requires SGX Launch Control unlocked mode to work, SGX virtualization doesn't, since how enclave is created is completely controlled by guest SGX software, which is not necessarily linux. Therefore, this series allows KVM to expose SGX to guest even SGX Launch Control is in locked mode, or is not present at all. The reason is the goal of SGX virtualization, or virtualization in general, is to expose hardware feature to guest, but not to make assumption how guest will use it. Therefore, KVM should support SGX guest as long as hardware is able to, to have chance to support more potential use cases in cloud environment.
- Support exposing SGX2
Due to the same reason above, SGX2 feature detection is added to core SGX code to allow KVM to expose SGX2 to guest, even currently SGX driver doesn't support SGX2, because SGX2 can work just fine in guest w/o any interaction to host SGX driver.
- Restricit SGX guest access to provisioning key
To grant guest being able to fully use SGX, guest needs to be able to create provisioning enclave. However provisioning key is sensitive and is restricted by /dev/sgx_provision in host SGX driver, therefore KVM SGX virtualization follows the same role: a new KVM_CAP_SGX_ATTRIBUTE is added to KVM uAPI, and only file descriptor of /dev/sgx_provision is passed to that CAP by usersppace hypervisor (Qemu) when creating the guest, it can access provisioning bit. This is done by making KVM trape ECREATE instruction from guest, and check the provisioning bit in ECREATE's attribute.
Kai Huang (1):
x86/sgx: Add helper to update SGX_LEPUBKEYHASHn MSRs
Sean Christopherson (22):
x86/sgx: Split out adding EPC page to free list to separate helper x86/sgx: Add enum for SGX_CHILD_PRESENT error code x86/sgx: Introduce virtual EPC for use by KVM guests x86/cpufeatures: Add SGX1 and SGX2 sub-features x86/cpu/intel: Allow SGX virtualization without Launch Control support x86/sgx: Expose SGX architectural definitions to the kernel x86/sgx: Move ENCLS leaf definitions to sgx_arch.h x86/sgx: Add SGX2 ENCLS leaf definitions (EAUG, EMODPR and EMODT) x86/sgx: Add encls_faulted() helper
x86/sgx: Add helpers to expose ECREATE and EINIT to KVM x86/sgx: Move provisioning device creation out of SGX driver
KVM: VMX: Convert vcpu_vmx.exit_reason to a union
KVM: x86: Export kvm_mmu_gva_to_gpa_{read,write}() for SGX (VMX)
KVM: x86: Define new #PF SGX error code bit
KVM: x86: Add SGX feature leaf to reverse CPUID lookup
KVM: VMX: Add basic handling of VM-Exit from SGX enclave
KVM: VMX: Frame in ENCLS handler for SGX virtualization
KVM: VMX: Add SGX ENCLS[ECREATE] handler to enforce CPUID restrictions
KVM: VMX: Add emulation of SGX Launch Control LE hash MSRs
KVM: VMX: Add ENCLS[EINIT] handler to support SGX Launch Control (LC)
KVM: VMX: Enable SGX virtualization for SGX1, SGX2 and LC
KVM: x86: Add capability to grant VM access to privileged SGX
attribute
Documentation/virt/kvm/api.rst | 23 +
arch/x86/Kconfig | 12 +
arch/x86/include/asm/cpufeature.h | 5 +-
arch/x86/include/asm/cpufeatures.h | 6 +-
arch/x86/include/asm/disabled-features.h | 7 +-
arch/x86/include/asm/kvm_host.h | 5 +
arch/x86/include/asm/required-features.h | 2 +-
arch/x86/include/asm/sgx.h | 19 +
.../cpu/sgx/arch.h => include/asm/sgx_arch.h} | 20 +
arch/x86/include/asm/vmx.h | 1 +
arch/x86/include/uapi/asm/vmx.h | 1 +
arch/x86/kernel/cpu/common.c | 4 +
arch/x86/kernel/cpu/feat_ctl.c | 50 +-
arch/x86/kernel/cpu/sgx/Makefile | 1 +
arch/x86/kernel/cpu/sgx/driver.c | 17 -
arch/x86/kernel/cpu/sgx/encl.c | 2 +-
arch/x86/kernel/cpu/sgx/encls.h | 29 +-
arch/x86/kernel/cpu/sgx/ioctl.c | 23 +-
arch/x86/kernel/cpu/sgx/main.c | 79 ++-
arch/x86/kernel/cpu/sgx/sgx.h | 5 +-
arch/x86/kernel/cpu/sgx/virt.c | 318 ++++++++++++
arch/x86/kernel/cpu/sgx/virt.h | 14 +
arch/x86/kvm/Makefile | 2 +
arch/x86/kvm/cpuid.c | 58 ++-
arch/x86/kvm/cpuid.h | 1 +
arch/x86/kvm/vmx/nested.c | 70 ++-
arch/x86/kvm/vmx/nested.h | 5 +
arch/x86/kvm/vmx/sgx.c | 462 ++++++++++++++++++
arch/x86/kvm/vmx/sgx.h | 34 ++
arch/x86/kvm/vmx/vmcs12.c | 1 +
arch/x86/kvm/vmx/vmcs12.h | 4 +-
arch/x86/kvm/vmx/vmx.c | 171 +++++--
arch/x86/kvm/vmx/vmx.h | 27 +-
arch/x86/kvm/x86.c | 24 +
include/uapi/linux/kvm.h | 1 +
tools/testing/selftests/sgx/defines.h | 2 +-
36 files changed, 1366 insertions(+), 139 deletions(-)
create mode 100644 arch/x86/include/asm/sgx.h
rename arch/x86/{kernel/cpu/sgx/arch.h => include/asm/sgx_arch.h} (96%)
create mode 100644 arch/x86/kernel/cpu/sgx/virt.c
create mode 100644 arch/x86/kernel/cpu/sgx/virt.h
create mode 100644 arch/x86/kvm/vmx/sgx.c
create mode 100644 arch/x86/kvm/vmx/sgx.hSignature · 1 line
2.29.2
RFC PATCH 1/23 x86/sgx: Split out adding EPC page to free list to separate helper
From: Sean Christopherson <[email protected]>
SGX virtualization requires to allocate "raw" EPC and use it as virtual EPC for SGX guest. Unlike EPC used by SGX driver, virtual EPC doesn't track how EPC pages are used in VM, e.g. (de)construction of enclaves, so it cannot guarantee EREMOVE success, e.g. it doesn't have a priori knowledge of which pages are SECS with non-zero child counts.
Split sgx_free_page() into two parts so that the "add to free list" part can be used by virtual EPC without having to modify the EREMOVE logic in sgx_free_page().
--- arch/x86/kernel/cpu/sgx/main.c | 24 ++++++++++++++++++------ arch/x86/kernel/cpu/sgx/sgx.h | 1 + 2 files changed, 19 insertions(+), 6 deletions(-)
arch/x86/kernel/cpu/sgx/main.c +18 -6
@@ -594,15 +594,30 @@ struct sgx_epc_page *sgx_alloc_epc_page(void *owner, bool reclaim)return page; } +/** +* __sgx_free_epc_page() - Free an EPC page +* @page: pointer to a previously allocated EPC page +* +* Insert an EPC page back to the list of free pages. +*/ +void __sgx_free_epc_page (struct sgx_epc_page *page) +{ +struct sgx_epc_section *section = &sgx_epc_sections[page->section]; + +spin_lock (§ion->lock); +list_add_tail (&page->list, §ion->page_list); + section->free_cnt++; +spin_unlock (§ion->lock); +} +/** * sgx_free_epc_page() - Free an EPC page -* @page: an EPC page +* @page: pointer to a previously allocated EPC page * * Call EREMOVE for an EPC page and insert it back to the list of free pages. */ void sgx_free_epc_page (struct sgx_epc_page *page) { -struct sgx_epc_section *section = &sgx_epc_sections[page->section];int ret;WARN_ON_ONCE (page->flags &SGX_EPC_PAGE_RECLAIMER_TRACKED );
@@ -611,10 +626,7 @@ void sgx_free_epc_page(struct sgx_epc_page *page)if (WARN_ONCE (ret,"EREMOVE returned %d (0x%x)" , ret, ret))return ; -spin_lock (§ion->lock); -list_add_tail (&page->list, §ion->page_list); - section->free_cnt++; -spin_unlock (§ion->lock); +__sgx_free_epc_page (page); } static bool __init sgx_setup_epc_section(u64 phys_addr, u64 size,
arch/x86/kernel/cpu/sgx/sgx.h +1 -0
@@ -77,6 +77,7 @@ static inline void *sgx_get_epc_virt_addr(struct sgx_epc_page *page) }struct sgx_epc_page *__sgx_alloc_epc_page (void ); +void __sgx_free_epc_page (struct sgx_epc_page *page);void sgx_free_epc_page (struct sgx_epc_page *page);void sgx_mark_page_reclaimable (struct sgx_epc_page *page);
Signature · 1 line
2.29.2
RFC PATCH 2/23 x86/sgx: Add enum for SGX_CHILD_PRESENT error code
In reply to Kai HuangFrom: Sean Christopherson <[email protected]>
SGX virtualization requires to allocate "raw" EPC and use it as "virtual EPC" for SGX guest. Unlike EPC used by SGX driver, virtual EPC doesn't track how EPC pages are used in VM, e.g. (de)construction of enclaves, so it cannot guarantee EREMOVE success, e.g. it doesn't have a priori knowledge of which pages are SECS with non-zero child counts.
Add SGX_CHILD_PRESENT for use by SGX virtualization to assert EREMOVE failures are expected, but only due to SGX_CHILD_PRESENT.
--- arch/x86/kernel/cpu/sgx/arch.h | 2 ++ 1 file changed, 2 insertions(+)
arch/x86/kernel/cpu/sgx/arch.h +2 -0
@@ -26,12 +26,14 @@* enum sgx_return_code - The return code type for ENCLS, ENCLU and ENCLV * %SGX_NOT_TRACKED: Previous ETRACK's shootdown sequence has not * been completed yet. +* %SGX_CHILD_PRESENT Enclave has child pages present in the EPC. * %SGX_INVALID_EINITTOKEN: EINITTOKEN is invalid and enclave signer's * public key does not match IA32_SGXLEPUBKEYHASH. * %SGX_UNMASKED_EVENT: An unmasked event, e.g. INTR, was received */ enum sgx_return_code {SGX_NOT_TRACKED =11 , +SGX_CHILD_PRESENT =13 ,SGX_INVALID_EINITTOKEN =16 ,SGX_UNMASKED_EVENT =128 , };
Signature · 1 line
2.29.2
RFC PATCH 3/23 x86/sgx: Introduce virtual EPC for use by KVM guests
In reply to Kai HuangFrom: Sean Christopherson <[email protected]>
Add a misc device /dev/sgx_virt_epc to allow userspace to allocate "raw" EPC without an associated enclave. The intended and only known use case for raw EPC allocation is to expose EPC to a KVM guest, hence the virt_epc moniker, virt.{c,h} files and X86_SGX_VIRTUALIZATION Kconfig.
Modify sgx_init() to always try to initialize virtual EPC driver, even when SGX driver is disabled due to SGX Launch Control is in locked mode, or not present at all, since SGX virtualization allows to expose SGX to guests that support non-LC configurations.
Implement the "raw" EPC allocation in the x86 core-SGX subsystem via /dev/sgx_virt_epc rather than in KVM. Doing so has two major advantages:
- Does not require changes to KVM's uAPI, e.g. EPC gets handled as just another memory backend for guests.
- EPC management is wholly contained in the SGX subsystem, e.g. SGX does not have to export any symbols, changes to reclaim flows don't need to be routed through KVM, SGX's dirty laundry doesn't have to get aired out for the world to see, and so on and so forth.
The virtual EPC allocated to guests is currently not reclaimable, due to oversubscription of EPC for KVM guests is not currently supported. Due to the complications of handling reclaim conflicts between guest and host, KVM EPC oversubscription is significantly more complex than basic support for SGX virtualization.
--- arch/x86/Kconfig | 12 ++ arch/x86/kernel/cpu/sgx/Makefile | 1 + arch/x86/kernel/cpu/sgx/main.c | 5 +- arch/x86/kernel/cpu/sgx/virt.c | 263 +++++++++++++++++++++++++++++++ arch/x86/kernel/cpu/sgx/virt.h | 14 ++ 5 files changed, 294 insertions(+), 1 deletion(-) create mode 100644 arch/x86/kernel/cpu/sgx/virt.c create mode 100644 arch/x86/kernel/cpu/sgx/virt.h
arch/x86/Kconfig +12 -0
@@ -1947,6 +1947,18 @@ config X86_SGX If unsure, say N. +config X86_SGX_VIRTUALIZATION + bool "Software Guard eXtensions (SGX) Virtualization" + depends on X86_SGX && KVM_INTEL + help + + Enables KVM guests to create SGX enclaves. + + This includes support to expose "raw" unreclaimable enclave memory to + guests via a device node, e.g. /dev/sgx_virt_epc. + + If unsure, say N. + config EFI bool "EFI runtime service support" depends on ACPI
arch/x86/kernel/cpu/sgx/Makefile +1 -0
@@ -3,3 +3,4 @@ obj-y += \ encl.o \ ioctl.o \ main.o +obj-$(CONFIG_X86_SGX_VIRTUALIZATION) += virt.o
arch/x86/kernel/cpu/sgx/main.c +4 -1
@@ -9,9 +9,11 @@#include <linux/sched/mm.h> #include <linux/sched/signal.h> #include <linux/slab.h> +#include "arch.h" #include "driver.h" #include "encl.h" #include "encls.h" +#include "virt.h" struct sgx_epc_section sgx_epc_sections[SGX_MAX_EPC_SECTIONS ];static int sgx_nr_epc_sections;
@@ -726,7 +728,8 @@ static void __init sgx_init(void)if (!sgx_page_reclaimer_init ()) goto err_page_cache; - ret =sgx_drv_init (); +/* Success if the native *or* virtual EPC driver initialized cleanly. */ + ret = !!sgx_drv_init () & !!sgx_virt_epc_init ();if (ret) goto err_kthread;
arch/x86/kernel/cpu/sgx/virt.c new file +263 -0
263 lines not on this page · raw message
arch/x86/kernel/cpu/sgx/virt.h new file +14 -0
@@ -0,0 +1,14 @@ +/* SPDX-License-Identifier: (GPL-2.0 OR BSD-3-Clause) */ +#ifndef _ASM_X86_SGX_VIRT_H +#define _ASM_X86_SGX_VIRT_H + +#ifdef CONFIG_X86_SGX_VIRTUALIZATION +int __init sgx_virt_epc_init (void ); +#else +static inline int __init sgx_virt_epc_init (void ) +{ +return -ENODEV ; +} +#endif + +#endif /* _ASM_X86_SGX_VIRT_H */
Signature · 1 line
2.29.2
RFC PATCH 4/23 x86/cpufeatures: Add SGX1 and SGX2 sub-features
In reply to Kai HuangFrom: Sean Christopherson <[email protected]>
Add a feature word to hold SGX features enumerated via CPUID.0x12.0x0, along with flags for SGX1 and SGX2. As part of virtualizing SGX, KVM needs to expose the SGX CPUID leafs to its guest. SGX1 and SGX2 need to be in a dedicated feature word so that they can be queried via KVM's reverse CPUID lookup to properly emulate the expected guest behavior.
--- arch/x86/include/asm/cpufeature.h | 5 +++-- arch/x86/include/asm/cpufeatures.h | 6 +++++- arch/x86/include/asm/disabled-features.h | 7 ++++++- arch/x86/include/asm/required-features.h | 2 +- arch/x86/kernel/cpu/common.c | 4 ++++ arch/x86/kernel/cpu/feat_ctl.c | 2 ++ 6 files changed, 21 insertions(+), 5 deletions(-)
arch/x86/include/asm/cpufeature.h +3 -2
@@ -30,6 +30,7 @@ enum cpuid_leafsCPUID_7_ECX ,CPUID_8000_0007_EBX ,CPUID_7_EDX , + CPUID_12_EAX, };#ifdef CONFIG_X86_FEATURE_NAMES
@@ -89,7 +90,7 @@ extern const char * const x86_bug_flags[NBUGINTS*32]; CHECK_BIT_IN_MASK_WORD(REQUIRED_MASK, 17, feature_bit) || \ CHECK_BIT_IN_MASK_WORD(REQUIRED_MASK, 18, feature_bit) || \ REQUIRED_MASK_CHECK || \ - BUILD_BUG_ON_ZERO(NCAPINTS != 19)) + BUILD_BUG_ON_ZERO(NCAPINTS != 20))#define DISABLED_MASK_BIT_SET (feature_bit) \ ( CHECK_BIT_IN_MASK_WORD(DISABLED_MASK, 0, feature_bit) || \
@@ -112,7 +113,7 @@ extern const char * const x86_bug_flags[NBUGINTS*32]; CHECK_BIT_IN_MASK_WORD(DISABLED_MASK, 17, feature_bit) || \ CHECK_BIT_IN_MASK_WORD(DISABLED_MASK, 18, feature_bit) || \ DISABLED_MASK_CHECK || \ - BUILD_BUG_ON_ZERO(NCAPINTS != 19)) + BUILD_BUG_ON_ZERO(NCAPINTS != 20))#define cpu_has (c, bit) \ (__builtin_constant_p(bit) && REQUIRED_MASK_BIT_SET(bit) ? 1 : \
arch/x86/include/asm/cpufeatures.h +5 -1
@@ -13,7 +13,7 @@/* * Defines x86 CPU feature bits */ -#define NCAPINTS 19/* N 32-bit words worth of info */ +#define NCAPINTS 20/* N 32-bit words worth of info */ #define NBUGINTS 1/* N 32-bit bug flags */ /*
@@ -383,6 +383,10 @@#define X86_FEATURE_CORE_CAPABILITIES (18*32+30)/* "" IA32_CORE_CAPABILITIES MSR */ #define X86_FEATURE_SPEC_CTRL_SSBD (18*32+31)/* "" Speculative Store Bypass Disable */ +/* Intel-defined SGX features, CPUID level 0x00000012:0 (EAX), word 19 */ +#define X86_FEATURE_SGX1 (19*32+ 0)/* SGX1 leaf functions */ +#define X86_FEATURE_SGX2 (19*32+ 1)/* SGX2 leaf functions */ +/* * BUG word(s) */
arch/x86/include/asm/disabled-features.h +6 -1
@@ -28,12 +28,16 @@# define DISABLE_CYRIX_ARR (1<<(X86_FEATURE_CYRIX_ARR & 31))# define DISABLE_CENTAUR_MCR (1<<(X86_FEATURE_CENTAUR_MCR & 31))# define DISABLE_PCID 0 +# define DISABLE_SGX1 0 +# define DISABLE_SGX2 0#else # define DISABLE_VME 0# define DISABLE_K6_MTRR 0# define DISABLE_CYRIX_ARR 0# define DISABLE_CENTAUR_MCR 0# define DISABLE_PCID (1<<(X86_FEATURE_PCID & 31)) +# define DISABLE_SGX1 (1<<(X86_FEATURE_SGX1 & 31)) +# define DISABLE_SGX2 (1<<(X86_FEATURE_SGX2 & 31))#endif /* CONFIG_X86_64 */ #ifdef CONFIG_X86_INTEL_MEMORY_PROTECTION_KEYS
@@ -91,6 +95,7 @@ DISABLE_ENQCMD)#define DISABLED_MASK17 0#define DISABLED_MASK18 0 -#define DISABLED_MASK_CHECK BUILD_BUG_ON_ZERO(NCAPINTS != 19) +#define DISABLED_MASK19 (DISABLE_SGX1|DISABLE_SGX2) +#define DISABLED_MASK_CHECK BUILD_BUG_ON_ZERO(NCAPINTS != 20)#endif /* _ASM_X86_DISABLED_FEATURES_H */
arch/x86/include/asm/required-features.h +1 -1
@@ -101,6 +101,6 @@#define REQUIRED_MASK16 0#define REQUIRED_MASK17 0#define REQUIRED_MASK18 0 -#define REQUIRED_MASK_CHECK BUILD_BUG_ON_ZERO(NCAPINTS != 19) +#define REQUIRED_MASK_CHECK BUILD_BUG_ON_ZERO(NCAPINTS != 20)#endif /* _ASM_X86_REQUIRED_FEATURES_H */
arch/x86/kernel/cpu/common.c +4 -0
@@ -932,6 +932,10 @@ void get_cpu_cap(struct cpuinfo_x86 *c) c->x86_capability[CPUID_D_1_EAX ] = eax; } +/* Additional Intel-defined SGX flags: level 0x00000012 */ +if (c->cpuid_level >=0x00000012 ) + c->x86_capability[CPUID_12_EAX ] =cpuid_eax (0x00000012 ); +/* AMD-defined flags: level 0x80000001 */ eax =cpuid_eax (0x80000000 ); c->extended_cpuid_level = eax;
arch/x86/kernel/cpu/feat_ctl.c +2 -0
@@ -97,6 +97,8 @@ static void clear_sgx_caps(void) {setup_clear_cpu_cap (X86_FEATURE_SGX );setup_clear_cpu_cap (X86_FEATURE_SGX_LC ); +setup_clear_cpu_cap (X86_FEATURE_SGX1 ); +setup_clear_cpu_cap (X86_FEATURE_SGX2 ); }static int __init nosgx (char *str)
Signature · 1 line
2.29.2
RFC PATCH 5/23 x86/cpu/intel: Allow SGX virtualization without Launch Control support
In reply to Kai HuangFrom: Sean Christopherson <[email protected]>
Allow SGX virtualization on systems without Launch Control support, i.e. allow KVM to expose SGX to guests that support non-LC configurations.
Introduce clear_sgx_lc() to clear SGX_LC feature bit only if SGX Launch Control is locked by BIOS when SGX virtualization is enabled, to prevent SGX driver being enabled.
Improve error message to distinguish three cases: 1) SGX disabled completely by BIOS; 2) SGX disabled completely due to SGX LC is locked by BIOS, and SGX virtualization is also disabled; 3) Only SGX driver is disabled due to SGX LC is locked by BIOS, but SGX virtualization is enabled.
--- arch/x86/kernel/cpu/feat_ctl.c | 48 +++++++++++++++++++++++++--------- 1 file changed, 36 insertions(+), 12 deletions(-)
arch/x86/kernel/cpu/feat_ctl.c +36 -12
@@ -101,6 +101,11 @@ static void clear_sgx_caps(void)setup_clear_cpu_cap (X86_FEATURE_SGX2 ); } +static void clear_sgx_lc (void ) +{ +setup_clear_cpu_cap (X86_FEATURE_SGX_LC ); +} +static int __init nosgx (char *str) {clear_sgx_caps ();
@@ -113,7 +118,7 @@ early_param("nosgx", nosgx);void init_ia32_feat_ctl (struct cpuinfo_x86 *c) {bool tboot =tboot_enabled (); -bool enable_sgx; +bool enable_sgx_virt, enable_sgx_driver;u64 msr;if (rdmsrl_safe (MSR_IA32_FEAT_CTL , &msr)) {
@@ -123,12 +128,19 @@ void init_ia32_feat_ctl(struct cpuinfo_x86 *c) }/* -* Enable SGX if and only if the kernel supports SGX and Launch Control -* is supported, i.e. disable SGX if the LE hash MSRs can't be written. +* Enable SGX if and only if the kernel supports SGX. Require Launch +* Control support if SGX virtualization is *not* supported, i.e. +* disable SGX if the LE hash MSRs can't be written and SGX can't be +* exposed to a KVM guest (which might support non-LC configurations). */ - enable_sgx =cpu_has (c,X86_FEATURE_SGX ) && -cpu_has (c,X86_FEATURE_SGX_LC ) && -IS_ENABLED (CONFIG_X86_SGX ); + enable_sgx_driver =cpu_has (c,X86_FEATURE_SGX ) && +cpu_has (c,X86_FEATURE_SGX1 ) && +IS_ENABLED (CONFIG_X86_SGX ) && +cpu_has (c,X86_FEATURE_SGX_LC ); + enable_sgx_virt =cpu_has (c,X86_FEATURE_SGX ) && +cpu_has (c,X86_FEATURE_SGX1 ) && +IS_ENABLED (CONFIG_X86_SGX ) && +IS_ENABLED (CONFIG_X86_SGX_VIRTUALIZATION );if (msr &FEAT_CTL_LOCKED ) goto update_caps;
@@ -151,8 +163,11 @@ void init_ia32_feat_ctl(struct cpuinfo_x86 *c) msr |=FEAT_CTL_VMX_ENABLED_INSIDE_SMX ; } -if (enable_sgx) - msr |=FEAT_CTL_SGX_ENABLED |FEAT_CTL_SGX_LC_ENABLED ; +if (enable_sgx_driver || enable_sgx_virt) { + msr |=FEAT_CTL_SGX_ENABLED ; +if (enable_sgx_driver) + msr |=FEAT_CTL_SGX_LC_ENABLED ; + }wrmsrl (MSR_IA32_FEAT_CTL , msr);
@@ -175,10 +190,19 @@ void init_ia32_feat_ctl(struct cpuinfo_x86 *c) } update_sgx: -if (!(msr &FEAT_CTL_SGX_ENABLED ) || - !(msr &FEAT_CTL_SGX_LC_ENABLED ) || !enable_sgx) { -if (enable_sgx) -pr_err_once ("SGX disabled by BIOS\n" ); +if (!(msr &FEAT_CTL_SGX_ENABLED )) { +if (enable_sgx_driver || enable_sgx_virt) +pr_err_once ("SGX disabled by BIOS.\n" );clear_sgx_caps (); } +if (!(msr &FEAT_CTL_SGX_LC_ENABLED ) && + (enable_sgx_driver || enable_sgx_virt)) { +if (!enable_sgx_virt) { +pr_err_once ("SGX Launch Control is locked. Disable SGX.\n" ); +clear_sgx_caps (); + }else if (enable_sgx_driver) { +pr_err_once ("SGX Launch Control is locked. Disable SGX driver.\n" ); +clear_sgx_lc (); + } + } }
Signature · 1 line
2.29.2
RFC PATCH 6/23 x86/sgx: Expose SGX architectural definitions to the kernel
In reply to Kai HuangFrom: Sean Christopherson <[email protected]>
KVM will use many of the architectural constants and structs to virtualize SGX.
---
arch/x86/{kernel/cpu/sgx/arch.h => include/asm/sgx_arch.h} | 0
arch/x86/kernel/cpu/sgx/encl.c | 2 +-
arch/x86/kernel/cpu/sgx/main.c | 2 +-
arch/x86/kernel/cpu/sgx/sgx.h | 2 +-
tools/testing/selftests/sgx/defines.h | 2 +-
5 files changed, 4 insertions(+), 4 deletions(-)
rename arch/x86/{kernel/cpu/sgx/arch.h => include/asm/sgx_arch.h} (100%)arch/x86/include/asm/sgx_arch.h renamed from arch/x86/kernel/cpu/sgx/arch.h +0 -0
arch/x86/kernel/cpu/sgx/encl.c +1 -1
@@ -7,7 +7,7 @@#include <linux/shmem_fs.h> #include <linux/suspend.h> #include <linux/sched/mm.h> -#include "arch.h" +#include <asm/sgx_arch.h> #include "encl.h" #include "encls.h" #include "sgx.h"
arch/x86/kernel/cpu/sgx/main.c +1 -1
@@ -9,7 +9,7 @@#include <linux/sched/mm.h> #include <linux/sched/signal.h> #include <linux/slab.h> -#include "arch.h" +#include <asm/sgx_arch.h> #include "driver.h" #include "encl.h" #include "encls.h"
arch/x86/kernel/cpu/sgx/sgx.h +1 -1
@@ -8,7 +8,7 @@#include <linux/rwsem.h> #include <linux/types.h> #include <asm/asm.h> -#include "arch.h" +#include <asm/sgx_arch.h> #undef pr_fmt#define pr_fmt (fmt) "sgx: " fmt
tools/testing/selftests/sgx/defines.h +1 -1
@@ -14,7 +14,7 @@#define __aligned (x) __attribute__((__aligned__(x)))#define __packed __attribute__((packed)) -#include "../../../../arch/x86/kernel/cpu/sgx/arch.h" +#include "../../../../arch/x86/include/asm/sgx_arch.h" #include "../../../../arch/x86/include/asm/enclu.h" #include "../../../../arch/x86/include/uapi/asm/sgx.h"
Signature · 1 line
2.29.2
RFC PATCH 7/23 x86/sgx: Move ENCLS leaf definitions to sgx_arch.h
In reply to Kai HuangFrom: Sean Christopherson <[email protected]>
Move the ENCLS leaf definitions to sgx_arch.h so that they can be used by KVM. And because they're architectural.
--- arch/x86/include/asm/sgx_arch.h | 15 +++++++++++++++ arch/x86/kernel/cpu/sgx/encls.h | 15 --------------- 2 files changed, 15 insertions(+), 15 deletions(-)
arch/x86/include/asm/sgx_arch.h +15 -0
@@ -22,6 +22,21 @@/* The bitmask for the EPC section type. */ #define SGX_CPUID_EPC_MASK GENMASK(3, 0) +enum sgx_encls_function { +ECREATE =0x00 , +EADD =0x01 , +EINIT =0x02 , +EREMOVE =0x03 , +EDGBRD =0x04 , +EDGBWR =0x05 , +EEXTEND =0x06 , +ELDU =0x08 , +EBLOCK =0x09 , +EPA =0x0A , +EWB =0x0B , +ETRACK =0x0C , +}; +/** * enum sgx_return_code - The return code type for ENCLS, ENCLU and ENCLV * %SGX_NOT_TRACKED: Previous ETRACK's shootdown sequence has not
arch/x86/kernel/cpu/sgx/encls.h +0 -15
@@ -11,21 +11,6 @@#include <asm/traps.h> #include "sgx.h" -enum sgx_encls_function { -ECREATE =0x00 , -EADD =0x01 , -EINIT =0x02 , -EREMOVE =0x03 , -EDGBRD =0x04 , -EDGBWR =0x05 , -EEXTEND =0x06 , -ELDU =0x08 , -EBLOCK =0x09 , -EPA =0x0A , -EWB =0x0B , -ETRACK =0x0C , -}; -/** * ENCLS_FAULT_FLAG - flag signifying an ENCLS return code is a trapnr *
Signature · 1 line
2.29.2
RFC PATCH 8/23 x86/sgx: Add SGX2 ENCLS leaf definitions (EAUG, EMODPR and EMODT)
In reply to Kai HuangFrom: Sean Christopherson <[email protected]>
Define the ENCLS leafs that are available with SGX2, also referred to as Enclave Dynamic Memory Management (EDMM). The leafs will be used by KVM to conditionally expose SGX2 capabilities to guests.
--- arch/x86/include/asm/sgx_arch.h | 3 +++ 1 file changed, 3 insertions(+)
arch/x86/include/asm/sgx_arch.h +3 -0
@@ -35,6 +35,9 @@ enum sgx_encls_function {EPA =0x0A ,EWB =0x0B ,ETRACK =0x0C , +EAUG =0x0D , +EMODPR =0x0E , + EMODT = 0x0F, };/**
Signature · 1 line
2.29.2
RFC PATCH 9/23 x86/sgx: Add encls_faulted() helper
In reply to Kai HuangFrom: Sean Christopherson <[email protected]>
Add a helper to extract the fault indicator from an encoded ENCLS return value. SGX virtualization will also need to detect ENCLS faults.
--- arch/x86/kernel/cpu/sgx/encls.h | 14 +++++++++++++- arch/x86/kernel/cpu/sgx/ioctl.c | 2 +- 2 files changed, 14 insertions(+), 2 deletions(-)
arch/x86/kernel/cpu/sgx/encls.h +13 -1
@@ -40,6 +40,18 @@ }while (0 ); \ } +/* +* encls_faulted() - Check if an ENCLS leaf faulted given an error code +* @ret the return value of an ENCLS leaf function call +* +* Return: +* %true if @ret indicates a fault, %false otherwise +*/ +static inline bool encls_faulted (int ret) +{ +return ret &ENCLS_FAULT_FLAG ; +} +/** * encls_failed() - Check if an ENCLS function failed * @ret: the return value of an ENCLS function call
@@ -50,7 +62,7 @@*/ static inline bool encls_failed (int ret) { -if (ret &ENCLS_FAULT_FLAG ) +if (encls_faulted (ret))return ENCLS_TRAPNR (ret) !=X86_TRAP_PF ;return !!ret;
arch/x86/kernel/cpu/sgx/ioctl.c +1 -1
@@ -568,7 +568,7 @@ static int sgx_encl_init(struct sgx_encl *encl, struct sgx_sigstruct *sigstruct, } } -if (ret &ENCLS_FAULT_FLAG ) { +if (encls_faulted (ret)) {if (encls_failed (ret))ENCLS_WARN (ret,"EINIT" );
Signature · 1 line
2.29.2
RFC PATCH 10/23 x86/sgx: Add helper to update SGX_LEPUBKEYHASHn MSRs
In reply to Kai HuangAdd a helper to update SGX_LEPUBKEYHASHn MSRs. SGX virtualization also needs to update those MSRs based on guest's "virtual" SGX_LEPUBKEYHASHn before EINIT from guest.
--- arch/x86/kernel/cpu/sgx/ioctl.c | 5 ++--- arch/x86/kernel/cpu/sgx/main.c | 8 ++++++++ arch/x86/kernel/cpu/sgx/sgx.h | 2 ++ 3 files changed, 12 insertions(+), 3 deletions(-)
arch/x86/kernel/cpu/sgx/ioctl.c +2 -3
@@ -495,7 +495,7 @@ static int sgx_encl_init(struct sgx_encl *encl, struct sgx_sigstruct *sigstruct, void *token) {u64 mrsigner[4 ]; -int i, j, k; +int i, j;void *addr;int ret;
@@ -544,8 +544,7 @@ static int sgx_encl_init(struct sgx_encl *encl, struct sgx_sigstruct *sigstruct,preempt_disable (); -for (k =0 ; k <4 ; k++) -wrmsrl (MSR_IA32_SGXLEPUBKEYHASH0 + k, mrsigner[k]); +sgx_update_lepubkeyhash (mrsigner); ret =__einit (sigstruct, token, addr);
arch/x86/kernel/cpu/sgx/main.c +8 -0
@@ -714,6 +714,14 @@ static bool __init sgx_page_cache_init(void)return true; } +void sgx_update_lepubkeyhash (u64 *lepubkeyhash) +{ +int i; + +for (i =0 ; i <4 ; i++) +wrmsrl (MSR_IA32_SGXLEPUBKEYHASH0 + i, lepubkeyhash[i]); +} +static void __init sgx_init (void ) {int ret;
arch/x86/kernel/cpu/sgx/sgx.h +2 -0
@@ -84,4 +84,6 @@ void sgx_mark_page_reclaimable(struct sgx_epc_page *page);int sgx_unmark_page_reclaimable (struct sgx_epc_page *page);struct sgx_epc_page *sgx_alloc_epc_page (void *owner,bool reclaim); +void sgx_update_lepubkeyhash (u64 *lepubkeyhash); +#endif /* _X86_SGX_H */
Signature · 1 line
2.29.2
RFC PATCH 11/23 x86/sgx: Add helpers to expose ECREATE and EINIT to KVM
In reply to Kai HuangFrom: Sean Christopherson <[email protected]>
Provide wrappers around __ecreate() and __einit() to hide the ugliness of overloading the ENCLS return value to encode multiple error formats in a single int. KVM will trap-and-execute ECREATE and EINIT as part of SGX virtualization, and on an exception, KVM needs the trapnr so that it can inject the correct fault into the guest.
--- arch/x86/include/asm/sgx.h | 16 ++++++++++ arch/x86/kernel/cpu/sgx/virt.c | 55 ++++++++++++++++++++++++++++++++++ 2 files changed, 71 insertions(+) create mode 100644 arch/x86/include/asm/sgx.h
arch/x86/include/asm/sgx.h new file +16 -0
@@ -0,0 +1,16 @@ +/* SPDX-License-Identifier: GPL-2.0 */ +#ifndef _ASM_X86_SGX_H +#define _ASM_X86_SGX_H + +#include <linux/types.h> + +#ifdef CONFIG_X86_SGX_VIRTUALIZATION +struct sgx_pageinfo ; + +int sgx_virt_ecreate (struct sgx_pageinfo *pageinfo,void __user *secs, +int *trapnr); +int sgx_virt_einit (void __user *sigstruct,void __user *token, +void __user *secs,u64 *lepubkeyhash,int *trapnr); +#endif + +#endif /* _ASM_X86_SGX_H */
arch/x86/kernel/cpu/sgx/virt.c +55 -0
58 lines not on this page · raw message
Signature · 1 line
2.29.2
RFC PATCH 12/23 x86/sgx: Move provisioning device creation out of SGX driver
In reply to Kai HuangFrom: Sean Christopherson <[email protected]>
And extract sgx_set_attribute() out of sgx_ioc_enclave_provision() and export it as symbol for KVM to use.
Provisioning key is sensitive. SGX driver only allows to create enclave which can access provisioning key when enclave creator has permission to open /dev/sgx_provision. It should apply to VM as well, as provisioning key is platform specific, thus unrestricted VM can also potentially compromise provisioning key.
Move provisioning device creation out of sgx_drv_init() to sgx_init() as preparation for adding SGX virtualization support, so that even SGX driver is not enabled due to flexible launch control is not available, SGX virtualization can still be enabled, and use it to restrict VM's capability of being able to access provisioning key.
--- arch/x86/include/asm/sgx.h | 3 +++ arch/x86/kernel/cpu/sgx/driver.c | 17 ------------ arch/x86/kernel/cpu/sgx/ioctl.c | 16 ++---------- arch/x86/kernel/cpu/sgx/main.c | 44 +++++++++++++++++++++++++++++++- 4 files changed, 48 insertions(+), 32 deletions(-)
arch/x86/include/asm/sgx.h +3 -0
@@ -4,6 +4,9 @@#include <linux/types.h> +int sgx_set_attribute (unsigned long *allowed_attributes, +unsigned int attribute_fd); +#ifdef CONFIG_X86_SGX_VIRTUALIZATION struct sgx_pageinfo ;
arch/x86/kernel/cpu/sgx/driver.c +0 -17
@@ -133,10 +133,6 @@ static const struct file_operations sgx_encl_fops = { .get_unmapped_area = sgx_get_unmapped_area, }; -const struct file_operations sgx_provision_fops = { - .owner =THIS_MODULE , -}; -static struct miscdevice sgx_dev_enclave = { .minor =MISC_DYNAMIC_MINOR , .name ="sgx_enclave" ,
@@ -144,13 +140,6 @@ static struct miscdevice sgx_dev_enclave = { .fops = &sgx_encl_fops, }; -static struct miscdevice sgx_dev_provision = { - .minor =MISC_DYNAMIC_MINOR , - .name ="sgx_provision" , - .nodename ="sgx_provision" , - .fops = &sgx_provision_fops, -}; -int __init sgx_drv_init (void ) {unsigned int eax, ebx, ecx, edx;
@@ -184,11 +173,5 @@ int __init sgx_drv_init(void)if (ret)return ret; - ret =misc_register (&sgx_dev_provision); -if (ret) { -misc_deregister (&sgx_dev_enclave); -return ret; - } -return 0 ; }
arch/x86/kernel/cpu/sgx/ioctl.c +2 -14
@@ -2,6 +2,7 @@/* Copyright(c) 2016-20 Intel Corporation. */ #include <asm/mman.h> +#include <asm/sgx.h> #include <linux/mman.h> #include <linux/delay.h> #include <linux/file.h>
@@ -664,24 +665,11 @@ static long sgx_ioc_enclave_init(struct sgx_encl *encl, void __user *arg)static long sgx_ioc_enclave_provision (struct sgx_encl *encl,void __user *arg) {struct sgx_enclave_provision params; -struct file *file;if (copy_from_user (¶ms, arg,sizeof (params)))return -EFAULT ; - file =fget (params.fd); -if (!file) -return -EINVAL ; - -if (file->f_op != &sgx_provision_fops) { -fput (file); -return -EINVAL ; - } - - encl->attributes_mask |=SGX_ATTR_PROVISIONKEY ; - -fput (file); -return 0 ; +return sgx_set_attribute (&encl->attributes_mask, params.fd); }long sgx_ioctl (struct file *filep,unsigned int cmd,unsigned long arg)
arch/x86/kernel/cpu/sgx/main.c +43 -1
77 lines not on this page · raw message
Signature · 1 line
2.29.2
RFC PATCH 13/23 KVM: VMX: Convert vcpu_vmx.exit_reason to a union
In reply to Kai HuangFrom: Sean Christopherson <[email protected]>
Convert vcpu_vmx.exit_reason from a u32 to a union (of size u32). The full VM_EXIT_REASON field is comprised of a 16-bit basic exit reason in bits 15:0, and single-bit modifiers in bits 31:16.
Historically, KVM has only had to worry about handling the "failed VM-Entry" modifier, which could only be set in very specific flows and required dedicated handling. I.e. manually stripping the FAILED_VMENTRY bit was a somewhat viable approach. But even with only a single bit to worry about, KVM has had several bugs related to comparing a basic exit reason against the full exit reason store in vcpu_vmx.
Upcoming Intel features, e.g. SGX, will add new modifier bits that can be set on more or less any VM-Exit, as opposed to the significantly more restricted FAILED_VMENTRY, i.e. correctly handling everything in one-off flows isn't scalable. Tracking exit reason in a union forces code to explicitly choose between consuming the full exit reason and the basic exit, and is a convenient way to document and access the modifiers.
No functional change intended.
--- arch/x86/kvm/vmx/nested.c | 42 +++++++++++++++--------- arch/x86/kvm/vmx/vmx.c | 68 ++++++++++++++++++++------------------- arch/x86/kvm/vmx/vmx.h | 25 +++++++++++++- 3 files changed, 86 insertions(+), 49 deletions(-)
arch/x86/kvm/vmx/nested.c +27 -15
@@ -3310,7 +3310,11 @@ enum nvmx_vmentry_status nested_vmx_enter_non_root_mode(struct kvm_vcpu *vcpu,struct vmcs12 *vmcs12 =get_vmcs12 (vcpu);enum vm_entry_failure_code entry_failure_code;bool evaluate_pending_interrupts; -u32 exit_reason, failed_index; +u32 failed_index; +union vmx_exit_reason exit_reason = { + .basic =-1 , + .failed_vmentry =1 , + };if (kvm_check_request (KVM_REQ_TLB_FLUSH_CURRENT , vcpu))kvm_vcpu_flush_tlb_current (vcpu);
@@ -3362,7 +3366,7 @@ enum nvmx_vmentry_status nested_vmx_enter_non_root_mode(struct kvm_vcpu *vcpu,if (nested_vmx_check_guest_state (vcpu, vmcs12, &entry_failure_code)) { - exit_reason =EXIT_REASON_INVALID_STATE ; + exit_reason.basic =EXIT_REASON_INVALID_STATE ; vmcs12->exit_qualification = entry_failure_code; goto vmentry_fail_vmexit; }
@@ -3373,7 +3377,7 @@ enum nvmx_vmentry_status nested_vmx_enter_non_root_mode(struct kvm_vcpu *vcpu, vcpu->arch.tsc_offset += vmcs12->tsc_offset;if (prepare_vmcs02 (vcpu, vmcs12, &entry_failure_code)) { - exit_reason =EXIT_REASON_INVALID_STATE ; + exit_reason.basic =EXIT_REASON_INVALID_STATE ; vmcs12->exit_qualification = entry_failure_code; goto vmentry_fail_vmexit_guest_mode; }
@@ -3383,7 +3387,7 @@ enum nvmx_vmentry_status nested_vmx_enter_non_root_mode(struct kvm_vcpu *vcpu, vmcs12->vm_entry_msr_load_addr, vmcs12->vm_entry_msr_load_count);if (failed_index) { - exit_reason =EXIT_REASON_MSR_LOAD_FAIL ; + exit_reason.basic =EXIT_REASON_MSR_LOAD_FAIL ; vmcs12->exit_qualification = failed_index; goto vmentry_fail_vmexit_guest_mode; }
@@ -3451,7 +3455,7 @@ enum nvmx_vmentry_status nested_vmx_enter_non_root_mode(struct kvm_vcpu *vcpu,return NVMX_VMENTRY_VMEXIT ;load_vmcs12_host_state (vcpu, vmcs12); - vmcs12->vm_exit_reason = exit_reason |VMX_EXIT_REASONS_FAILED_VMENTRY ; + vmcs12->vm_exit_reason = exit_reason.full;if (enable_shadow_vmcs || vmx->nested.hv_evmcs) vmx->nested.need_vmcs12_to_shadow_sync = true;return NVMX_VMENTRY_VMEXIT ;
@@ -5512,7 +5516,12 @@ static int handle_vmfunc(struct kvm_vcpu *vcpu)return kvm_skip_emulated_instruction (vcpu); fail: -nested_vmx_vmexit (vcpu, vmx->exit_reason, +/* +* This is effectively a reflected VM-Exit, as opposed to a synthesized +* nested VM-Exit. Pass the original exit reason, i.e. don't hardcode +* EXIT_REASON_VMFUNC as the exit reason. +*/ +nested_vmx_vmexit (vcpu, vmx->exit_reason.full,vmx_get_intr_info (vcpu),vmx_get_exit_qual (vcpu));return 1 ;
@@ -5580,7 +5589,8 @@ static bool nested_vmx_exit_handled_io(struct kvm_vcpu *vcpu,* MSR bitmap. This may be the case even when L0 doesn't use MSR bitmaps. */ static bool nested_vmx_exit_handled_msr (struct kvm_vcpu *vcpu, -struct vmcs12 *vmcs12,u32 exit_reason) +struct vmcs12 *vmcs12, +union vmx_exit_reason exit_reason) {u32 msr_index =kvm_rcx_read (vcpu);gpa_t bitmap;
@@ -5594,7 +5604,7 @@ static bool nested_vmx_exit_handled_msr(struct kvm_vcpu *vcpu,* First we need to figure out which of the four to use: */ bitmap = vmcs12->msr_bitmap; -if (exit_reason ==EXIT_REASON_MSR_WRITE ) +if (exit_reason.basic ==EXIT_REASON_MSR_WRITE ) bitmap +=2048 ;if (msr_index >=0xc0000000 ) { msr_index -=0xc0000000 ;
@@ -5731,11 +5741,12 @@ static bool nested_vmx_exit_handled_mtf(struct vmcs12 *vmcs12)* Return true if L0 wants to handle an exit from L2 regardless of whether or not * L1 wants the exit. Only call this when in is_guest_mode (L2). */ -static bool nested_vmx_l0_wants_exit (struct kvm_vcpu *vcpu,u32 exit_reason) +static bool nested_vmx_l0_wants_exit (struct kvm_vcpu *vcpu, +union vmx_exit_reason exit_reason) {u32 intr_info; -switch ((u16 )exit_reason) { +switch (exit_reason.basic) {case EXIT_REASON_EXCEPTION_NMI : intr_info =vmx_get_intr_info (vcpu);if (is_nmi (intr_info))
@@ -5791,12 +5802,13 @@ static bool nested_vmx_l0_wants_exit(struct kvm_vcpu *vcpu, u32 exit_reason)* Return 1 if L1 wants to intercept an exit from L2. Only call this when in * is_guest_mode (L2). */ -static bool nested_vmx_l1_wants_exit (struct kvm_vcpu *vcpu,u32 exit_reason) +static bool nested_vmx_l1_wants_exit (struct kvm_vcpu *vcpu, +union vmx_exit_reason exit_reason) {struct vmcs12 *vmcs12 =get_vmcs12 (vcpu);u32 intr_info; -switch ((u16 )exit_reason) { +switch (exit_reason.basic) {case EXIT_REASON_EXCEPTION_NMI : intr_info =vmx_get_intr_info (vcpu);if (is_nmi (intr_info))
@@ -5915,7 +5927,7 @@ static bool nested_vmx_l1_wants_exit(struct kvm_vcpu *vcpu, u32 exit_reason)bool nested_vmx_reflect_vmexit (struct kvm_vcpu *vcpu) {struct vcpu_vmx *vmx =to_vmx (vcpu); -u32 exit_reason = vmx->exit_reason; +union vmx_exit_reason exit_reason = vmx->exit_reason;unsigned long exit_qual;u32 exit_intr_info;
@@ -5934,7 +5946,7 @@ bool nested_vmx_reflect_vmexit(struct kvm_vcpu *vcpu) goto reflect_vmexit; } -trace_kvm_nested_vmexit (exit_reason, vcpu,KVM_ISA_VMX ); +trace_kvm_nested_vmexit (exit_reason.full, vcpu,KVM_ISA_VMX );/* If L0 (KVM) wants the exit, it trumps L1's desires. */ if (nested_vmx_l0_wants_exit (vcpu, exit_reason))
@@ -5960,7 +5972,7 @@ bool nested_vmx_reflect_vmexit(struct kvm_vcpu *vcpu) exit_qual =vmx_get_exit_qual (vcpu); reflect_vmexit: -nested_vmx_vmexit (vcpu, exit_reason, exit_intr_info, exit_qual); +nested_vmx_vmexit (vcpu, exit_reason.full, exit_intr_info, exit_qual);return true; }
arch/x86/kvm/vmx/vmx.c +35 -33
157 lines not on this page · raw message
arch/x86/kvm/vmx/vmx.h +24 -1
@@ -70,6 +70,29 @@ struct pt_desc {struct pt_ctx guest; }; +union vmx_exit_reason { +struct { +u32 basic :16 ; +u32 reserved16 :1 ; +u32 reserved17 :1 ; +u32 reserved18 :1 ; +u32 reserved19 :1 ; +u32 reserved20 :1 ; +u32 reserved21 :1 ; +u32 reserved22 :1 ; +u32 reserved23 :1 ; +u32 reserved24 :1 ; +u32 reserved25 :1 ; +u32 reserved26 :1 ; +u32 sgx_enclave_mode :1 ; +u32 smi_pending_mtf :1 ; +u32 smi_from_vmx_root :1 ; +u32 reserved30 :1 ; +u32 failed_vmentry :1 ; + }; +u32 full; +}; +/* * The nested_vmx structure is part of vcpu_vmx, and holds information we need * for correct emulation of VMX (i.e., nested VMX) on this vcpu.
@@ -244,7 +267,7 @@ struct vcpu_vmx {int vpid;bool emulation_required; -u32 exit_reason; +union vmx_exit_reason exit_reason;/* Posted interrupt descriptor */ struct pi_desc pi_desc;
Signature · 1 line
2.29.2
RFC PATCH 14/23 KVM: x86: Export kvm_mmu_gva_to_gpa_{read,write}() for SGX (VMX)
In reply to Kai HuangFrom: Sean Christopherson <[email protected]>
Export the gva_to_gpa() helpers for use by SGX virtualization when executing ENCLS[ECREATE] and ENCLS[EINIT] on behalf of the guest. To execute ECREATE and EINIT, KVM must obtain the GPA of the target Secure Enclave Control Structure (SECS) in order to get its corresponding HVA.
Because the SECS must reside in the Enclave Page Cache (EPC), copying the SECS's data to a host-controlled buffer via existing exported helpers is not a viable option as the EPC is not readable or writable by the kernel.
SGX virtualization will also use gva_to_gpa() to obtain HVAs for non-EPC pages in order to pass user pointers directly to ECREATE and EINIT, which avoids having to copy pages worth of data into the kernel.
--- arch/x86/kvm/x86.c | 2 ++ 1 file changed, 2 insertions(+)
arch/x86/kvm/x86.c +2 -0
@@ -5866,6 +5866,7 @@ gpa_t kvm_mmu_gva_to_gpa_read(struct kvm_vcpu *vcpu, gva_t gva,u32 access = (kvm_x86_ops.get_cpl (vcpu) ==3 ) ?PFERR_USER_MASK :0 ;return vcpu->arch.walk_mmu->gva_to_gpa (vcpu, gva, access, exception); } +EXPORT_SYMBOL_GPL (kvm_mmu_gva_to_gpa_read);gpa_t kvm_mmu_gva_to_gpa_fetch (struct kvm_vcpu *vcpu,gva_t gva,struct x86_exception *exception)
@@ -5882,6 +5883,7 @@ gpa_t kvm_mmu_gva_to_gpa_write(struct kvm_vcpu *vcpu, gva_t gva, access |=PFERR_WRITE_MASK ;return vcpu->arch.walk_mmu->gva_to_gpa (vcpu, gva, access, exception); } +EXPORT_SYMBOL_GPL (kvm_mmu_gva_to_gpa_write);/* uses this to access any guest's mapped memory without checking CPL */ gpa_t kvm_mmu_gva_to_gpa_system(struct kvm_vcpu *vcpu, gva_t gva,
Signature · 1 line
2.29.2
RFC PATCH 15/23 KVM: x86: Define new #PF SGX error code bit
In reply to Kai HuangFrom: Sean Christopherson <[email protected]>
Page faults that are signaled by the SGX Enclave Page Cache Map (EPCM), as opposed to the traditional IA32/EPT page tables, set an SGX bit in the error code to indicate that the #PF was induced by SGX. KVM will need to emulate this behavior as part of its trap-and-execute scheme for virtualizing SGX Launch Control, e.g. to inject SGX-induced #PFs if EINIT faults in the host, and to support live migration.
--- arch/x86/include/asm/kvm_host.h | 2 ++ 1 file changed, 2 insertions(+)
arch/x86/include/asm/kvm_host.h +2 -0
@@ -216,6 +216,7 @@ enum x86_intercept_stage;#define PFERR_RSVD_BIT 3#define PFERR_FETCH_BIT 4#define PFERR_PK_BIT 5 +#define PFERR_SGX_BIT 15#define PFERR_GUEST_FINAL_BIT 32#define PFERR_GUEST_PAGE_BIT 33
@@ -225,6 +226,7 @@ enum x86_intercept_stage;#define PFERR_RSVD_MASK (1U << PFERR_RSVD_BIT)#define PFERR_FETCH_MASK (1U << PFERR_FETCH_BIT)#define PFERR_PK_MASK (1U << PFERR_PK_BIT) +#define PFERR_SGX_MASK (1U << PFERR_SGX_BIT)#define PFERR_GUEST_FINAL_MASK (1ULL << PFERR_GUEST_FINAL_BIT)#define PFERR_GUEST_PAGE_MASK (1ULL << PFERR_GUEST_PAGE_BIT)
Signature · 1 line
2.29.2
RFC PATCH 16/23 KVM: x86: Add SGX feature leaf to reverse CPUID lookup
In reply to Kai HuangFrom: Sean Christopherson <[email protected]>
Add SGX's sub-features leaf to the reverse CPUID lookup table in preparation for adding SGX virtualization.
--- arch/x86/kvm/cpuid.h | 1 + 1 file changed, 1 insertion(+)
arch/x86/kvm/cpuid.h +1 -0
@@ -63,6 +63,7 @@ static const struct cpuid_reg reverse_cpuid[] = { [CPUID_8000_0007_EBX] = {0x80000007, 0, CPUID_EBX}, [CPUID_7_EDX] = { 7, 0, CPUID_EDX}, [CPUID_7_1_EAX] = { 7, 1, CPUID_EAX}, + [CPUID_12_EAX] = { 0x12, 0, CPUID_EAX}, };/*
Signature · 1 line
2.29.2
RFC PATCH 17/23 KVM: VMX: Add basic handling of VM-Exit from SGX enclave
In reply to Kai HuangFrom: Sean Christopherson <[email protected]>
Add support for handling VM-Exits that originate from a guest SGX enclave. In SGX, an "enclave" is a new CPL3-only execution environment, wherein the CPU and memory state is protected by hardware to make the state inaccesible to code running outside of the enclave. When exiting an enclave due to an asynchronous event (from the perspective of the enclave), e.g. exceptions, interrupts, and VM-Exits, the enclave's state is automatically saved and scrubbed (the CPU loads synthetic state), and then reloaded when re-entering the enclave. E.g. after an instruction based VM-Exit from an enclave, vmcs.GUEST_RIP will not contain the RIP of the enclave instruction that trigered VM-Exit, but will instead point to a RIP in the enclave's untrusted runtime (the guest userspace code that coordinates entry/exit to/from the enclave).
To help a VMM recognize and handle exits from enclaves, SGX adds bits to existing VMCS fields, VM_EXIT_REASON.VMX_EXIT_REASON_FROM_ENCLAVE and GUEST_INTERRUPTIBILITY_INFO.GUEST_INTR_STATE_ENCLAVE_INTR. Define the new architectural bits, and add a boolean to struct vcpu_vmx to cache VMX_EXIT_REASON_FROM_ENCLAVE. Clear the bit in exit_reason so that checks against exit_reason do not need to account for SGX, e.g. "if (exit_reason == EXIT_REASON_EXCEPTION_NMI)" continues to work.
KVM is a largely a passive observer of the new bits, e.g. KVM needs to account for the bits when propagating information to a nested VMM, but otherwise doesn't need to act differently for the majority of VM-Exits from enclaves.
The one scenario that is directly impacted is emulation, which is for all intents and purposes impossible[1] since KVM does not have access to the RIP or instruction stream that triggered the VM-Exit. The inability to emulate is a non-issue for KVM, as most instructions that might trigger VM-Exit unconditionally #UD in an enclave (before the VM-Exit check. For the few instruction that conditionally #UD, KVM either never sets the exiting control, e.g. PAUSE_EXITING[2], or sets it if and only if the feature is not exposed to the guest in order to inject a #UD, e.g. RDRAND_EXITING.
But, because it is still possible for a guest to trigger emulation, e.g. MMIO, inject a #UD if KVM ever attempts emulation after a VM-Exit from an enclave. This is architecturally accurate for instruction VM-Exits, and for MMIO it's the least bad choice, e.g. it's preferable to killing the VM. In practice, only broken or particularly stupid guests should ever encounter this behavior.
Add a WARN in skip_emulated_instruction to detect any attempt to modify the guest's RIP during an SGX enclave VM-Exit as all such flows should either be unreachable or must handle exits from enclaves before getting to skip_emulated_instruction.
[1] Impossible for all practical purposes. Not truly impossible since KVM could implement some form of para-virtualization scheme.
[2] PAUSE_LOOP_EXITING only affects CPL0 and enclaves exist only at CPL3, so we also don't need to worry about that interaction.
--- arch/x86/include/asm/vmx.h | 1 + arch/x86/include/uapi/asm/vmx.h | 1 + arch/x86/kvm/vmx/nested.c | 2 ++ arch/x86/kvm/vmx/vmx.c | 38 +++++++++++++++++++++++++++++++-- 4 files changed, 40 insertions(+), 2 deletions(-)
arch/x86/include/asm/vmx.h +1 -0
@@ -371,6 +371,7 @@ enum vmcs_field {#define GUEST_INTR_STATE_MOV_SS 0x00000002#define GUEST_INTR_STATE_SMI 0x00000004#define GUEST_INTR_STATE_NMI 0x00000008 +#define GUEST_INTR_STATE_ENCLAVE_INTR 0x00000010/* GUEST_ACTIVITY_STATE flags */ #define GUEST_ACTIVITY_ACTIVE 0
arch/x86/include/uapi/asm/vmx.h +1 -0
@@ -27,6 +27,7 @@#define VMX_EXIT_REASONS_FAILED_VMENTRY 0x80000000 +#define VMX_EXIT_REASONS_SGX_ENCLAVE_MODE 0x08000000#define EXIT_REASON_EXCEPTION_NMI 0#define EXIT_REASON_EXTERNAL_INTERRUPT 1
arch/x86/kvm/vmx/nested.c +2 -0
@@ -4100,6 +4100,8 @@ static void prepare_vmcs12(struct kvm_vcpu *vcpu, struct vmcs12 *vmcs12, {/* update exit information fields: */ vmcs12->vm_exit_reason = vm_exit_reason; +if (to_vmx (vcpu)->exit_reason.sgx_enclave_mode) + vmcs12->vm_exit_reason |=VMX_EXIT_REASONS_SGX_ENCLAVE_MODE ; vmcs12->exit_qualification = exit_qualification; vmcs12->vm_exit_intr_info = exit_intr_info;
arch/x86/kvm/vmx/vmx.c +36 -2
@@ -1562,12 +1562,18 @@ static int vmx_rtit_ctl_check(struct kvm_vcpu *vcpu, u64 data)static bool vmx_can_emulate_instruction (struct kvm_vcpu *vcpu,void *insn,int insn_len) { +if (to_vmx (vcpu)->exit_reason.sgx_enclave_mode) { +kvm_queue_exception (vcpu,UD_VECTOR ); +return false; + }return true; }static int skip_emulated_instruction (struct kvm_vcpu *vcpu) { +union vmx_exit_reason exit_reason =to_vmx (vcpu)->exit_reason;unsigned long rip, orig_rip; +u32 instr_len;/* * Using VMCS.VM_EXIT_INSTRUCTION_LEN on EPT misconfig depends on
@@ -1578,9 +1584,33 @@ static int skip_emulated_instruction(struct kvm_vcpu *vcpu)* i.e. we end up advancing IP with some random value. */ if (!static_cpu_has (X86_FEATURE_HYPERVISOR ) || -to_vmx (vcpu)->exit_reason.basic !=EXIT_REASON_EPT_MISCONFIG ) { + exit_reason.basic !=EXIT_REASON_EPT_MISCONFIG ) { + instr_len =vmcs_read32 (VM_EXIT_INSTRUCTION_LEN ); + +/* +* Emulating an enclave's instructions isn't supported as KVM +* cannot access the enclave's memory or its true RIP, e.g. the +* vmcs.GUEST_RIP points at the exit point of the enclave, not +* the RIP that actually triggered the VM-Exit. But, because +* most instructions that cause VM-Exit will #UD in an enclave, +* most instruction-based VM-Exits simply do not occur. +* +* There are a few exceptions, notably the debug instructions +* INT1ICEBRK and INT3, as they are allowed in debug enclaves +* and generate #DB/#BP as expected, which KVM might intercept. +* But again, the CPU does the dirty work and saves an instr +* length of zero so VMMs don't shoot themselves in the foot. +* WARN if KVM tries to skip a non-zero length instruction on +* a VM-Exit from an enclave. +*/ +if (!instr_len) + goto rip_updated; + +WARN (exit_reason.sgx_enclave_mode, +"KVM: skipping instruction after SGX enclave VM-Exit" ); + orig_rip =kvm_rip_read (vcpu); - rip = orig_rip +vmcs_read32 (VM_EXIT_INSTRUCTION_LEN ); + rip = orig_rip + instr_len;#ifdef CONFIG_X86_64 /* * We need to mask out the high 32 bits of RIP if not in 64-bit
@@ -1596,6 +1626,7 @@ static int skip_emulated_instruction(struct kvm_vcpu *vcpu)return 0 ; } +rip_updated:/* skipping an emulated instruction also counts */ vmx_set_interrupt_shadow (vcpu,0 );
@@ -5361,6 +5392,9 @@ static int handle_ept_misconfig(struct kvm_vcpu *vcpu) {gpa_t gpa; +if (!vmx_can_emulate_instruction (vcpu,NULL ,0 )) +return 1 ; +/* * A nested guest cannot optimize MMIO vmexits, because we have an * nGPA here instead of the required GPA.
Signature · 1 line
2.29.2
RFC PATCH 18/23 KVM: VMX: Frame in ENCLS handler for SGX virtualization
In reply to Kai HuangFrom: Sean Christopherson <[email protected]>
Introduce sgx.c and sgx.h, along with the framework for handling ENCLS VM-Exits. Add a bool, enable_sgx, that will eventually be wired up to a module param to control whether or not SGX virtualization is enabled at runtime.
--- arch/x86/kvm/Makefile | 2 ++ arch/x86/kvm/vmx/sgx.c | 51 ++++++++++++++++++++++++++++++++++++++++++ arch/x86/kvm/vmx/sgx.h | 15 +++++++++++++ arch/x86/kvm/vmx/vmx.c | 9 +++++--- 4 files changed, 74 insertions(+), 3 deletions(-) create mode 100644 arch/x86/kvm/vmx/sgx.c create mode 100644 arch/x86/kvm/vmx/sgx.h
arch/x86/kvm/Makefile +2 -0
@@ -20,6 +20,8 @@ kvm-y += x86.o emulate.o i8259.o irq.o lapic.o \ kvm-intel-y += vmx/vmx.o vmx/vmenter.o vmx/pmu_intel.o vmx/vmcs12.o \ vmx/evmcs.o vmx/nested.o vmx/posted_intr.o +kvm-intel-$(CONFIG_X86_SGX_VIRTUALIZATION) += vmx/sgx.o + kvm-amd-y += svm/svm.o svm/vmenter.o svm/pmu.o svm/nested.o svm/avic.o svm/sev.o obj-$(CONFIG_KVM) += kvm.o
arch/x86/kvm/vmx/sgx.c new file +51 -0
51 lines not on this page · raw message
arch/x86/kvm/vmx/sgx.h new file +15 -0
@@ -0,0 +1,15 @@ +/* SPDX-License-Identifier: GPL-2.0 */ +#ifndef __KVM_X86_SGX_H +#define __KVM_X86_SGX_H + +#include <linux/kvm_host.h> + +#ifdef CONFIG_X86_SGX_VIRTUALIZATION +extern bool __read_mostly enable_sgx; + +int handle_encls (struct kvm_vcpu *vcpu); +#else +#define enable_sgx 0 +#endif + +#endif /* __KVM_X86_SGX_H */
arch/x86/kvm/vmx/vmx.c +6 -3
@@ -57,6 +57,7 @@#include "mmu.h" #include "nested.h" #include "pmu.h" +#include "sgx.h" #include "trace.h" #include "vmcs.h" #include "vmcs12.h"
@@ -5643,16 +5644,18 @@ static int handle_vmx_instruction(struct kvm_vcpu *vcpu)return 1 ; } +#ifndef CONFIG_X86_SGX_VIRTUALIZATION static int handle_encls (struct kvm_vcpu *vcpu) {/* -* SGX virtualization is not yet supported. There is no software -* enable bit for SGX, so we have to trap ENCLS and inject a #UD -* to prevent the guest from executing ENCLS. +* SGX virtualization is disabled. There is no software enable bit for +* SGX, so KVM intercepts all ENCLS leafs and injects a #UD to prevent +* the guest from executing ENCLS (when SGX is supported by hardware). */ kvm_queue_exception (vcpu,UD_VECTOR );return 1 ; } +#endif /* CONFIG_X86_SGX_VIRTUALIZATION */ /* * The exit handlers return 1 if the exit was handled fully and guest execution
Signature · 1 line
2.29.2
RFC PATCH 19/23 KVM: VMX: Add SGX ENCLS[ECREATE] handler to enforce CPUID restrictions
In reply to Kai HuangFrom: Sean Christopherson <[email protected]>
Add an ECREATE handler that will be used to intercept ECREATE for the purpose of enforcing and enclave's MISCSELECT, ATTRIBUTES and XFRM, i.e. to allow userspace to restrict SGX features via CPUID. ECREATE will be intercepted when any of the aforementioned masks diverges from hardware in order to enforce the desired CPUID model, i.e. inject #GP if the guest attempts to set a bit that hasn't been enumerated as allowed-1 in CPUID.
Note, access to the PROVISIONKEY is not yet supported.
--- arch/x86/include/asm/kvm_host.h | 3 + arch/x86/kvm/vmx/sgx.c | 243 ++++++++++++++++++++++++++++++++ 2 files changed, 246 insertions(+)
arch/x86/include/asm/kvm_host.h +3 -0
@@ -996,6 +996,9 @@ struct kvm_arch {struct msr_bitmap_range ranges[16 ]; } msr_filter; +/* Guest can access the SGX PROVISIONKEY. */ +bool sgx_provisioning_allowed; +struct kvm_pmu_event_filter *pmu_event_filter;struct task_struct *nx_lpage_recovery_thread;
arch/x86/kvm/vmx/sgx.c +243 -0
255 lines not on this page · raw message
Signature · 1 line
2.29.2
RFC PATCH 20/23 KVM: VMX: Add emulation of SGX Launch Control LE hash MSRs
In reply to Kai HuangFrom: Sean Christopherson <[email protected]>
Emulate the four Launch Enclave public key hash MSRs (LE hash MSRs) that exist on CPUs that support SGX Launch Control (LC). SGX LC modifies the behavior of ENCLS[EINIT] to use the LE hash MSRs when verifying the key used to sign an enclave. On CPUs without LC support, the LE hash is hardwired into the CPU to an Intel controlled key (the Intel key is also the reset value of the LE hash MSRs). Track the guest's desired hash so that a future patch can stuff the hash into the hardware MSRs when executing EINIT on behalf of the guest, when those MSRs are writable in host.
Note, KVM allows writes to the LE hash MSRs if IA32_FEATURE_CONTROL is unlocked. This is technically not architectural behavior, but it's roughly equivalent to the arch behavior of the MSRs being writable prior to activating SGX[1]. Emulating SGX activation is feasible, but adds no tangible benefits and would just create extra work for KVM and guest firmware.
[1] SGX related bits in IA32_FEATURE_CONTROL cannot be set until SGX is activated, e.g. by firmware. SGX activation is triggered by setting bit 0 in MSR 0x7a. Until SGX is activated, the LE hash MSRs are writable, e.g. to allow firmware to lock down the LE root key with a non-Intel value.
--- arch/x86/kvm/vmx/sgx.c | 35 +++++++++++++++++++++++++++++++++++ arch/x86/kvm/vmx/sgx.h | 6 ++++++ arch/x86/kvm/vmx/vmx.c | 20 ++++++++++++++++++++ arch/x86/kvm/vmx/vmx.h | 2 ++ 4 files changed, 63 insertions(+)
arch/x86/kvm/vmx/sgx.c +35 -0
@@ -12,6 +12,9 @@bool __read_mostly enable_sgx; +/* Initial value of guest's virtual SGX_LEPUBKEYHASHn MSRs */ +static u64 sgx_pubkey_hash[4 ]__ro_after_init ; +/* * ENCLS's memory operands use a fixed segment (DS) and a fixed * address size based on the mode. Related prefixes are ignored.
@@ -292,3 +295,35 @@ int handle_encls(struct kvm_vcpu *vcpu) }return 1 ; } + +void setup_default_sgx_lepubkeyhash (void ) +{ +/* +* Use Intel's default value for Skylake hardware if Launch Control is +* not supported, i.e. Intel's hash is hardcoded into silicon, or if +* Launch Control is supported and enabled, i.e. mimic the reset value +* and let the guest write the MSRs at will. If Launch Control is +* supported but disabled, then use the current MSR values as the hash +* MSRs exist but are read-only (locked and not writable). +*/ +if (!enable_sgx || !boot_cpu_has (X86_FEATURE_SGX_LC ) || +rdmsrl_safe (MSR_IA32_SGXLEPUBKEYHASH0 , &sgx_pubkey_hash[0 ])) { + sgx_pubkey_hash[0 ] =0xa6053e051270b7acULL ; + sgx_pubkey_hash[1 ] =0x6cfbe8ba8b3b413dULL ; + sgx_pubkey_hash[2 ] =0xc4916d99f2b3735dULL ; + sgx_pubkey_hash[3 ] =0xd4f8c05909f9bb3bULL ; + }else { +/* MSR_IA32_SGXLEPUBKEYHASH0 is read above */ +rdmsrl (MSR_IA32_SGXLEPUBKEYHASH1 , sgx_pubkey_hash[1 ]); +rdmsrl (MSR_IA32_SGXLEPUBKEYHASH2 , sgx_pubkey_hash[2 ]); +rdmsrl (MSR_IA32_SGXLEPUBKEYHASH3 , sgx_pubkey_hash[3 ]); + } +} + +void vcpu_setup_sgx_lepubkeyhash (struct kvm_vcpu *vcpu) +{ +struct vcpu_vmx *vmx =to_vmx (vcpu); + +memcpy (vmx->msr_ia32_sgxlepubkeyhash, sgx_pubkey_hash, +sizeof (sgx_pubkey_hash)); +}
arch/x86/kvm/vmx/sgx.h +6 -0
@@ -8,8 +8,14 @@extern bool __read_mostly enable_sgx;int handle_encls (struct kvm_vcpu *vcpu); + +void setup_default_sgx_lepubkeyhash (void ); +void vcpu_setup_sgx_lepubkeyhash (struct kvm_vcpu *vcpu);#else #define enable_sgx 0 + +static inline void setup_default_sgx_lepubkeyhash (void ) { } +static inline void vcpu_setup_sgx_lepubkeyhash (struct kvm_vcpu *vcpu) { }#endif #endif /* __KVM_X86_SGX_H */
arch/x86/kvm/vmx/vmx.c +20 -0
@@ -1889,6 +1889,13 @@ static int vmx_get_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)case MSR_IA32_FEAT_CTL : msr_info->data = vmx->msr_ia32_feature_control;break ; + case MSR_IA32_SGXLEPUBKEYHASH0 ... MSR_IA32_SGXLEPUBKEYHASH3: +if (!msr_info->host_initiated && + !guest_cpuid_has (vcpu,X86_FEATURE_SGX_LC )) +return 1 ; + msr_info->data =to_vmx (vcpu)->msr_ia32_sgxlepubkeyhash + [msr_info->index -MSR_IA32_SGXLEPUBKEYHASH0 ]; +break ; case MSR_IA32_VMX_BASIC ... MSR_IA32_VMX_VMFUNC:if (!nested_vmx_allowed (vcpu))return 1 ;
@@ -2155,6 +2162,15 @@ static int vmx_set_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)if (msr_info->host_initiated && data ==0 )vmx_leave_nested (vcpu);break ; + case MSR_IA32_SGXLEPUBKEYHASH0 ... MSR_IA32_SGXLEPUBKEYHASH3: +if (!msr_info->host_initiated && + (!guest_cpuid_has (vcpu,X86_FEATURE_SGX_LC ) || + ((vmx->msr_ia32_feature_control &FEAT_CTL_LOCKED ) && + !(vmx->msr_ia32_feature_control &FEAT_CTL_SGX_LC_ENABLED )))) +return 1 ; + vmx->msr_ia32_sgxlepubkeyhash + [msr_index -MSR_IA32_SGXLEPUBKEYHASH0 ] = data; +break ; case MSR_IA32_VMX_BASIC ... MSR_IA32_VMX_VMFUNC:if (!msr_info->host_initiated)return 1 ;/* they are read-only */
@@ -6973,6 +6989,8 @@ static int vmx_create_vcpu(struct kvm_vcpu *vcpu)else memset (&vmx->nested.msrs,0 ,sizeof (vmx->nested.msrs)); +vcpu_setup_sgx_lepubkeyhash (vcpu); + vmx->nested.posted_intr_nv =-1 ; vmx->nested.current_vmptr =-1ull ;
@@ -7912,6 +7930,8 @@ static __init int hardware_setup(void)if (!enable_ept || !cpu_has_vmx_intel_pt ()) pt_mode =PT_MODE_SYSTEM ; +setup_default_sgx_lepubkeyhash (); +if (nested) {nested_vmx_setup_ctls_msrs (&vmcs_config.nested, vmx_capability.ept);
arch/x86/kvm/vmx/vmx.h +2 -0
@@ -299,6 +299,8 @@ struct vcpu_vmx {*/ u64 msr_ia32_feature_control;u64 msr_ia32_feature_control_valid_bits; +/* SGX Launch Control public key hash */ +u64 msr_ia32_sgxlepubkeyhash[4 ];u64 ept_pointer;struct pt_desc pt_desc;
Signature · 1 line
2.29.2
RFC PATCH 22/23 KVM: VMX: Enable SGX virtualization for SGX1, SGX2 and LC
In reply to Kai HuangFrom: Sean Christopherson <[email protected]>
Enable SGX virtualization now that KVM has the VM-Exit handlers needed to trap-and-execute ENCLS to ensure correctness and/or enforce the CPU model exposed to the guest. Add a KVM module param, "sgx", to allow an admin to disable SGX virtualization independent of the kernel.
When supported in hardware and the kernel, advertise SGX1, SGX2 and SGX LC to userspace via CPUID and wire up the ENCLS_EXITING bitmap based on the guest's SGX capabilities, i.e. to allow ENCLS to be executed in an SGX-enabled guest. With the exception of the provision key, all SGX attribute bits may be exposed to the guest. Guest access to the provision key, which is controlled via securityfs, will be added in a future patch.
Note, KVM does not yet support exposing ENCLS_C leafs or ENCLV leafs.
--- arch/x86/kvm/cpuid.c | 58 +++++++++++++++++++++++++++- arch/x86/kvm/vmx/nested.c | 26 +++++++++++-- arch/x86/kvm/vmx/nested.h | 5 +++ arch/x86/kvm/vmx/sgx.c | 80 ++++++++++++++++++++++++++++++++++++++- arch/x86/kvm/vmx/sgx.h | 13 +++++++ arch/x86/kvm/vmx/vmcs12.c | 1 + arch/x86/kvm/vmx/vmcs12.h | 4 +- arch/x86/kvm/vmx/vmx.c | 38 ++++++++++++++++++- 8 files changed, 216 insertions(+), 9 deletions(-)
arch/x86/kvm/cpuid.c +56 -2
@@ -18,6 +18,7 @@#include <asm/processor.h> #include <asm/user.h> #include <asm/fpu/xstate.h> +#include <asm/sgx_arch.h> #include "cpuid.h" #include "lapic.h" #include "mmu.h"
@@ -169,6 +170,21 @@ static void kvm_vcpu_after_set_cpuid(struct kvm_vcpu *vcpu) vcpu->arch.guest_supported_xcr0 = (best->eax | ((u64 )best->edx <<32 )) & supported_xcr0; +/* +* Bits 127:0 of the allowed SECS.ATTRIBUTES (CPUID.0x12.0x1) enumerate +* the supported XSAVE Feature Request Mask (XFRM), i.e. the enclave's +* requested XCR0 value. The enclave's XFRM must be a subset of XCRO +* at the time of EENTER, thus adjust the allowed XFRM by the guest's +* supported XCR0. Similar to XCR0 handling, FP and SSE are forced to +* '1' even on CPUs that don't support XSAVE. +*/ + best =kvm_find_cpuid_entry (vcpu,0x12 ,0x1 ); +if (best) { + best->ecx &= vcpu->arch.guest_supported_xcr0 &0xffffffff ; + best->edx &= vcpu->arch.guest_supported_xcr0 >>32 ; + best->ecx |=XFEATURE_MASK_FPSSE ; + } +kvm_update_pv_runtime (vcpu); vcpu->arch.maxphyaddr =cpuid_query_maxphyaddr (vcpu);
@@ -390,7 +406,7 @@ void kvm_set_cpu_caps(void) ); kvm_cpu_cap_mask(CPUID_7_0_EBX, - F(FSGSBASE) | F(BMI1) | F(HLE) | F(AVX2) | F(SMEP) | + F(FSGSBASE) | F(SGX) | F(BMI1) | F(HLE) | F(AVX2) | F(SMEP) | F(BMI2) | F(ERMS) | 0 /*INVPCID*/ | F(RTM) | 0 /*MPX*/ | F(RDSEED) | F(ADX) | F(SMAP) | F(AVX512IFMA) | F(AVX512F) | F(AVX512PF) | F(AVX512ER) | F(AVX512CD) | F(CLFLUSHOPT) | F(CLWB) | F(AVX512DQ) |
@@ -401,7 +417,8 @@ void kvm_set_cpu_caps(void)F (AVX512VBMI ) |F (LA57 ) |F (PKU ) |0 /*OSPKE*/ |F (RDPID ) |F (AVX512_VPOPCNTDQ ) |F (UMIP ) |F (AVX512_VBMI2 ) |F (GFNI ) |F (VAES ) |F (VPCLMULQDQ ) |F (AVX512_VNNI ) |F (AVX512_BITALG ) | -F (CLDEMOTE ) |F (MOVDIRI ) |F (MOVDIR64B ) |0 /*WAITPKG*/ +F (CLDEMOTE ) |F (MOVDIRI ) |F (MOVDIR64B ) |0 /*WAITPKG*/ | +F (SGX_LC ) );/* Set LA57 based on hardware capability. */ if (cpuid_ecx (7 ) &F (LA57 ))
@@ -440,6 +457,11 @@ void kvm_set_cpu_caps(void)F (XSAVEOPT ) |F (XSAVEC ) |F (XGETBV1 ) |F (XSAVES ) ); +kvm_cpu_cap_mask (CPUID_12_EAX , +F (SGX1 ) |F (SGX2 ) |0 /* Reserved */ |0 /* Reserved */ | +0 /* Reserved */ |0 /* ENCLV */ |0 /* ENCLS_C */ + ); + kvm_cpu_cap_mask(CPUID_8000_0001_ECX, F(LAHF_LM) | F(CMP_LEGACY) | 0 /*SVM*/ | 0 /* ExtApicSpace */ | F(CR8_LEGACY) | F(ABM) | F(SSE4A) | F(MISALIGNSSE) |
@@ -761,6 +783,38 @@ static inline int __do_cpuid_func(struct kvm_cpuid_array *array, u32 function) entry->edx =0 ; }break ; +case 0x12 : +/* Intel SGX */ +if (!kvm_cpu_cap_has (X86_FEATURE_SGX )) { + entry->eax = entry->ebx = entry->ecx = entry->edx =0 ; +break ; + } + +/* +* Index 0: Sub-features, MISCSELECT (a.k.a extended features) +* and max enclave sizes. The SGX sub-features and MISCSELECT +* are restricted by kernel and KVM capabilities (like most +* feature flags), while enclave size is unrestricted. +*/ +cpuid_entry_override (entry,CPUID_12_EAX ); + entry->ebx &=SGX_MISC_EXINFO ; + + entry =do_host_cpuid (array, function,1 ); +if (!entry) + goto out; + +/* +* Index 1: SECS.ATTRIBUTES. ATTRIBUTES are restricted a la +* feature flags. Advertise all supported flags, including +* privileged attributes that require explicit opt-in from +* userspace. ATTRIBUTES.XFRM is not adjusted as userspace is +* expected to derive it from supported XCR0. +*/ + entry->eax &=SGX_ATTR_DEBUG |SGX_ATTR_MODE64BIT | +/* PROVISIONKEY | */ SGX_ATTR_EINITTOKENKEY | +SGX_ATTR_KSS ; + entry->ebx &=0 ; +break ;/* Intel PT */ case 0x14 :if (!kvm_cpu_cap_has (X86_FEATURE_INTEL_PT )) {
arch/x86/kvm/vmx/nested.c +23 -3
@@ -11,6 +11,7 @@#include "mmu.h" #include "nested.h" #include "pmu.h" +#include "sgx.h" #include "trace.h" #include "x86.h"
@@ -2318,6 +2319,9 @@ static void prepare_vmcs02_early(struct vcpu_vmx *vmx, struct vmcs12 *vmcs12)if (!nested_cpu_has2 (vmcs12,SECONDARY_EXEC_UNRESTRICTED_GUEST )) exec_control &= ~SECONDARY_EXEC_UNRESTRICTED_GUEST ; +if (exec_control &SECONDARY_EXEC_ENCLS_EXITING ) +vmx_write_encls_bitmap (&vmx->vcpu, vmcs12); +secondary_exec_controls_set (vmx, exec_control); }
@@ -5698,6 +5702,20 @@ static bool nested_vmx_exit_handled_cr(struct kvm_vcpu *vcpu,return false; } +static bool nested_vmx_exit_handled_encls (struct kvm_vcpu *vcpu, +struct vmcs12 *vmcs12) +{ +u32 encls_leaf; + +if (!nested_cpu_has2 (vmcs12,SECONDARY_EXEC_ENCLS_EXITING )) +return false; + + encls_leaf =kvm_rax_read (vcpu); +if (encls_leaf >62 ) + encls_leaf =63 ; +return vmcs12->encls_exiting_bitmap &BIT_ULL (encls_leaf); +} +static bool nested_vmx_exit_handled_vmcs_access (struct kvm_vcpu *vcpu,struct vmcs12 *vmcs12,gpa_t bitmap) {
@@ -5791,9 +5809,6 @@ static bool nested_vmx_l0_wants_exit(struct kvm_vcpu *vcpu,case EXIT_REASON_VMFUNC :/* VM functions are emulated through L2->L0 vmexits. */ return true; -case EXIT_REASON_ENCLS : -/* SGX is never exposed to L1 */ -return true;default :break ; }
@@ -5917,6 +5932,8 @@ static bool nested_vmx_l1_wants_exit(struct kvm_vcpu *vcpu,case EXIT_REASON_TPAUSE :return nested_cpu_has2 (vmcs12,SECONDARY_EXEC_ENABLE_USR_WAIT_PAUSE ); +case EXIT_REASON_ENCLS : +return nested_vmx_exit_handled_encls (vcpu, vmcs12);default :return true; }
@@ -6489,6 +6506,9 @@ void nested_vmx_setup_ctls_msrs(struct nested_vmx_msrs *msrs, u32 ept_caps) msrs->secondary_ctls_high |=SECONDARY_EXEC_VIRTUALIZE_APIC_ACCESSES ; +if (enable_sgx) + msrs->secondary_ctls_high |=SECONDARY_EXEC_ENCLS_EXITING ; +/* miscellaneous data */ rdmsr(MSR_IA32_VMX_MISC, msrs->misc_low,
arch/x86/kvm/vmx/nested.h +5 -0
@@ -244,6 +244,11 @@ static inline bool nested_exit_on_intr(struct kvm_vcpu *vcpu)PIN_BASED_EXT_INTR_MASK ; } +static inline bool nested_cpu_has_encls_exit (struct vmcs12 *vmcs12) +{ +return nested_cpu_has2 (vmcs12,SECONDARY_EXEC_ENCLS_EXITING ); +} +/* * if fixed0[i] == 1: val[i] must be 1 * if fixed1[i] == 0: val[i] must be 0
arch/x86/kvm/vmx/sgx.c +79 -1
93 lines not on this page · raw message
arch/x86/kvm/vmx/sgx.h +13 -0
@@ -4,6 +4,9 @@#include <linux/kvm_host.h> +#include "capabilities.h" +#include "vmx_ops.h" +#ifdef CONFIG_X86_SGX_VIRTUALIZATION extern bool __read_mostly enable_sgx;
@@ -11,11 +14,21 @@ int handle_encls(struct kvm_vcpu *vcpu);void setup_default_sgx_lepubkeyhash (void );void vcpu_setup_sgx_lepubkeyhash (struct kvm_vcpu *vcpu); + +void vmx_write_encls_bitmap (struct kvm_vcpu *vcpu,struct vmcs12 *vmcs12);#else #define enable_sgx 0static inline void setup_default_sgx_lepubkeyhash (void ) { }static inline void vcpu_setup_sgx_lepubkeyhash (struct kvm_vcpu *vcpu) { } + +static inline void vmx_write_encls_bitmap (struct kvm_vcpu *vcpu, +struct vmcs12 *vmcs12) +{ +/* Nothing to do if hardware doesn't support SGX */ +if (cpu_has_vmx_encls_vmexit ()) +vmcs_write64 (ENCLS_EXITING_BITMAP ,-1ull ); +}#endif #endif /* __KVM_X86_SGX_H */
arch/x86/kvm/vmx/vmcs12.c +1 -0
@@ -50,6 +50,7 @@ const unsigned short vmcs_field_to_offset_table[] = {FIELD64 (VMREAD_BITMAP , vmread_bitmap),FIELD64 (VMWRITE_BITMAP , vmwrite_bitmap),FIELD64 (XSS_EXIT_BITMAP , xss_exit_bitmap), +FIELD64 (ENCLS_EXITING_BITMAP , encls_exiting_bitmap),FIELD64 (GUEST_PHYSICAL_ADDRESS , guest_physical_address),FIELD64 (VMCS_LINK_POINTER , vmcs_link_pointer), FIELD64(GUEST_IA32_DEBUGCTL, guest_ia32_debugctl),
arch/x86/kvm/vmx/vmcs12.h +3 -1
@@ -69,7 +69,8 @@ struct __packed vmcs12 {u64 vm_function_control;u64 eptp_list_address;u64 pml_address; -u64 padding64[3 ];/* room for future expansion */ +u64 encls_exiting_bitmap; +u64 padding64[2 ];/* room for future expansion */ /* * To allow migration of L1 (complete with its L2 guests) between * machines of different natural widths (32 or 64 bit), we cannot have
@@ -256,6 +257,7 @@ static inline void vmx_check_vmcs12_offsets(void)CHECK_OFFSET (vm_function_control,296 );CHECK_OFFSET (eptp_list_address,304 );CHECK_OFFSET (pml_address,312 ); +CHECK_OFFSET (encls_exiting_bitmap,320 );CHECK_OFFSET (cr0_guest_host_mask,344 );CHECK_OFFSET (cr4_guest_host_mask,352 );CHECK_OFFSET (cr0_read_shadow,360 );
arch/x86/kvm/vmx/vmx.c +36 -2
@@ -2161,6 +2161,9 @@ static int vmx_set_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info) vmx->msr_ia32_feature_control = data;if (msr_info->host_initiated && data ==0 )vmx_leave_nested (vcpu); + +/* SGX may be enabled/disabled by guest's firmware */ +vmx_write_encls_bitmap (vcpu,NULL );break ; case MSR_IA32_SGXLEPUBKEYHASH0 ... MSR_IA32_SGXLEPUBKEYHASH3: if (!msr_info->host_initiated &&
@@ -4320,6 +4323,15 @@ static void vmx_compute_secondary_exec_control(struct vcpu_vmx *vmx)vmx_adjust_sec_exec_control (vmx, &exec_control, waitpkg,WAITPKG ,ENABLE_USR_WAIT_PAUSE , false); +if (cpu_has_vmx_encls_vmexit () && nested) { +if (guest_cpuid_has (vcpu,X86_FEATURE_SGX )) + vmx->nested.msrs.secondary_ctls_high |= +SECONDARY_EXEC_ENCLS_EXITING ; +else + vmx->nested.msrs.secondary_ctls_high &= + ~SECONDARY_EXEC_ENCLS_EXITING ; + } + vmx->secondary_exec_control = exec_control; }
@@ -4419,8 +4431,7 @@ static void init_vmcs(struct vcpu_vmx *vmx)vmcs_write16 (GUEST_PML_INDEX ,PML_ENTITY_NUM -1 ); } -if (cpu_has_vmx_encls_vmexit ()) -vmcs_write64 (ENCLS_EXITING_BITMAP ,-1ull ); +vmx_write_encls_bitmap (&vmx->vcpu,NULL );if (vmx_pt_mode_is_host_guest ()) {memset (&vmx->pt_desc,0 ,sizeof (vmx->pt_desc));
@@ -7317,6 +7328,22 @@ static void vmx_vcpu_after_set_cpuid(struct kvm_vcpu *vcpu)set_cr4_guest_host_mask (vmx); +vmx_write_encls_bitmap (vcpu,NULL ); +if (guest_cpuid_has (vcpu,X86_FEATURE_SGX )) + vmx->msr_ia32_feature_control_valid_bits |=FEAT_CTL_SGX_ENABLED ; +else + vmx->msr_ia32_feature_control_valid_bits &= ~FEAT_CTL_SGX_ENABLED ; +/* +* Only allow guest to write its virtual SGX_LEPUBKEYHASHn MSRs when +* host is writable, otherwise it is meaningless. +*/ +if (guest_cpuid_has (vcpu,X86_FEATURE_SGX_LC )) + vmx->msr_ia32_feature_control_valid_bits |= +FEAT_CTL_SGX_LC_ENABLED ; +else + vmx->msr_ia32_feature_control_valid_bits &= + ~FEAT_CTL_SGX_LC_ENABLED ; +/* Refresh #PF interception to account for MAXPHYADDR changes. */ update_exception_bitmap (vcpu); }
@@ -7337,6 +7364,13 @@ static __init void vmx_set_cpu_caps(void)if (vmx_pt_mode_is_host_guest ())kvm_cpu_cap_check_and_set (X86_FEATURE_INTEL_PT ); +if (!enable_sgx) { +kvm_cpu_cap_clear (X86_FEATURE_SGX ); +kvm_cpu_cap_clear (X86_FEATURE_SGX_LC ); +kvm_cpu_cap_clear (X86_FEATURE_SGX1 ); +kvm_cpu_cap_clear (X86_FEATURE_SGX2 ); + } +if (vmx_umip_emulated ())kvm_cpu_cap_set (X86_FEATURE_UMIP );
Signature · 1 line
2.29.2
RFC PATCH 21/23 KVM: VMX: Add ENCLS[EINIT] handler to support SGX Launch Control (LC)
In reply to Kai HuangFrom: Sean Christopherson <[email protected]>
Add a VM-Exit handler to trap-and-execute EINIT when SGX LC is enabled in the host. When SGX LC is enabled, the host kernel may rewrite the hardware values at will, e.g. to launch enclaves with different signers, thus KVM needs to intercept EINIT to ensure it is executed with the correct LE hash (even if the guest sees a hardwired hash).
Switching the LE hash MSRs on VM-Enter/VM-Exit is not a viable option as writing the MSRs is prohibitively expensive, e.g. on SKL hardware each WRMSR is ~400 cycles. And because EINIT takes tens of thousands of cycles to execute, the ~1500 cycle overhead to trap-and-execute EINIT is unlikely to be noticed by the guest, let alone impact its overall SGX performance.
--- arch/x86/kvm/vmx/sgx.c | 55 ++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 55 insertions(+)
arch/x86/kvm/vmx/sgx.c +55 -0
@@ -256,6 +256,59 @@ static int handle_encls_ecreate(struct kvm_vcpu *vcpu)return kvm_skip_emulated_instruction (vcpu); } +static int handle_encls_einit (struct kvm_vcpu *vcpu) +{ +unsigned long sig_hva, secs_hva, token_hva, rflags; +struct vcpu_vmx *vmx =to_vmx (vcpu); +gva_t sig_gva, secs_gva, token_gva; +gpa_t sig_gpa, secs_gpa, token_gpa; +int ret, trapnr; + +if (sgx_get_encls_gva (vcpu,kvm_rbx_read (vcpu),1808 ,4096 , &sig_gva) || +sgx_get_encls_gva (vcpu,kvm_rcx_read (vcpu),4096 ,4096 , &secs_gva) || +sgx_get_encls_gva (vcpu,kvm_rdx_read (vcpu),304 ,512 , &token_gva)) +return 1 ; + +/* +* Translate the SIGSTRUCT, SECS and TOKEN pointers from GVA to GPA. +* Resume the guest on failure to inject a #PF. +*/ +if (sgx_gva_to_gpa (vcpu, sig_gva, false, &sig_gpa) || +sgx_gva_to_gpa (vcpu, secs_gva, true, &secs_gpa) || +sgx_gva_to_gpa (vcpu, token_gva, false, &token_gpa)) +return 1 ; + +/* +* ...and then to HVA. The order of accesses isn't architectural, i.e. +* KVM doesn't have to fully process one address at a time. Exit to +* userspace if a GPA is invalid. Note, all structures are aligned and +* cannot split pages. +*/ +if (sgx_gpa_to_hva (vcpu, sig_gpa, &sig_hva) || +sgx_gpa_to_hva (vcpu, secs_gpa, &secs_hva) || +sgx_gpa_to_hva (vcpu, token_gpa, &token_hva)) +return 0 ; + + ret =sgx_virt_einit ((void __user *)sig_hva, (void __user *)token_hva, + (void __user *)secs_hva, + vmx->msr_ia32_sgxlepubkeyhash, &trapnr); + +if (ret == -EFAULT ) +return sgx_inject_fault (vcpu, secs_gva, trapnr); + + rflags =vmx_get_rflags (vcpu) & ~(X86_EFLAGS_CF |X86_EFLAGS_PF | +X86_EFLAGS_AF |X86_EFLAGS_SF | +X86_EFLAGS_OF ); +if (ret) + rflags |=X86_EFLAGS_ZF ; +else + rflags &= ~X86_EFLAGS_ZF ; +vmx_set_rflags (vcpu, rflags); + +kvm_rax_write (vcpu, ret); +return kvm_skip_emulated_instruction (vcpu); +} +static inline bool encls_leaf_enabled_in_guest (struct kvm_vcpu *vcpu,u32 leaf) {if (!enable_sgx || !guest_cpuid_has (vcpu,X86_FEATURE_SGX ))
@@ -288,6 +341,8 @@ int handle_encls(struct kvm_vcpu *vcpu) }else {if (leaf ==ECREATE )return handle_encls_ecreate (vcpu); +if (leaf ==EINIT ) +return handle_encls_einit (vcpu);WARN (1 ,"KVM: unexpected exit on ENCLS[%u]" , leaf); vcpu->run->exit_reason =KVM_EXIT_UNKNOWN ; vcpu->run->hw.hardware_exit_reason =EXIT_REASON_ENCLS ;
Signature · 1 line
2.29.2
RFC PATCH 23/23 KVM: x86: Add capability to grant VM access to privileged SGX attribute
In reply to Kai HuangFrom: Sean Christopherson <[email protected]>
Add a capability, KVM_CAP_SGX_ATTRIBUTE, that can be used by userspace to grant a VM access to a priveleged attribute, with args[0] holding a file handle to a valid SGX attribute file.
The SGX subsystem restricts access to a subset of enclave attributes to provide additional security for an uncompromised kernel, e.g. to prevent malware from using the PROVISIONKEY to ensure its nodes are running inside a geniune SGX enclave and/or to obtain a stable fingerprint.
To prevent userspace from circumventing such restrictions by running an enclave in a VM, KVM restricts guest access to privileged attributes by default.
--- Documentation/virt/kvm/api.rst | 23 +++++++++++++++++++++++ arch/x86/kvm/cpuid.c | 2 +- arch/x86/kvm/x86.c | 22 ++++++++++++++++++++++ include/uapi/linux/kvm.h | 1 + 4 files changed, 47 insertions(+), 1 deletion(-)
Documentation/virt/kvm/api.rst +23 -0
@@ -6014,6 +6014,29 @@ KVM_EXIT_X86_RDMSR and KVM_EXIT_X86_WRMSR exit notifications which user space can then handle to implement model specific MSR handling and/or user notifications to inform a user that an MSR was not handled. +7.22 KVM_CAP_SGX_ATTRIBUTE +---------------------- + +:Architectures: x86 +:Target: VM +:Parameters: args[0] is a file handle of a SGX attribute file in securityfs +:Returns: 0 on success, -EINVAL if the file handle is invalid or if a requested + attribute is not supported by KVM. + +KVM_CAP_SGX_ATTRIBUTE enables a userspace VMM to grant a VM access to one or +more priveleged enclave attributes. args[0] must hold a file handle to a valid +SGX attribute file corresponding to an attribute that is supported/restricted +by KVM (currently only PROVISIONKEY). + +The SGX subsystem restricts access to a subset of enclave attributes to provide +additional security for an uncompromised kernel, e.g. use of the PROVISIONKEY +is restricted to deter malware from using the PROVISIONKEY to obtain a stable +system fingerprint. To prevent userspace from circumventing such restrictions +by running an enclave in a VM, KVM prevents access to privileged attributes by +default. + +See Documentation/x86/sgx/2.Kernel-internals.rst for more details. + 8. Other capabilities. ======================
arch/x86/kvm/cpuid.c +1 -1
@@ -811,7 +811,7 @@ static inline int __do_cpuid_func(struct kvm_cpuid_array *array, u32 function)* expected to derive it from supported XCR0. */ entry->eax &=SGX_ATTR_DEBUG |SGX_ATTR_MODE64BIT | -/* PROVISIONKEY | */ SGX_ATTR_EINITTOKENKEY | +SGX_ATTR_PROVISIONKEY |SGX_ATTR_EINITTOKENKEY |SGX_ATTR_KSS ; entry->ebx &=0 ;break ;
arch/x86/kvm/x86.c +22 -0
@@ -74,6 +74,8 @@#include <asm/tlbflush.h> #include <asm/intel_pt.h> #include <asm/emulate_prefix.h> +#include <asm/sgx.h> +#include <asm/sgx_arch.h> #include <clocksource/hyperv_timer.h> #define CREATE_TRACE_POINTS
@@ -3739,6 +3741,9 @@ int kvm_vm_ioctl_check_extension(struct kvm *kvm, long ext)case KVM_CAP_X86_USER_SPACE_MSR :case KVM_CAP_X86_MSR_FILTER :case KVM_CAP_ENFORCE_PV_FEATURE_CPUID : +#ifdef CONFIG_X86_SGX_VIRTUALIZATION +case KVM_CAP_SGX_ATTRIBUTE : +#endif r =1 ;break ;case KVM_CAP_SYNC_REGS :
@@ -5270,6 +5275,23 @@ int kvm_vm_ioctl_enable_cap(struct kvm *kvm, kvm->arch.user_space_msr_mask = cap->args[0 ]; r =0 ;break ; +#ifdef CONFIG_X86_SGX_VIRTUALIZATION +case KVM_CAP_SGX_ATTRIBUTE : { +unsigned long allowed_attributes =0 ; + + r =sgx_set_attribute (&allowed_attributes, cap->args[0 ]); +if (r) +break ; + +/* KVM only supports the PROVISIONKEY privileged attribute. */ +if ((allowed_attributes &SGX_ATTR_PROVISIONKEY ) && + !(allowed_attributes & ~SGX_ATTR_PROVISIONKEY )) + kvm->arch.sgx_provisioning_allowed = true; +else + r = -EINVAL ; +break ; + } +#endif default : r = -EINVAL ;break ;
include/uapi/linux/kvm.h +1 -0
@@ -1053,6 +1053,7 @@ struct kvm_ppc_resize_hpt {#define KVM_CAP_X86_USER_SPACE_MSR 188#define KVM_CAP_X86_MSR_FILTER 189#define KVM_CAP_ENFORCE_PV_FEATURE_CPUID 190 +#define KVM_CAP_SGX_ATTRIBUTE 200#ifdef KVM_CAP_IRQ_ROUTING
Signature · 1 line
2.29.2
Sorry that I made mistake when copy and paste Jim's email address :( Remove the wrong email address ([email protected]) and add the correct one ([email protected]). Really apologize for the noise.
Thanks,
-Kai
Quoted 188 lines · Kai Huang
--- Disclaimer ---
These patches were originally written by Sean Christopherson while at Intel. Now that Sean has left Intel, I (Kai) have taken over getting them upstream. This series needs more review before it can be merged. It is being posted publicly and under RFC so Sean and others can review it. Maintainers are safe ignoring it for now.
------------------
Hi all,
This series adds KVM SGX virtualization support. The first 12 patches starting with x86/sgx or x86/cpu.. are necessary changes to x86 and SGX core/driver to support KVM SGX virtualization, while the rest are patches to KVM subsystem.
Please help to review this series. Also I'd like to hear what is the proper way to merge this series, since it contains change to both x86/SGX and KVM subsystem. Any feedback is highly appreciated. And please let me know if I forgot to CC anyone, or anyone wants to be removed from CC. Thanks in advance!
This series is based against latest tip tree's x86/sgx branch. You can also get the code from tip branch of kvm-sgx repo on github:
https://github.com/intel/kvm-sgx.git tip
It also requires Qemu changes to create VM with SGX support. You can find Qemu repo here:
https://github.com/intel/qemu-sgx.git next
Please refer to README.md of above qemu-sgx repo for detail on how to create guest with SGX support. At meantime, for your quick reference you can use below command to create SGX guest:
#qemu-system-x86_64 -smp 4 -m 2G -drive file=<your_vm_image>,if=virtio \ -cpu host,+sgx_provisionkey \ -sgx-epc id=epc1,memdev=mem1 \ -object memory-backend-epc,id=mem1,size=64M,preallocPlease note that the SGX relevant part is:
-cpu host,+sgx_provisionkey \ -sgx-epc id=epc1,memdev=mem1 \ -object memory-backend-epc,id=mem1,size=64M,preallocAnd you can change other parameters of your qemu command based on your needs.
=========
KVM SGX virtualization Overview- Virtual EPC
"Virtual EPC" is the EPC section exposed by KVM to guest so SGX software in guest can discover it and use it to create SGX enclaves. KVM exposes SGX to guest via CPUID, and exposes one or more "virtual EPC" sections for guest. The size of "virtual EPC" is passed as Qemu parameter when creating the guest, and the base address is calcualted internally according to guest's configuration.
To support virtual EPC, add a new misc device
/dev/sgx_virt_epcto SGX core/driver to allow userspace (Qemu) to allocate "raw" EPC, and use it as "virtual EPC" for guest. Obviously, unlike EPC allocated for host SGX driver, virtual EPC allocated via/dev/sgx_virt_epcdoesn't have enclave associated, and how virtual EPC is used by guest is compeletely controlled by guest's SGX software.Implement the "raw" EPC allocation in the x86 core-SGX subsystem via
/dev/sgx_virt_epcrather than in KVM. Doing so has two major advantages:- Does not require changes to KVM's uAPI, e.g. EPC gets handled as just another memory backend for guests.
- EPC management is wholly contained in the SGX subsystem, e.g. SGX does not have to export any symbols, changes to reclaim flows don't need to be routed through KVM, SGX's dirty laundry doesn't have to get aired out for the world to see, and so on and so forth.
The virtual EPC allocated to guests is currently not reclaimable, due to reclaiming EPC from KVM guests is not currently supported. Due to the complications of handling reclaim conflicts between guest and host, KVM EPC oversubscription, which allows total virtual EPC size greater than physical EPC by being able to reclaiming guests' EPC, is significantly more complex than basic support for SGX virtualization.
- Support SGX virtualization without SGX Launch Control unlocked mode
Although SGX driver requires SGX Launch Control unlocked mode to work, SGX virtualization doesn't, since how enclave is created is completely controlled by guest SGX software, which is not necessarily linux. Therefore, this series allows KVM to expose SGX to guest even SGX Launch Control is in locked mode, or is not present at all. The reason is the goal of SGX virtualization, or virtualization in general, is to expose hardware feature to guest, but not to make assumption how guest will use it. Therefore, KVM should support SGX guest as long as hardware is able to, to have chance to support more potential use cases in cloud environment.
- Support exposing SGX2
Due to the same reason above, SGX2 feature detection is added to core SGX code to allow KVM to expose SGX2 to guest, even currently SGX driver doesn't support SGX2, because SGX2 can work just fine in guest w/o any interaction to host SGX driver.
- Restricit SGX guest access to provisioning key
To grant guest being able to fully use SGX, guest needs to be able to create provisioning enclave. However provisioning key is sensitive and is restricted by
/dev/sgx_provisionin host SGX driver, therefore KVM SGX virtualization follows the same role: a newKVM_CAP_SGX_ATTRIBUTEis added to KVM uAPI, and only file descriptor of/dev/sgx_provisionis passed to that CAP by usersppace hypervisor (Qemu) when creating the guest, it can access provisioning bit. This is done by making KVM trape ECREATE instruction from guest, and check the provisioning bit in ECREATE's attribute.Kai Huang (1):
x86/sgx: Add helper to updateSGX_LEPUBKEYHASHnMSRsSean Christopherson (22):
x86/sgx: Split out adding EPC page to free list to separate helper x86/sgx: Add enum forSGX_CHILD_PRESENTerror code x86/sgx: Introduce virtual EPC for use by KVM guests x86/cpufeatures: Add SGX1 and SGX2 sub-features x86/cpu/intel: Allow SGX virtualization without Launch Control support x86/sgx: Expose SGX architectural definitions to the kernel x86/sgx: Move ENCLS leaf definitions tosgx_arch.hx86/sgx: Add SGX2 ENCLS leaf definitions (EAUG, EMODPR and EMODT) x86/sgx: Addencls_faulted()helper
x86/sgx: Add helpers to expose ECREATE and EINIT to KVM x86/sgx: Move provisioning device creation out of SGX driver
KVM: VMX: Convert vcpu_vmx.exit_reason to a union
KVM: x86: Export kvm_mmu_gva_to_gpa_{read,write}() for SGX (VMX)
KVM: x86: Define new #PF SGX error code bit
KVM: x86: Add SGX feature leaf to reverse CPUID lookup
KVM: VMX: Add basic handling of VM-Exit from SGX enclave
KVM: VMX: Frame in ENCLS handler for SGX virtualization
KVM: VMX: Add SGX ENCLS[ECREATE] handler to enforce CPUID restrictions
KVM: VMX: Add emulation of SGX Launch Control LE hash MSRs
KVM: VMX: Add ENCLS[EINIT] handler to support SGX Launch Control (LC)
KVM: VMX: Enable SGX virtualization for SGX1, SGX2 and LC
KVM: x86: Add capability to grant VM access to privileged SGX attributeDocumentation/virt/kvm/api.rst | 23 + arch/x86/Kconfig | 12 + arch/x86/include/asm/cpufeature.h | 5 +- arch/x86/include/asm/cpufeatures.h | 6 +- arch/x86/include/asm/disabled-features.h | 7 +- arch/x86/include/asm/kvm_host.h | 5 + arch/x86/include/asm/required-features.h | 2 +- arch/x86/include/asm/sgx.h | 19 + .../cpu/sgx/arch.h => include/asm/sgx_arch.h} | 20 + arch/x86/include/asm/vmx.h | 1 + arch/x86/include/uapi/asm/vmx.h | 1 + arch/x86/kernel/cpu/common.c | 4 + arch/x86/kernel/cpu/feat_ctl.c | 50 +- arch/x86/kernel/cpu/sgx/Makefile | 1 + arch/x86/kernel/cpu/sgx/driver.c | 17 - arch/x86/kernel/cpu/sgx/encl.c | 2 +- arch/x86/kernel/cpu/sgx/encls.h | 29 +- arch/x86/kernel/cpu/sgx/ioctl.c | 23 +- arch/x86/kernel/cpu/sgx/main.c | 79 ++- arch/x86/kernel/cpu/sgx/sgx.h | 5 +- arch/x86/kernel/cpu/sgx/virt.c | 318 ++++++++++++ arch/x86/kernel/cpu/sgx/virt.h | 14 + arch/x86/kvm/Makefile | 2 + arch/x86/kvm/cpuid.c | 58 ++- arch/x86/kvm/cpuid.h | 1 + arch/x86/kvm/vmx/nested.c | 70 ++- arch/x86/kvm/vmx/nested.h | 5 + arch/x86/kvm/vmx/sgx.c | 462 ++++++++++++++++++ arch/x86/kvm/vmx/sgx.h | 34 ++ arch/x86/kvm/vmx/vmcs12.c | 1 + arch/x86/kvm/vmx/vmcs12.h | 4 +- arch/x86/kvm/vmx/vmx.c | 171 +++++-- arch/x86/kvm/vmx/vmx.h | 27 +- arch/x86/kvm/x86.c | 24 + include/uapi/linux/kvm.h | 1 + tools/testing/selftests/sgx/defines.h | 2 +- 36 files changed, 1366 insertions(+), 139 deletions(-)create mode 100644
arch/x86/include/asm/sgx.h
renamearch/x86/{kernel/cpu/sgx/arch.h=>include/asm/sgx_arch.h} (96%) create mode 100644arch/x86/kernel/cpu/sgx/virt.c
create mode 100644arch/x86/kernel/cpu/sgx/virt.h
create mode 100644arch/x86/kvm/vmx/sgx.c
create mode 100644arch/x86/kvm/vmx/sgx.h
Quoted 7 lines · Kai Huang
- Virtual EPC
"Virtual EPC" is the EPC section exposed by KVM to guest so SGX software in guest can discover it and use it to create SGX enclaves. KVM exposes SGX to guest via CPUID, and exposes one or more "virtual EPC" sections for guest. The size of "virtual EPC" is passed as Qemu parameter when creating the guest, and the base address is calcualted internally according to guest's
^ calculated
Quoted 1 line
configuration.
This is not a great first paragraph to introduce me to this feature.
Please remind us what EPC *is*, then you can go and talk about why we have to virtualize it, and how "virtual EPC" is different from normal EPC. For instance:
SGX enclave memory is special and is reserved specifically for enclave use. In bare-metal SGX enclaves, the kernel allocates enclave pages, copies data into the pages with privileged instructions, then allows the enclave to start. In this scenario, only initialized pages already assigned to an enclave are mapped to userspace.
In virtualized environments, the hypervisor still needs to do the physical enclave page allocation. The guest kernel is responsible for the data copying (among other things). This means that the job of starting an enclave is now split between hypervisor and guest.
This series introduces a new misc device: /dev/sgx_virt_epc. This device allows the host to map *uninitialized* enclave memory into userspace, which can then be passed into a guest.
While it might be *possible* to start a host-side enclave with /dev/sgx_enclave and pass its memory into a guest, it would be wasteful and convoluted.
Quoted 4 lines
core/driver to allow userspace (Qemu) to allocate "raw" EPC, and use it as "virtual EPC" for guest. Obviously, unlike EPC allocated for host SGX driver, virtual EPC allocated via
/dev/sgx_virt_epcdoesn't have enclave associated, and how virtual EPC is used by guest is compeletely controlled by guest's SGX
^ completely
Please run a spell checker on this thing.
Quoted 13 lines
software.
Implement the "raw" EPC allocation in the x86 core-SGX subsystem via
/dev/sgx_virt_epcrather than in KVM. Doing so has two major advantages:- Does not require changes to KVM's uAPI, e.g. EPC gets handled as just another memory backend for guests.
- EPC management is wholly contained in the SGX subsystem, e.g. SGX does not have to export any symbols, changes to reclaim flows don't need to be routed through KVM, SGX's dirty laundry doesn't have to get aired out for the world to see, and so on and so forth.
Quoted 6 lines
The virtual EPC allocated to guests is currently not reclaimable, due to reclaiming EPC from KVM guests is not currently supported. Due to the complications of handling reclaim conflicts between guest and host, KVM EPC oversubscription, which allows total virtual EPC size greater than physical EPC by being able to reclaiming guests' EPC, is significantly more complex than basic support for SGX virtualization.
It would also help here to remind the reader that enclave pages have a special reclaim mechanism separtae from normal page reclaim, and that mechanism is disabled for these pages.
Does the *ABI* here preclude doing oversubscription in the future?
Quoted 3 lines
- Support SGX virtualization without SGX Launch Control unlocked mode
Although SGX driver requires SGX Launch Control unlocked mode to work, SGX
Although the bare-metal SGX driver requires...
Also, didn't we call this "Flexible Launch Control"?
Quoted 3 lines
virtualization doesn't, since how enclave is created is completely controlled by guest SGX software, which is not necessarily linux. Therefore, this series allows KVM to expose SGX to guest even SGX Launch Control is in locked mode,
... "expose SGX to guests even if" ...
Quoted 5 lines
or is not present at all. The reason is the goal of SGX virtualization, or virtualization in general, is to expose hardware feature to guest, but not to make assumption how guest will use it. Therefore, KVM should support SGX guest as long as hardware is able to, to have chance to support more potential use cases in cloud environment.
This is kinda long-winded and misses a lot of important context. How about:
SGX hardware supports two "launch control" modes to limit which enclaves can run. In the "locked" mode, the hardware prevents enclaves from running unless they are blessed by a third party. In the unlocked mode, the kernel is in full control of which enclaves can run. The bare-metal SGX code refuses to launch enclaves unless it is in the unlocked mode.
This sgx_virt_epc driver does not have such a restriction. This allows guests which are OK with the locked mode to use SGX, even if the host kernel refuses to.
Quoted 11 lines
- Support exposing SGX2
Due to the same reason above, SGX2 feature detection is added to core SGX code to allow KVM to expose SGX2 to guest, even currently SGX driver doesn't support SGX2, because SGX2 can work just fine in guest w/o any interaction to host SGX driver.
- Restricit SGX guest access to provisioning key
To grant guest being able to fully use SGX, guest needs to be able to create provisioning enclave.
"enclave" or "enclaves"?
Quoted 1 line
However provisioning key is sensitive and is restricted by
^ the
Quoted 5 lines
/dev/sgx_provisionin host SGX driver, therefore KVM SGX virtualization follows the same role: a newKVM_CAP_SGX_ATTRIBUTEis added to KVM uAPI, and only file descriptor of/dev/sgx_provisionis passed to that CAP by usersppace hypervisor (Qemu) when creating the guest, it can access provisioning bit. This is done by making KVM trape ECREATE instruction from guest, and check the provisioning bit
^ trap
Quoted 1 line
in ECREATE's attribute.
The grammar in that paragraph is really off to me. Can you give it another go?
Re: [RFC PATCH 02/23] x86/sgx: Add enum for SGX_CHILD_PRESENT error code
In reply to Kai HuangQuoted 2 lines · Kai Huang
Add
SGX_CHILD_PRESENTfor use by SGX virtualization to assert EREMOVE failures are expected, but only due toSGX_CHILD_PRESENT.
This dances around the fact that this is an architectural error-code. Could that be explicit? Maybe the subject should be:
Add SGX_CHILD_PRESENT hardware error code
Re: [RFC PATCH 03/23] x86/sgx: Introduce virtual EPC for use by KVM guests
In reply to Kai HuangQuoted 11 lines · Kai Huang
From: Sean Christopherson <[email protected]>
Add a misc device
/dev/sgx_virt_epcto allow userspace to allocate "raw" EPC without an associated enclave. The intended and only known use case for raw EPC allocation is to expose EPC to a KVM guest, hence thevirt_epcmoniker, virt.{c,h} files andX86_SGX_VIRTUALIZATIONKconfig.Modify
sgx_init()to always try to initialize virtual EPC driver, even when SGX driver is disabled due to SGX Launch Control is in locked mode, or not present at all, since SGX virtualization allows to expose SGX to guests that support non-LC configurations.
The grammar here is a bit off. Here's a rewrite:
Modify sgx_init() to always try to initialize the virtual EPC driver, even if the bare-metal SGX driver is disabled. The bare-metal driver might be disabled if SGX Launch Control is in locked mode, or not supported in the hardware at all. This allows (non-Linux) guests that support non-LC configurations to use SGX.
Quoted 26 lines
arch/x86/kernel/cpu/sgx/Makefile +1
-0@@ -3,3 +3,4 @@ obj-y += \ encl.o \ ioctl.o \ main.o +obj-$(CONFIG_X86_SGX_VIRTUALIZATION) += virt.oarch/x86/kernel/cpu/sgx/main.c +2
-0@@ -9,9 +9,11 @@#include <linux/sched/mm.h> #include <linux/sched/signal.h> #include <linux/slab.h> +#include "arch.h" #include "driver.h" #include "encl.h" #include "encls.h" +#include "virt.h" struct sgx_epc_section sgx_epc_sections[SGX_MAX_EPC_SECTIONS ];static int sgx_nr_epc_sections;
arch/x86/kernel/cpu/sgx/main.c static void __init sgx_init(void) · line 734
Quoted 8 lines
arch/x86/kernel/cpu/sgx/main.c +2
-1@@ -726,7 +728,8 @@ static void __init sgx_init(void)if (!sgx_page_reclaimer_init ()) goto err_page_cache; - ret =sgx_drv_init (); +/* Success if the native *or* virtual EPC driver initialized cleanly. */ + ret = !!sgx_drv_init () & !!sgx_virt_epc_init ();if (ret) goto err_kthread;
FWIW, I hate that conditional. But, I tried to write to to be something more sane and failed.
Quoted 21 lines
arch/x86/kernel/cpu/sgx/virt.c new file +15
-0@@ -0,0 +1,263 @@ +// SPDX-License-Identifier: GPL-2.0 +/* Copyright(c) 2016-20 Intel Corporation. */ + +#include <linux/miscdevice.h> +#include <linux/mm.h> +#include <linux/mman.h> +#include <linux/sched/mm.h> +#include <linux/sched/signal.h> +#include <linux/slab.h> +#include <linux/xarray.h> +#include <asm/sgx.h> +#include <uapi/asm/sgx.h> + +#include "encls.h" +#include "sgx.h"
arch/x86/kernel/cpu/sgx/virt.c line 25
Quoted 10 lines
arch/x86/kernel/cpu/sgx/virt.c new file +10
-0+#include "virt.h" + +struct sgx_virt_epc { +struct xarray page_array; +struct mutex lock; +struct mm_struct *mm; +}; + +static struct mutex virt_epc_lock; +static struct list_head virt_epc_zombie_pages;
What does the lock protect?
What are zombie pages?
BTW, if zombies are SECS-only, shouldn't that be in the name rather than "epc"?
arch/x86/kernel/cpu/sgx/virt.c line 34
Quoted 8 lines
arch/x86/kernel/cpu/sgx/virt.c new file +8
-0+static int __sgx_virt_epc_fault (struct sgx_virt_epc *epc, +struct vm_area_struct *vma,unsigned long addr) +{ +struct sgx_epc_page *epc_page; +unsigned long index, pfn; +int ret; + +/* epc->lock must already have been hold */
/* epc->lock must already be held */
Wouldn't this be better as:
WARN_ON(!mutex_is_locked(&epc->lock));
?
Quoted 60 lines
arch/x86/kernel/cpu/sgx/virt.c new file +60
-0+/* Calculate index of EPC page in virtual EPC's page_array */ + index = vma->vm_pgoff +PFN_DOWN (addr - vma->vm_start); + + epc_page =xa_load (&epc->page_array, index); +if (epc_page) +return 0 ; + + epc_page =sgx_alloc_epc_page (epc, false); +if (IS_ERR (epc_page)) +return PTR_ERR (epc_page); + + ret =xa_err (xa_store (&epc->page_array, index, epc_page,GFP_KERNEL )); +if (ret) + goto err_free; + + pfn =PFN_DOWN (sgx_get_epc_phys_addr (epc_page)); + + ret =vmf_insert_pfn (vma, addr, pfn); +if (ret !=VM_FAULT_NOPAGE ) { + ret = -EFAULT ; + goto err_delete; + } + +return 0 ; + +err_delete: +xa_erase (&epc->page_array, index); +err_free: +sgx_free_epc_page (epc_page); +return ret; +} + +static vm_fault_t sgx_virt_epc_fault (struct vm_fault *vmf) +{ +struct vm_area_struct *vma = vmf->vma; +struct sgx_virt_epc *epc = vma->vm_private_data; +int ret; + +mutex_lock (&epc->lock); + ret =__sgx_virt_epc_fault (epc, vma, vmf->address); +mutex_unlock (&epc->lock); + +if (!ret) +return VM_FAULT_NOPAGE ; + +if (ret == -EBUSY && (vmf->flags &FAULT_FLAG_ALLOW_RETRY )) { +mmap_read_unlock (vma->vm_mm); +return VM_FAULT_RETRY ; + } + +return VM_FAULT_SIGBUS ; +} + +const struct vm_operations_struct sgx_virt_epc_vm_ops = { + .fault = sgx_virt_epc_fault, +}; + +static int sgx_virt_epc_mmap (struct file *file,struct vm_area_struct *vma) +{ +struct sgx_virt_epc *epc = file->private_data;
arch/x86/kernel/cpu/sgx/virt.c line 105
Quoted 10 lines
arch/x86/kernel/cpu/sgx/virt.c new file +10
-0+ +if (!(vma->vm_flags &VM_SHARED )) +return -EINVAL ; + +/* +* Don't allow mmap() from child after fork(), since child and parent +* cannot map to the same EPC. +*/ +if (vma->vm_mm != epc->mm) +return -EINVAL ;
I mentioned this below, but I'm not buying this logic. I know it would be *bad*, but I don't see why the kernel needs to keep it from happening.
Quoted 4 lines
arch/x86/kernel/cpu/sgx/virt.c new file +4
-0+ vma->vm_ops = &sgx_virt_epc_vm_ops; +/* Don't copy VMA in fork() */ + vma->vm_flags |=VM_PFNMAP |VM_IO |VM_DONTDUMP |VM_DONTCOPY ; + vma->vm_private_data = file->private_data;
arch/x86/kernel/cpu/sgx/virt.c line 120
Quoted 10 lines
arch/x86/kernel/cpu/sgx/virt.c new file +10
-0+ +return 0 ; +} + +static int sgx_virt_epc_free_page (struct sgx_epc_page *epc_page) +{ +int ret; + +if (!epc_page) +return 0 ;
I always worry about these. Why is passing NULL around OK?
arch/x86/kernel/cpu/sgx/virt.c line 126
Quoted 5 lines
arch/x86/kernel/cpu/sgx/virt.c new file +5
-0+/* +* Explicitly EREMOVE virtual EPC page. Virtual EPC is only used by +* guest, and in normal condition guest should have done EREMOVE for +* all EPC pages before they are freed here. But it's possible guest +* is killed or crashed unnormally in which case EREMOVE has not been
"abnormally"
I don't think "unnormally" is a word. Also, this isn't just about crashing or being killed. The guest could simply have a bug.
arch/x86/kernel/cpu/sgx/virt.c line 131
Quoted 5 lines
arch/x86/kernel/cpu/sgx/virt.c new file +5
-0+* done. Do EREMOVE unconditionally here to cover both cases, because +* it's not possible to tell whether guest has done EREMOVE, since +* virtual EPC page status is not tracked. And it is fine to EREMOVE +* EPC page multiple times. +*/
Surprise! I dislike this comment.
/*
* Take a previously guest-owned EPC page and return it to the
* general EPC page pool.
*
* Guests can not be trusted to have left this page in a good
* state, so run EREMOVE on the page unconditionally. In the
* case that a guest properly EREMOVE'd this page, a
* superfluous EREMOVE is harmless.
*/
arch/x86/kernel/cpu/sgx/virt.c line 143
Quoted 12 lines
arch/x86/kernel/cpu/sgx/virt.c new file +12
-0+ ret =__eremove (sgx_get_epc_virt_addr (epc_page)); +if (ret) { +/* +* Only SGX_CHILD_PRESENT is expected, which is because of +* EREMOVE-ing an SECS still with child, in which case it can +* be handled by EREMOVE-ing the SECS again after all pages in +* virtual EPC have been EREMOVE-ed. See comments in below in +* sgx_virt_epc_release(). +*/ +WARN_ON_ONCE (ret !=SGX_CHILD_PRESENT ); +return ret; + }
I find myself wondering what errors could cause the WARN_ON_ONCE() to be hit. The SDM indicates that it's only:
SGX_ENCLAVE_ACT If there are still logical processors executing
inside the enclave.
Should that be mentioned in the comment?
arch/x86/kernel/cpu/sgx/virt.c line 151
Quoted 8 lines
arch/x86/kernel/cpu/sgx/virt.c new file +8
-0+ +__sgx_free_epc_page (epc_page); +return 0 ; +} + +static int sgx_virt_epc_release (struct inode *inode,struct file *file) +{ +struct sgx_virt_epc *epc = file->private_data;
FWIW, I hate the "struct sgx_virt_epc *epc" name. "epc" here is really an instance
Quoted 4 lines
arch/x86/kernel/cpu/sgx/virt.c new file +4
-0+struct sgx_epc_page *epc_page, *tmp, *entry; +unsigned long index; + +LIST_HEAD (secs_pages);
arch/x86/kernel/cpu/sgx/virt.c line 165
Quoted 10 lines
arch/x86/kernel/cpu/sgx/virt.c new file +10
-0+ +mmdrop (epc->mm); + +xa_for_each (&epc->page_array, index, entry) { +/* +* Virtual EPC pages are not tracked, so it's possible for +* EREMOVE to fail due to, e.g. a SECS page still has children +* if guest was shutdown unexpectedly. If it is the case, leave +* it in the xarray and retry EREMOVE below later. +*/
I don't know what it is about the comments, but I cringe every time I see an "i.e." or "e.g.".
I'd rewrite the comment as:
/*
* Remove all normal, child pages. sgx_virt_epc_free_page()
* will fail if EREMOVE fails, but this is OK and expected on
* SECS pages. Those can only be EREMOVE'd *after* all their
* child pages. Retries below will clean them up. */
arch/x86/kernel/cpu/sgx/virt.c line 177
Quoted 12 lines
arch/x86/kernel/cpu/sgx/virt.c new file +12
-0+if (sgx_virt_epc_free_page (entry)) +continue ; + +xa_erase (&epc->page_array, index); + } + +/* +* Retry all failed pages after iterating through the entire tree, at +* which point all children should be removed and the SECS pages can be +* nuked as well...unless userspace has exposed multiple instance of +* virtual EPC to a single VM. +*/
I'm just a comment grouch today I guess. That's a horrible run-on sentence. Let's just state the goal of the loop in the comment above it:
Retry EREMOVE'ing pages. This will clean up any SECS pages that only had children in this 'epc' area.
arch/x86/kernel/cpu/sgx/virt.c line 175
Quoted 2 lines
arch/x86/kernel/cpu/sgx/virt.c new file +2
-0+xa_for_each (&epc->page_array, index, entry) { + epc_page = entry;
Then, talk about the error condition here:
arch/x86/kernel/cpu/sgx/virt.c line 186
Quoted 7 lines
arch/x86/kernel/cpu/sgx/virt.c new file +7
-0+/* +* Error here means that EREMOVE failed due to a SECS page +* still has child on *another* EPC instance. Put it to a +* temporary SECS list which will be spliced to 'zombie page +* list' and will be EREMOVE-ed again when freeing another +* virtual EPC instance. +*/
Surprise, I've got another rewrite:
/*
* An EREMOVE failure here means that the SECS page
* still has children. But, since all children in this
* 'sgx_virt_epc' have been removed, the SECS page must
* have a child on another instance.
*/
arch/x86/kernel/cpu/sgx/virt.c line 184
Quoted 2 lines
arch/x86/kernel/cpu/sgx/virt.c new file +2
-0+if (sgx_virt_epc_free_page (epc_page)) +list_add_tail (&epc_page->list, &secs_pages);
Why move these over to &secs_list here? I think it's to avoid another xa_for_each() below, but it's not clear.
arch/x86/kernel/cpu/sgx/virt.c
Quoted 5 lines
arch/x86/kernel/cpu/sgx/virt.c new file +5
-0+xa_erase (&epc->page_array, index); + } + +/* +* Third time's a charm.
This is confusing. This section is *NOT* retrying a third time. This is a cute comment, but it's actually, logically different from the two tries above. I say remove it. In fact, I'd even concentrate the comment here to explain that this is a logically *TOALLY* disconnected from what happened above.
arch/x86/kernel/cpu/sgx/virt.c line 197
Quoted 4 lines
Try to EREMOVE zombie SECS pages from virtual
arch/x86/kernel/cpu/sgx/virt.c new file +3
-0+* EPC instances that were previously released, i.e. free SECS pages +* that were in limbo due to having children in *this* EPC instance. +*/
This is as close as this code gets to telling me what a zombie page is.
I don't think it gets close enough, or does it in the right spot.
I think it probably needs explicit discussion in the changelog. I think Sean explained this to me once, but I've forgotten by now. The code needs to be understandable without getting Sean on the phone anyway. :)
I'd probably just say:
/*
* SECS pages are "pinned" by child pages, an unpinned once all
* children have been EREMOVE'd. A child page in this instance
* may have pinned an SECS page encountered in an earlier
* release(), creating a zombie. Since some children were
* EREMOVE'd above, try to EREMOVE all zombies in the hopes that
* one was unpinned.
*/
arch/x86/kernel/cpu/sgx/virt.c line 209
Quoted 12 lines
arch/x86/kernel/cpu/sgx/virt.c new file +12
-0+mutex_lock (&virt_epc_lock); +list_for_each_entry_safe (epc_page, tmp, &virt_epc_zombie_pages, list) { +/* +* Speculatively remove the page from the list of zombies, if +* the page is successfully EREMOVE it will be added to the +* list of free pages. If EREMOVE fails, throw the page on the +* local list, which will be spliced on at the end. +*/ +list_del (&epc_page->list); + +if (sgx_virt_epc_free_page (epc_page)) +list_add_tail (&epc_page->list, &secs_pages);
I don't get this. Couldn't you do without the unconditional list_del() and instead just do:
if (!sgx_virt_epc_free_page(epc_page)) list_del(&epc_page->list);
Or does the free() code clobber the list_head? If that's the case, maybe you should say that explicitly.
Quoted 22 lines
arch/x86/kernel/cpu/sgx/virt.c new file +22
-0+ } + +if (!list_empty (&secs_pages)) +list_splice_tail (&secs_pages, &virt_epc_zombie_pages); +mutex_unlock (&virt_epc_lock); + +kfree (epc); + +return 0 ; +} + +static int sgx_virt_epc_open (struct inode *inode,struct file *file) +{ +struct sgx_virt_epc *epc; + + epc =kzalloc (sizeof (struct sgx_virt_epc ),GFP_KERNEL ); +if (!epc) +return -ENOMEM ; +/* +* Keep the current->mm to virtual EPC. It will be checked in +* sgx_virt_epc_mmap() to prevent, in case of fork, child being +* able to mmap() to the same virtual EPC pages.
arch/x86/kernel/cpu/sgx/virt.c line 241
Quoted 10 lines
arch/x86/kernel/cpu/sgx/virt.c new file +10
-0+*/ +mmgrab (current->mm); + epc->mm = current->mm; +mutex_init (&epc->lock); +xa_init (&epc->page_array); + + file->private_data = epc; + +return 0 ; +}
I understand why this made sense for regular enclaves, but I'm having a harder time here. If you mmap(fd, MAP_SHARED), fork(), and then pass that mapping through to two different guests, you get to hold the pieces, just like if you did the same with normal memory.
Why does the kernel need to enforce this policy?
Re: [RFC PATCH 04/23] x86/cpufeatures: Add SGX1 and SGX2 sub-features
In reply to Kai Huangarch/x86/kernel/cpu/feat_ctl.c static void clear_sgx_caps(void) · line 102
Quoted 9 lines · Kai Huang
arch/x86/kernel/cpu/feat_ctl.c +2
-0@@ -97,6 +97,8 @@ static void clear_sgx_caps(void) {setup_clear_cpu_cap (X86_FEATURE_SGX );setup_clear_cpu_cap (X86_FEATURE_SGX_LC ); +setup_clear_cpu_cap (X86_FEATURE_SGX1 ); +setup_clear_cpu_cap (X86_FEATURE_SGX2 ); }
Logically, I think you want this *after* the "Allow SGX virtualization without Launch Control support" patch. As it stands, this will totally disable SGX (including virtualization) if launch control is unavailable.
Re: [RFC PATCH 05/23] x86/cpu/intel: Allow SGX virtualization without Launch Control support
In reply to Kai HuangQuoted 4 lines · Kai Huang
From: Sean Christopherson <[email protected]>
Allow SGX virtualization on systems without Launch Control support, i.e. allow KVM to expose SGX to guests that support non-LC configurations.
Context, please.
The kernel will currently disable all SGX support if the hardware does not support launch control. Make it more permissive to allow SGX virtualization on systems without Launch Control support. This will allow KVM to expose SGX to guests that have less-strict requirements on the availability of flexible launch control.
Quoted 3 lines
Introduce
clear_sgx_lc()to clearSGX_LCfeature bit only if SGX Launch Control is locked by BIOS when SGX virtualization is enabled, to prevent SGX driver being enabled.
This is another run-on, and it makes it really hard to figure out what it is trying to say.
Quoted 5 lines
Improve error message to distinguish three cases: 1) SGX disabled completely by BIOS; 2) SGX disabled completely due to SGX LC is locked by BIOS, and SGX virtualization is also disabled; 3) Only SGX driver is disabled due to SGX LC is locked by BIOS, but SGX virtualization is enabled.
Editing for grammar and clarity again...
Improve error message to distinguish between three cases. There are two cases where SGX support is completely disabled:
1) SGX has been disabled completely by the BIOS
2) SGX LC is locked by the BIOS. Bare-metal support is disabled because of LC unavailability. SGX virtualization is unavailable (because of Kconfig).
One where it is partially available:
3) SGX LC is locked by the BIOS. Bare-metal support is disabled because of LC unavailability. SGX virtualization is supported.
Quoted 44 lines
Signed-off-by Sean Christopherson <[email protected]>Co-developed-by Kai Huang <[email protected]>Signed-off-by Kai Huang <[email protected]>--- arch/x86/kernel/cpu/feat_ctl.c | 48 +++++++++++++++++++++++++--------- 1 file changed, 36 insertions(+), 12 deletions(-)arch/x86/kernel/cpu/feat_ctl.c +10
-4@@ -101,6 +101,11 @@ static void clear_sgx_caps(void)setup_clear_cpu_cap (X86_FEATURE_SGX2 ); } +static void clear_sgx_lc (void ) +{ +setup_clear_cpu_cap (X86_FEATURE_SGX_LC ); +} +static int __init nosgx (char *str) {clear_sgx_caps ();@@ -113,7 +118,7 @@ early_param("nosgx", nosgx);void init_ia32_feat_ctl (struct cpuinfo_x86 *c) {bool tboot =tboot_enabled (); -bool enable_sgx; +bool enable_sgx_virt, enable_sgx_driver;u64 msr;if (rdmsrl_safe (MSR_IA32_FEAT_CTL , &msr)) {@@ -123,12 +128,19 @@ void init_ia32_feat_ctl(struct cpuinfo_x86 *c) }/* -* Enable SGX if and only if the kernel supports SGX and Launch Control -* is supported, i.e. disable SGX if the LE hash MSRs can't be written. +* Enable SGX if and only if the kernel supports SGX. Require Launch +* Control support if SGX virtualization is *not* supported, i.e. +* disable SGX if the LE hash MSRs can't be written and SGX can't be +* exposed to a KVM guest (which might support non-LC configurations). */ - enable_sgx = cpu_has(c, X86_FEATURE_SGX) &&
arch/x86/kernel/cpu/feat_ctl.c void init_ia32_feat_ctl(struct cpuinfo_x86 *c) · line 143
Quoted 10 lines
arch/x86/kernel/cpu/feat_ctl.c +8
-2-cpu_has (c,X86_FEATURE_SGX_LC ) && -IS_ENABLED (CONFIG_X86_SGX ); + enable_sgx_driver =cpu_has (c,X86_FEATURE_SGX ) && +cpu_has (c,X86_FEATURE_SGX1 ) && +IS_ENABLED (CONFIG_X86_SGX ) && +cpu_has (c,X86_FEATURE_SGX_LC ); + enable_sgx_virt =cpu_has (c,X86_FEATURE_SGX ) && +cpu_has (c,X86_FEATURE_SGX1 ) && +IS_ENABLED (CONFIG_X86_SGX ) && +IS_ENABLED (CONFIG_X86_SGX_VIRTUALIZATION );
Don't we also need some runtime checks here? What if we boot on hardware that doesn't support KVM?
Quoted 26 lines
arch/x86/kernel/cpu/feat_ctl.c +7
-6if (msr &FEAT_CTL_LOCKED ) goto update_caps;@@ -151,8 +163,11 @@ void init_ia32_feat_ctl(struct cpuinfo_x86 *c) msr |=FEAT_CTL_VMX_ENABLED_INSIDE_SMX ; } -if (enable_sgx) - msr |=FEAT_CTL_SGX_ENABLED |FEAT_CTL_SGX_LC_ENABLED ; +if (enable_sgx_driver || enable_sgx_virt) { + msr |=FEAT_CTL_SGX_ENABLED ; +if (enable_sgx_driver) + msr |=FEAT_CTL_SGX_LC_ENABLED ; + }wrmsrl (MSR_IA32_FEAT_CTL , msr);@@ -175,10 +190,19 @@ void init_ia32_feat_ctl(struct cpuinfo_x86 *c) } update_sgx: -if (!(msr &FEAT_CTL_SGX_ENABLED ) || - !(msr &FEAT_CTL_SGX_LC_ENABLED ) || !enable_sgx) { -if (enable_sgx) -pr_err_once ("SGX disabled by BIOS\n" ); +if (!(msr &FEAT_CTL_SGX_ENABLED )) { +if (enable_sgx_driver || enable_sgx_virt)
arch/x86/kernel/cpu/feat_ctl.c void init_ia32_feat_ctl(struct cpuinfo_x86 *c) · line 204
Quoted 10 lines
arch/x86/kernel/cpu/feat_ctl.c +8
-0+pr_err_once ("SGX disabled by BIOS.\n" );clear_sgx_caps (); } +if (!(msr &FEAT_CTL_SGX_LC_ENABLED ) && + (enable_sgx_driver || enable_sgx_virt)) { +if (!enable_sgx_virt) { +pr_err_once ("SGX Launch Control is locked. Disable SGX.\n" ); +clear_sgx_caps (); + }else if (enable_sgx_driver) { +pr_err_once ("SGX Launch Control is locked. Disable SGX driver.\n" );
Should we have an explicit message for enabling virtualization? I'm not sure how many people will understand that "SGX driver" actually doesn't mean /dev/sgx_epc_virt.
Quoted 5 lines
arch/x86/kernel/cpu/feat_ctl.c +3
-0+clear_sgx_lc (); + } + } }
Re: [RFC PATCH 10/23] x86/sgx: Add helper to update SGX_LEPUBKEYHASHn MSRs
In reply to Kai HuangQuoted 5 lines · Kai Huang
Add a helper to update
SGX_LEPUBKEYHASHnMSRs. SGX virtualization also needs to update those MSRs based on guest's "virtual"SGX_LEPUBKEYHASHnbefore EINIT from guest.Signed-off-by Kai Huang <[email protected]>
This one is fine from a core x86 perspective:
Re: [RFC PATCH 11/23] x86/sgx: Add helpers to expose ECREATE and EINIT to KVM
In reply to Kai HuangQuoted 7 lines · Kai Huang
From: Sean Christopherson <[email protected]>
Provide wrappers around
__ecreate()and__einit()to hide the ugliness of overloading the ENCLS return value to encode multiple error formats in a single int. KVM will trap-and-execute ECREATE and EINIT as part of SGX virtualization, and on an exception, KVM needs the trapnr so that it can inject the correct fault into the guest.
This is missing a bit of a step about how and why ECREATE needs to be run in the host in the first place.
Quoted 29 lines
arch/x86/include/asm/sgx.h new file +16
-0@@ -0,0 +1,16 @@ +/* SPDX-License-Identifier: GPL-2.0 */ +#ifndef _ASM_X86_SGX_H +#define _ASM_X86_SGX_H + +#include <linux/types.h> + +#ifdef CONFIG_X86_SGX_VIRTUALIZATION +struct sgx_pageinfo ; + +int sgx_virt_ecreate (struct sgx_pageinfo *pageinfo,void __user *secs, +int *trapnr); +int sgx_virt_einit (void __user *sigstruct,void __user *token, +void __user *secs,u64 *lepubkeyhash,int *trapnr); +#endif + +#endif /* _ASM_X86_SGX_H */ arch/x86/kernel/cpu/sgx/virt.c +0
-0@@ -261,3 +261,58 @@ int __init sgx_virt_epc_init(void)return misc_register (&sgx_virt_epc_dev);
arch/x86/kernel/cpu/sgx/virt.c int __init sgx_virt_epc_init(void) · line 272
Quoted 10 lines
arch/x86/kernel/cpu/sgx/virt.c +9
-0} + +int sgx_virt_ecreate (struct sgx_pageinfo *pageinfo,void __user *secs, +int *trapnr) +{ +int ret; + +__uaccess_begin (); + ret =__ecreate (pageinfo, (void *)secs); +__uaccess_end ();
The __uaccess_begin/end() worries me. There are *very* few of these in the kernel and it seems like something we want to use as sparingly as possible.
Why don't we just use the kernel mapping for 'secs' and not have to deal with stac/clac?
I'm also just generally worried about casting away an __user without doing any checking. How is that OK?
Quoted 11 lines
arch/x86/kernel/cpu/sgx/virt.c +11
-0+if (encls_faulted (ret)) { + *trapnr =ENCLS_TRAPNR (ret); +return -EFAULT ; + } + +/* ECREATE doesn't return an error code, it faults or succeeds. */ +WARN_ON_ONCE (ret); +return 0 ; +} +EXPORT_SYMBOL_GPL (sgx_virt_ecreate); +
arch/x86/kernel/cpu/sgx/virt.c int __init sgx_virt_epc_init(void) · line 294
Quoted 10 lines
arch/x86/kernel/cpu/sgx/virt.c +10
-0+static int __sgx_virt_einit (void __user *sigstruct,void __user *token, +void __user *secs) +{ +int ret; + +__uaccess_begin (); + ret =__einit ((void *)sigstruct, (void *)token, (void *)secs); +__uaccess_end (); +return ret; +}
If casting away one __user wasn't good enough, we get three! :)
We need some more background in the changelog on why this is OK.
Quoted 24 lines
arch/x86/kernel/cpu/sgx/virt.c +23
-0+int sgx_virt_einit (void __user *sigstruct,void __user *token, +void __user *secs,u64 *lepubkeyhash,int *trapnr) +{ +int ret; + +if (!boot_cpu_has (X86_FEATURE_SGX_LC )) { + ret =__sgx_virt_einit (sigstruct, token, secs); + }else { +preempt_disable (); + +sgx_update_lepubkeyhash (lepubkeyhash); + + ret =__sgx_virt_einit (sigstruct, token, secs); +preempt_enable (); + } + +if (encls_faulted (ret)) { + *trapnr =ENCLS_TRAPNR (ret); +return -EFAULT ; + } +return ret; +} +EXPORT_SYMBOL_GPL (sgx_virt_einit);
Sean Christopherson
maintainerRe: [RFC PATCH 03/23] x86/sgx: Introduce virtual EPC for use by KVM guests
In reply to Dave HansenQuoted 22 lines · Dave Hansen
Quoted 21 lines · Kai Huang
From: Sean Christopherson <[email protected]>
arch/x86/kernel/cpu/sgx/main.c +2
-0@@ -9,9 +9,11 @@#include <linux/sched/mm.h> #include <linux/sched/signal.h> #include <linux/slab.h> +#include "arch.h" #include "driver.h" #include "encl.h" #include "encls.h" +#include "virt.h" struct sgx_epc_section sgx_epc_sections[SGX_MAX_EPC_SECTIONS ];static int sgx_nr_epc_sections;@@ -726,7 +728,8 @@ static void __init sgx_init(void)if (!sgx_page_reclaimer_init ()) goto err_page_cache;
Quoted 8 lines
arch/x86/kernel/cpu/sgx/main.c static void __init sgx_init(void) · line 734
Quoted 5 lines
arch/x86/kernel/cpu/sgx/main.c +2
-1- ret =sgx_drv_init (); +/* Success if the native *or* virtual EPC driver initialized cleanly. */ + ret = !!sgx_drv_init () & !!sgx_virt_epc_init ();if (ret) goto err_kthread;FWIW, I hate that conditional. But, I tried to write to to be something more sane and failed.
Heh, you're welcome :-D
Quoted 25 lines
Quoted 25 lines
arch/x86/kernel/cpu/sgx/virt.c new file +19
-0@@ -0,0 +1,263 @@ +// SPDX-License-Identifier: GPL-2.0 +/* Copyright(c) 2016-20 Intel Corporation. */ + +#include <linux/miscdevice.h> +#include <linux/mm.h> +#include <linux/mman.h> +#include <linux/sched/mm.h> +#include <linux/sched/signal.h> +#include <linux/slab.h> +#include <linux/xarray.h> +#include <asm/sgx.h> +#include <uapi/asm/sgx.h> + +#include "encls.h" +#include "sgx.h" +#include "virt.h" + +struct sgx_virt_epc { +struct xarray page_array;
Quoted 8 lines
arch/x86/kernel/cpu/sgx/virt.c line 25
Quoted 6 lines
arch/x86/kernel/cpu/sgx/virt.c new file +6
-0+struct mutex lock; +struct mm_struct *mm; +}; + +static struct mutex virt_epc_lock; +static struct list_head virt_epc_zombie_pages;What does the lock protect?
Effectively, the list of zombie SECS pages. Not sure why I used a generic name.
Quoted 1 line
What are zombie pages?
My own terminology for SECS pages whose virtual EPC has been destroyed but can't be reclaimed due to them having child EPC pages in other virtual EPCs.
Quoted 2 lines
BTW, if zombies are SECS-only, shouldn't that be in the name rather than "epc"?
I used the virt_epc prefix/namespace to tag it as a global list. I've no argument against something like zombie_secs_pages.
Quoted 11 lines
arch/x86/kernel/cpu/sgx/virt.c line 34
Quoted 8 lines
arch/x86/kernel/cpu/sgx/virt.c new file +8
-0+static int __sgx_virt_epc_fault (struct sgx_virt_epc *epc, +struct vm_area_struct *vma,unsigned long addr) +{ +struct sgx_epc_page *epc_page; +unsigned long index, pfn; +int ret; + +/* epc->lock must already have been hold */ /* epc->lock must already be held */
Quoted 5 lines
Wouldn't this be better as:
WARN_ON(!mutex_is_locked(&epc->lock));
?
Or just proper lockdep?
Quoted 65 lines
Quoted 65 lines
arch/x86/kernel/cpu/sgx/virt.c new file +65
-0+/* Calculate index of EPC page in virtual EPC's page_array */ + index = vma->vm_pgoff +PFN_DOWN (addr - vma->vm_start); + + epc_page =xa_load (&epc->page_array, index); +if (epc_page) +return 0 ; + + epc_page =sgx_alloc_epc_page (epc, false); +if (IS_ERR (epc_page)) +return PTR_ERR (epc_page); + + ret =xa_err (xa_store (&epc->page_array, index, epc_page,GFP_KERNEL )); +if (ret) + goto err_free; + + pfn =PFN_DOWN (sgx_get_epc_phys_addr (epc_page)); + + ret =vmf_insert_pfn (vma, addr, pfn); +if (ret !=VM_FAULT_NOPAGE ) { + ret = -EFAULT ; + goto err_delete; + } + +return 0 ; + +err_delete: +xa_erase (&epc->page_array, index); +err_free: +sgx_free_epc_page (epc_page); +return ret; +} + +static vm_fault_t sgx_virt_epc_fault (struct vm_fault *vmf) +{ +struct vm_area_struct *vma = vmf->vma; +struct sgx_virt_epc *epc = vma->vm_private_data; +int ret; + +mutex_lock (&epc->lock); + ret =__sgx_virt_epc_fault (epc, vma, vmf->address); +mutex_unlock (&epc->lock); + +if (!ret) +return VM_FAULT_NOPAGE ; + +if (ret == -EBUSY && (vmf->flags &FAULT_FLAG_ALLOW_RETRY )) { +mmap_read_unlock (vma->vm_mm); +return VM_FAULT_RETRY ; + } + +return VM_FAULT_SIGBUS ; +} + +const struct vm_operations_struct sgx_virt_epc_vm_ops = { + .fault = sgx_virt_epc_fault, +}; + +static int sgx_virt_epc_mmap (struct file *file,struct vm_area_struct *vma) +{ +struct sgx_virt_epc *epc = file->private_data; + +if (!(vma->vm_flags &VM_SHARED )) +return -EINVAL ; + +/*
Quoted 8 lines
arch/x86/kernel/cpu/sgx/virt.c line 105
Quoted 5 lines
arch/x86/kernel/cpu/sgx/virt.c new file +5
-0+* Don't allow mmap() from child after fork(), since child and parent +* cannot map to the same EPC. +*/ +if (vma->vm_mm != epc->mm) +return -EINVAL ;I mentioned this below, but I'm not buying this logic. I know it would be *bad*, but I don't see why the kernel needs to keep it from happening.
There's no known use case (KVM doesn't support sharing a VM across multiple mm structs), and supporting multiple mm structs is a nightmare; see the driver for the amount of pain incurred.
And IIRC, supporting VMM (KVM) EPC oversubscription, which may or may not ever happen, was borderline impossible if virtual EPC supports multiple mm structs as the interaction between KVM and virtual EPC is a disaster in that case.
Quoted 8 lines
Quoted 8 lines
arch/x86/kernel/cpu/sgx/virt.c new file +8
-0+ vma->vm_ops = &sgx_virt_epc_vm_ops; +/* Don't copy VMA in fork() */ + vma->vm_flags |=VM_PFNMAP |VM_IO |VM_DONTDUMP |VM_DONTCOPY ; + vma->vm_private_data = file->private_data; + +return 0 ; +} +
Quoted 8 lines
arch/x86/kernel/cpu/sgx/virt.c line 120
Quoted 6 lines
arch/x86/kernel/cpu/sgx/virt.c new file +6
-0+static int sgx_virt_epc_free_page (struct sgx_epc_page *epc_page) +{ +int ret; + +if (!epc_page) +return 0 ;I always worry about these. Why is passing NULL around OK?
I suspect I did it to mimic kfree() behavior. I don't _think_ the radix (now xarray) usage will ever encounter a NULL entry.
Quoted 17 lines
arch/x86/kernel/cpu/sgx/virt.c line 143
Quoted 13 lines
arch/x86/kernel/cpu/sgx/virt.c new file +12
-0+ ret =__eremove (sgx_get_epc_virt_addr (epc_page)); +if (ret) { +/* +* Only SGX_CHILD_PRESENT is expected, which is because of +* EREMOVE-ing an SECS still with child, in which case it can +* be handled by EREMOVE-ing the SECS again after all pages in +* virtual EPC have been EREMOVE-ed. See comments in below in +* sgx_virt_epc_release(). +*/ +WARN_ON_ONCE (ret !=SGX_CHILD_PRESENT ); +return ret; + }I find myself wondering what errors could cause the
WARN_ON_ONCE()to be hit. The SDM indicates that it's only:
Quoted 4 lines
SGX_ENCLAVE_ACTIf there are still logical processors executing
inside the enclave.Should that be mentioned in the comment?
And faults, which are also spliced into the return value by the ENCLS macros. I do remember hitting this WARN when I broke things, though I can't remember whether it was a fault or the SGX_ENCLAVE_ACT scenario. Probably the latter?
Quoted 5 lines
Quoted 5 lines
arch/x86/kernel/cpu/sgx/virt.c new file +5
-0+ +__sgx_free_epc_page (epc_page); +return 0 ; +} +
...
Quoted 22 lines
arch/x86/kernel/cpu/sgx/virt.c
Quoted 2 lines
arch/x86/kernel/cpu/sgx/virt.c new file +2
-0+xa_for_each (&epc->page_array, index, entry) { + epc_page = entry;Then, talk about the error condition here:
arch/x86/kernel/cpu/sgx/virt.c line 186
Quoted 7 lines
arch/x86/kernel/cpu/sgx/virt.c new file +7
-0+/* +* Error here means that EREMOVE failed due to a SECS page +* still has child on *another* EPC instance. Put it to a +* temporary SECS list which will be spliced to 'zombie page +* list' and will be EREMOVE-ed again when freeing another +* virtual EPC instance. +*/ Surprise, I've got another rewrite:
/*
* An EREMOVE failure here means that the SECS page
* still has children. But, since all children in this
* 'sgx_virt_epc' have been removed, the SECS page must
* have a child on another instance.
*/
Quoted 5 lines
arch/x86/kernel/cpu/sgx/virt.c
Quoted 2 lines
arch/x86/kernel/cpu/sgx/virt.c new file +2
-0+if (sgx_virt_epc_free_page (epc_page)) +list_add_tail (&epc_page->list, &secs_pages);Why move these over to
&secs_listhere? I think it's to avoid anotherxa_for_each()below, but it's not clear.
Yes? IIRC, the sole motivation is to make the list_split_tail() operation as short as possible while holding the global virt_epc_lock.
Quoted 3 lines
Quoted 3 lines
arch/x86/kernel/cpu/sgx/virt.c new file +3
-0+xa_erase (&epc->page_array, index); + } +
...
Quoted 16 lines
arch/x86/kernel/cpu/sgx/virt.c line 209
Quoted 12 lines
arch/x86/kernel/cpu/sgx/virt.c new file +12
-0+mutex_lock (&virt_epc_lock); +list_for_each_entry_safe (epc_page, tmp, &virt_epc_zombie_pages, list) { +/* +* Speculatively remove the page from the list of zombies, if +* the page is successfully EREMOVE it will be added to the +* list of free pages. If EREMOVE fails, throw the page on the +* local list, which will be spliced on at the end. +*/ +list_del (&epc_page->list); + +if (sgx_virt_epc_free_page (epc_page)) +list_add_tail (&epc_page->list, &secs_pages);I don't get this. Couldn't you do without the unconditional
list_del()and instead just do:
Quoted 5 lines
if (!sgx_virt_epc_free_page(epc_page)) list_del(&epc_page->list);Or does the
free()code clobber thelist_head? If that's the case, maybe you should say that explicitly.
More or less. EPC pages need to be removed from their list before freeing, once a page is freed it is owned by the allocator. Deleting after freeing leads to list corruption if a different thread allocates the page and adds it to a different list.
Quoted 10 lines
Quoted 10 lines
arch/x86/kernel/cpu/sgx/virt.c new file +10
-0+ } + +if (!list_empty (&secs_pages)) +list_splice_tail (&secs_pages, &virt_epc_zombie_pages); +mutex_unlock (&virt_epc_lock); + +kfree (epc); + +return 0 ; +}
Sean Christopherson
maintainerRe: [RFC PATCH 11/23] x86/sgx: Add helpers to expose ECREATE and EINIT to KVM
In reply to Dave HansenQuoted 11 lines · Dave Hansen
Quoted 7 lines · Kai Huang
From: Sean Christopherson <[email protected]>
Provide wrappers around
__ecreate()and__einit()to hide the ugliness of overloading the ENCLS return value to encode multiple error formats in a single int. KVM will trap-and-execute ECREATE and EINIT as part of SGX virtualization, and on an exception, KVM needs the trapnr so that it can inject the correct fault into the guest.This is missing a bit of a step about how and why ECREATE needs to be run in the host in the first place.
There's (hopefully) good info in the KVM usage patch that can be borrowed:
Add an ECREATE handler that will be used to intercept ECREATE for the purpose of enforcing and enclave's MISCSELECT, ATTRIBUTES and XFRM, i.e. to allow userspace to restrict SGX features via CPUID. ECREATE will be intercepted when any of the aforementioned masks diverges from hardware in order to enforce the desired CPUID model, i.e. inject #GP if the guest attempts to set a bit that hasn't been enumerated as allowed-1 in CPUID.
Quoted 44 lines
Quoted 29 lines
arch/x86/include/asm/sgx.h new file +16
-0@@ -0,0 +1,16 @@ +/* SPDX-License-Identifier: GPL-2.0 */ +#ifndef _ASM_X86_SGX_H +#define _ASM_X86_SGX_H + +#include <linux/types.h> + +#ifdef CONFIG_X86_SGX_VIRTUALIZATION +struct sgx_pageinfo ; + +int sgx_virt_ecreate (struct sgx_pageinfo *pageinfo,void __user *secs, +int *trapnr); +int sgx_virt_einit (void __user *sigstruct,void __user *token, +void __user *secs,u64 *lepubkeyhash,int *trapnr); +#endif + +#endif /* _ASM_X86_SGX_H */ arch/x86/kernel/cpu/sgx/virt.c +0
-0@@ -261,3 +261,58 @@ int __init sgx_virt_epc_init(void)return misc_register (&sgx_virt_epc_dev);arch/x86/kernel/cpu/sgx/virt.c int __init sgx_virt_epc_init(void) · line 272
Quoted 10 lines
arch/x86/kernel/cpu/sgx/virt.c +9
-0} + +int sgx_virt_ecreate (struct sgx_pageinfo *pageinfo,void __user *secs, +int *trapnr) +{ +int ret; + +__uaccess_begin (); + ret =__ecreate (pageinfo, (void *)secs); +__uaccess_end ();The __uaccess_begin/end() worries me. There are *very* few of these in the kernel and it seems like something we want to use as sparingly as possible.
Quoted 2 lines
Why don't we just use the kernel mapping for 'secs' and not have to deal with stac/clac?
The kernel mapping isn't readily available. At this point, it's not even guaranteed that @secs points at an EPC page. Unlike the driver code, where the EPC page is allocated on-demand by the kernel, the pointer here is userspace (technically guest) controlled. The caller (KVM) is responsible for ensuring it's a valid userspace address, but the SGX/EPC specific checks are mostly deferred to hardware.
It's also possible to either retrieve the existing kernel mapping or to generate a new mapping by resolving the PFN; this is/was simpler.
Quoted 2 lines
I'm also just generally worried about casting away an
__userwithout doing any checking. How is that OK?
Short answer, KVM validates the virtual addresses.
KVM validates the host virtual addresses (HVA) when creating a memslot (maps GPA->HVA). The HVAs that are passed to these helpers are generated/retrieved by KVM translating GVA->GPA->HVA; the GPA->HVA stage ensures the address is in a valid memslot, and thus a valid user address.
That being said, these aren't exactly fast operations, adding access_ok() checks is probably a good idea.
Re: [RFC PATCH 11/23] x86/sgx: Add helpers to expose ECREATE and EINIT to KVM
Quoted 15 lines · Sean Christopherson
Quoted 11 lines · Dave Hansen
Quoted 7 lines · Kai Huang
From: Sean Christopherson <[email protected]>
Provide wrappers around
__ecreate()and__einit()to hide the ugliness of overloading the ENCLS return value to encode multiple error formats in a single int. KVM will trap-and-execute ECREATE and EINIT as part of SGX virtualization, and on an exception, KVM needs the trapnr so that it can inject the correct fault into the guest.This is missing a bit of a step about how and why ECREATE needs to be run in the host in the first place.
There's (hopefully) good info in the KVM usage patch that can be borrowed:
Quoted 7 lines
Add an ECREATE handler that will be used to intercept ECREATE for the purpose of enforcing and enclave's MISCSELECT, ATTRIBUTES and XFRM, i.e. to allow userspace to restrict SGX features via CPUID. ECREATE will be intercepted when any of the aforementioned masks diverges from hardware in order to enforce the desired CPUID model, i.e. inject #GP if the guest attempts to set a bit that hasn't been enumerated as allowed-1 in CPUID.
OK, so in plain language: the bare-metal kernel must intercept ECREATE to be able to impose policies on guests. When it does this, the bare-metal kernel runs ECREATE against the userspace mapping of the virtualized EPC.
Quoted 44 lines
Quoted 44 lines
Quoted 29 lines
arch/x86/include/asm/sgx.h new file +16
-0@@ -0,0 +1,16 @@ +/* SPDX-License-Identifier: GPL-2.0 */ +#ifndef _ASM_X86_SGX_H +#define _ASM_X86_SGX_H + +#include <linux/types.h> + +#ifdef CONFIG_X86_SGX_VIRTUALIZATION +struct sgx_pageinfo ; + +int sgx_virt_ecreate (struct sgx_pageinfo *pageinfo,void __user *secs, +int *trapnr); +int sgx_virt_einit (void __user *sigstruct,void __user *token, +void __user *secs,u64 *lepubkeyhash,int *trapnr); +#endif + +#endif /* _ASM_X86_SGX_H */ arch/x86/kernel/cpu/sgx/virt.c +0
-0@@ -261,3 +261,58 @@ int __init sgx_virt_epc_init(void)return misc_register (&sgx_virt_epc_dev);arch/x86/kernel/cpu/sgx/virt.c int __init sgx_virt_epc_init(void) · line 272
Quoted 10 lines
arch/x86/kernel/cpu/sgx/virt.c +9
-0} + +int sgx_virt_ecreate (struct sgx_pageinfo *pageinfo,void __user *secs, +int *trapnr) +{ +int ret; + +__uaccess_begin (); + ret =__ecreate (pageinfo, (void *)secs); +__uaccess_end ();The __uaccess_begin/end() worries me. There are *very* few of these in the kernel and it seems like something we want to use as sparingly as possible.
Quoted 4 lines
Quoted 2 lines
Why don't we just use the kernel mapping for 'secs' and not have to deal with stac/clac?
The kernel mapping isn't readily available.
Oh, duh. There's no kernel mapping for EPC... it's not RAM in the first place.
Quoted 6 lines
At this point, it's not even
guaranteed that @secs points at an EPC page. Unlike the driver code, where the EPC page is allocated on-demand by the kernel, the pointer here is userspace (technically guest) controlled. The caller (KVM) is responsible for ensuring it's a valid userspace address, but the SGX/EPC specific checks are mostly deferred to hardware.
Ahh, got it. Kai, could we get some of this into comments or the changelog?
Quoted 9 lines
Quoted 2 lines
I'm also just generally worried about casting away an
__userwithout doing any checking. How is that OK?Short answer, KVM validates the virtual addresses.
KVM validates the host virtual addresses (HVA) when creating a memslot (maps
GPA->HVA). The HVAs that are passed to these helpers are generated/retrieved by KVM translatingGVA->GPA->HVA; theGPA->HVAstage ensures the address is in a valid memslot, and thus a valid user address.
There is something a *bit* unpalatable about having KVM fill an 'unsigned long' only to cast it to a (void __user *), the to cast it back to a (void *) to pass it to the SGX inlines.
I guess sparse would catch us in the window that it is __user if someone tried to dereference it.
Adding access_ok()'s sounds like a good idea to me. Or, at *least* commenting why they're not necessary.
Re: [RFC PATCH 02/23] x86/sgx: Add enum for SGX_CHILD_PRESENT error code
In reply to Dave HansenQuoted 8 lines · On Wed, 6 Jan 2021 10:28:55 -0800 Dave Hansen wrote:
Quoted 2 lines · Kai Huang
Add
SGX_CHILD_PRESENTfor use by SGX virtualization to assert EREMOVE failures are expected, but only due toSGX_CHILD_PRESENT.This dances around the fact that this is an architectural error-code. Could that be explicit? Maybe the subject should be:
Add
SGX_CHILD_PRESENThardware error code
Sure. I'll do that.
Re: [RFC PATCH 04/23] x86/cpufeatures: Add SGX1 and SGX2 sub-features
In reply to Dave HansenQuoted 14 lines · On Wed, 6 Jan 2021 11:39:46 -0800 Dave Hansen wrote:
arch/x86/kernel/cpu/feat_ctl.c static void clear_sgx_caps(void) · line 102
Quoted 9 lines · Kai Huang
arch/x86/kernel/cpu/feat_ctl.c +2
-0@@ -97,6 +97,8 @@ static void clear_sgx_caps(void) {setup_clear_cpu_cap (X86_FEATURE_SGX );setup_clear_cpu_cap (X86_FEATURE_SGX_LC ); +setup_clear_cpu_cap (X86_FEATURE_SGX1 ); +setup_clear_cpu_cap (X86_FEATURE_SGX2 ); }Logically, I think you want this *after* the "Allow SGX virtualization without Launch Control support" patch. As it stands, this will totally disable SGX (including virtualization) if launch control is unavailable.
To me it is better to be here, since clear_sgx_caps(), which disables SGX totally, should logically clear all SGX feature bits, no matter later patch's behavior. So when new SGX bits are introduced, clear_sgx_caps() should clear them too. Otherwise the logic of this patch (adding new SGX feature bits) is not complete IMHO.
And actually in later patch "Allow SGX virtualization without Launch Control support", a new clear_sgx_lc() is added, and is called when LC is not available but SGX virtualization is enabled, to make sure only SGX_LC bit is cleared in this case. I don't quite understand why we need to clear SGX1 and SGX2 in clear_sgx_caps() after the later patch.
Borislav Petkov
maintainerRe: [RFC PATCH 04/23] x86/cpufeatures: Add SGX1 and SGX2 sub-features
In reply to Kai Huangarch/x86/include/asm/cpufeatures.h line 388
Quoted 3 lines · Kai Huang
arch/x86/include/asm/cpufeatures.h +3
-0+/* Intel-defined SGX features, CPUID level 0x00000012:0 (EAX), word 19 */ +#define X86_FEATURE_SGX1 (19*32+ 0)/* SGX1 leaf functions */ +#define X86_FEATURE_SGX2 (19*32+ 1)/* SGX2 leaf functions */
Is anything else from that leaf going to be added later? Bit 5 is "supports ENCLV instruction leaves", 6 is ENCLS insn leaves... are those going to be used in the kernel too eventually?
Rest of them is reserved in the SDM which probably means internal only for now.
Signature · 4 lines
Regards/Gruss,
Boris.
https://people.kernel.org/tglx/notes-about-netiquetteRe: [RFC PATCH 04/23] x86/cpufeatures: Add SGX1 and SGX2 sub-features
In reply to Kai HuangQuoted 20 lines · Kai Huang
Quoted 13 lines · On Wed, 6 Jan 2021 11:39:46 -0800 Dave Hansen wrote:
arch/x86/kernel/cpu/feat_ctl.c static void clear_sgx_caps(void) · line 102
Quoted 9 lines · Kai Huang
arch/x86/kernel/cpu/feat_ctl.c +2
-0@@ -97,6 +97,8 @@ static void clear_sgx_caps(void) {setup_clear_cpu_cap (X86_FEATURE_SGX );setup_clear_cpu_cap (X86_FEATURE_SGX_LC ); +setup_clear_cpu_cap (X86_FEATURE_SGX1 ); +setup_clear_cpu_cap (X86_FEATURE_SGX2 ); }Logically, I think you want this *after* the "Allow SGX virtualization without Launch Control support" patch. As it stands, this will totally disable SGX (including virtualization) if launch control is unavailable.
To me it is better to be here, since
clear_sgx_caps(), which disables SGX totally, should logically clear all SGX feature bits, no matter later patch's behavior. So when new SGX bits are introduced,clear_sgx_caps()should clear them too. Otherwise the logic of this patch (adding new SGX feature bits) is not complete IMHO.
Quoted 5 lines
And actually in later patch "Allow SGX virtualization without Launch Control support", a new
clear_sgx_lc()is added, and is called when LC is not available but SGX virtualization is enabled, to make sure onlySGX_LCbit is cleared in this case. I don't quite understand why we need to clear SGX1 and SGX2 inclear_sgx_caps()after the later patch.
I was talking about patch ordering. It could be argued that this goes after the content of patch 05/23. Please _consider_ changing the ordering.
If that doesn't work for some reason, please at least call out in the changelog that it leaves a temporarily funky situation.
Re: [RFC PATCH 05/23] x86/cpu/intel: Allow SGX virtualization without Launch Control support
In reply to Dave HansenQuoted 6 lines · On Wed, 6 Jan 2021 11:54:52 -0800 Dave Hansen wrote:
Quoted 4 lines · Kai Huang
From: Sean Christopherson <[email protected]>
Allow SGX virtualization on systems without Launch Control support, i.e. allow KVM to expose SGX to guests that support non-LC configurations.
Quoted 7 lines
Context, please.
The kernel will currently disable all SGX support if the hardware does not support launch control. Make it more permissive to allow SGX virtualization on systems without Launch Control support. This will allow KVM to expose SGX to guests that have less-strict requirements on the availability of flexible launch control.
OK. I'll add this.
Quoted 7 lines
Quoted 4 lines
Introduce
clear_sgx_lc()to clearSGX_LCfeature bit only if SGX Launch Control is locked by BIOS when SGX virtualization is enabled, to prevent SGX driver being enabled.This is another run-on, and it makes it really hard to figure out what it is trying to say.
How about just removing this paragraph? It is a little bit detail anyway. We can add some comment in the code.
Quoted 9 lines
Quoted 6 lines
Improve error message to distinguish three cases: 1) SGX disabled completely by BIOS; 2) SGX disabled completely due to SGX LC is locked by BIOS, and SGX virtualization is also disabled; 3) Only SGX driver is disabled due to SGX LC is locked by BIOS, but SGX virtualization is enabled.
Editing for grammar and clarity again...
Quoted 9 lines
Improve error message to distinguish between three cases. There are two cases where SGX support is completely disabled:
1) SGX has been disabled completely by the BIOS
2) SGX LC is locked by the BIOS. Bare-metal support is disabled because of LC unavailability. SGX virtualization is unavailable (because ofKconfig).
One where it is partially available:
3) SGX LC is locked by the BIOS. Bare-metal support is disabled because of LC unavailability. SGX virtualization is supported.
OK. Thanks for help here.
Quoted 50 lines
Quoted 50 lines
Signed-off-by Sean Christopherson <[email protected]>Co-developed-by Kai Huang <[email protected]>Signed-off-by Kai Huang <[email protected]>--- arch/x86/kernel/cpu/feat_ctl.c | 48 +++++++++++++++++++++++++--------- 1 file changed, 36 insertions(+), 12 deletions(-)arch/x86/kernel/cpu/feat_ctl.c +13
-6@@ -101,6 +101,11 @@ static void clear_sgx_caps(void)setup_clear_cpu_cap (X86_FEATURE_SGX2 ); } +static void clear_sgx_lc (void ) +{ +setup_clear_cpu_cap (X86_FEATURE_SGX_LC ); +} +static int __init nosgx (char *str) {clear_sgx_caps ();@@ -113,7 +118,7 @@ early_param("nosgx", nosgx);void init_ia32_feat_ctl (struct cpuinfo_x86 *c) {bool tboot =tboot_enabled (); -bool enable_sgx; +bool enable_sgx_virt, enable_sgx_driver;u64 msr;if (rdmsrl_safe (MSR_IA32_FEAT_CTL , &msr)) {@@ -123,12 +128,19 @@ void init_ia32_feat_ctl(struct cpuinfo_x86 *c) }/* -* Enable SGX if and only if the kernel supports SGX and Launch Control -* is supported, i.e. disable SGX if the LE hash MSRs can't be written. +* Enable SGX if and only if the kernel supports SGX. Require Launch +* Control support if SGX virtualization is *not* supported, i.e. +* disable SGX if the LE hash MSRs can't be written and SGX can't be +* exposed to a KVM guest (which might support non-LC configurations). */ - enable_sgx =cpu_has (c,X86_FEATURE_SGX ) && -cpu_has (c,X86_FEATURE_SGX_LC ) && -IS_ENABLED (CONFIG_X86_SGX ); + enable_sgx_driver =cpu_has (c,X86_FEATURE_SGX ) && +cpu_has (c,X86_FEATURE_SGX1 ) && + IS_ENABLED(CONFIG_X86_SGX) &&
Quoted 8 lines
arch/x86/kernel/cpu/feat_ctl.c void init_ia32_feat_ctl(struct cpuinfo_x86 *c) · line 143
Quoted 5 lines
arch/x86/kernel/cpu/feat_ctl.c +5
-0+cpu_has (c,X86_FEATURE_SGX_LC ); + enable_sgx_virt =cpu_has (c,X86_FEATURE_SGX ) && +cpu_has (c,X86_FEATURE_SGX1 ) && +IS_ENABLED (CONFIG_X86_SGX ) && +IS_ENABLED (CONFIG_X86_SGX_VIRTUALIZATION );Don't we also need some runtime checks here? What if we boot on hardware that doesn't support KVM?
Yeah I kinda agree here. KVM will be available if X86_FEATURE_VMX is available. I am OK to add additional check right after 'update_sgx' label:
update_sgx:
if (!cpu_has(c, X86_FEATURE_VMX)) enable_sgx_driver = 0;
The rest logic should just work. If necessary, we can also add some message to say SGX virtualization is disabled due to VMX is not available.
Sean, what is your opinion?
Quoted 33 lines
Quoted 33 lines
arch/x86/kernel/cpu/feat_ctl.c +11
-6if (msr &FEAT_CTL_LOCKED ) goto update_caps;@@ -151,8 +163,11 @@ void init_ia32_feat_ctl(struct cpuinfo_x86 *c) msr |=FEAT_CTL_VMX_ENABLED_INSIDE_SMX ; } -if (enable_sgx) - msr |=FEAT_CTL_SGX_ENABLED |FEAT_CTL_SGX_LC_ENABLED ; +if (enable_sgx_driver || enable_sgx_virt) { + msr |=FEAT_CTL_SGX_ENABLED ; +if (enable_sgx_driver) + msr |=FEAT_CTL_SGX_LC_ENABLED ; + }wrmsrl (MSR_IA32_FEAT_CTL , msr);@@ -175,10 +190,19 @@ void init_ia32_feat_ctl(struct cpuinfo_x86 *c) } update_sgx: -if (!(msr &FEAT_CTL_SGX_ENABLED ) || - !(msr &FEAT_CTL_SGX_LC_ENABLED ) || !enable_sgx) { -if (enable_sgx) -pr_err_once ("SGX disabled by BIOS\n" ); +if (!(msr &FEAT_CTL_SGX_ENABLED )) { +if (enable_sgx_driver || enable_sgx_virt) +pr_err_once ("SGX disabled by BIOS.\n" );clear_sgx_caps (); } +if (!(msr &FEAT_CTL_SGX_LC_ENABLED ) && + (enable_sgx_driver || enable_sgx_virt)) { +if (!enable_sgx_virt) {
Quoted 8 lines
arch/x86/kernel/cpu/feat_ctl.c void init_ia32_feat_ctl(struct cpuinfo_x86 *c) · line 204
Quoted 4 lines
arch/x86/kernel/cpu/feat_ctl.c +4
-0+pr_err_once ("SGX Launch Control is locked. Disable SGX.\n" ); +clear_sgx_caps (); + }else if (enable_sgx_driver) { +pr_err_once ("SGX Launch Control is locked. Disable SGX driver.\n" );Should we have an explicit message for enabling virtualization? I'm not sure how many people will understand that "SGX driver" actually doesn't mean
/dev/sgx_epc_virt.
OK. I'll add an explicit message for that. Let me see how I can refine this.
Thanks for comments.
Quoted 7 lines
Quoted 6 lines
arch/x86/kernel/cpu/feat_ctl.c +3
-0+clear_sgx_lc (); + } + } }
Re: [RFC PATCH 05/23] x86/cpu/intel: Allow SGX virtualization without Launch Control support
Quoted 7 lines · Kai Huang
Quoted 5 lines
Quoted 3 lines
Introduce
clear_sgx_lc()to clearSGX_LCfeature bit only if SGX Launch Control is locked by BIOS when SGX virtualization is enabled, to prevent SGX driver being enabled.This is another run-on, and it makes it really hard to figure out what it is trying to say.
How about just removing this paragraph? It is a little bit detail anyway. We can add some comment in the code.
Fine with me.
Re: [RFC PATCH 04/23] x86/cpufeatures: Add SGX1 and SGX2 sub-features
In reply to Dave HansenQuoted 27 lines · On Wed, 6 Jan 2021 14:21:39 -0800 Dave Hansen wrote:
Quoted 20 lines · Kai Huang
Quoted 13 lines · On Wed, 6 Jan 2021 11:39:46 -0800 Dave Hansen wrote:
arch/x86/kernel/cpu/feat_ctl.c static void clear_sgx_caps(void) · line 102
Quoted 9 lines · Kai Huang
arch/x86/kernel/cpu/feat_ctl.c +2
-0@@ -97,6 +97,8 @@ static void clear_sgx_caps(void) {setup_clear_cpu_cap (X86_FEATURE_SGX );setup_clear_cpu_cap (X86_FEATURE_SGX_LC ); +setup_clear_cpu_cap (X86_FEATURE_SGX1 ); +setup_clear_cpu_cap (X86_FEATURE_SGX2 ); }Logically, I think you want this *after* the "Allow SGX virtualization without Launch Control support" patch. As it stands, this will totally disable SGX (including virtualization) if launch control is unavailable.
To me it is better to be here, since
clear_sgx_caps(), which disables SGX totally, should logically clear all SGX feature bits, no matter later patch's behavior. So when new SGX bits are introduced,clear_sgx_caps()should clear them too. Otherwise the logic of this patch (adding new SGX feature bits) is not complete IMHO.Quoted 5 lines
And actually in later patch "Allow SGX virtualization without Launch Control support", a new
clear_sgx_lc()is added, and is called when LC is not available but SGX virtualization is enabled, to make sure onlySGX_LCbit is cleared in this case. I don't quite understand why we need to clear SGX1 and SGX2 inclear_sgx_caps()after the later patch.
Quoted 6 lines
I was talking about patch ordering. It could be argued that this goes after the content of patch 05/23. Please _consider_ changing the ordering.
If that doesn't work for some reason, please at least call out in the changelog that it leaves a temporarily funky situation.
The later patch currently uses SGX1 bit, which is the reason that this patch needs be before later patch.
Sean,
I think it is OK to remove SGX1 bit check in later patch, since I have never seen a machine with SGX bit in CPUID, but w/o SGX1. If we remove SGX1 bit check in later, we can put this patch after the later patch.
Do you have comment here? If you are OK, I'll remove SGX1 bit check in later patch and reorder the patch.
Re: [RFC PATCH 11/23] x86/sgx: Add helpers to expose ECREATE and EINIT to KVM
In reply to Dave HansenQuoted 24 lines · On Wed, 6 Jan 2021 13:23:37 -0800 Dave Hansen wrote:
Quoted 15 lines · Sean Christopherson
Quoted 11 lines · Dave Hansen
Quoted 7 lines · Kai Huang
From: Sean Christopherson <[email protected]>
Provide wrappers around
__ecreate()and__einit()to hide the ugliness of overloading the ENCLS return value to encode multiple error formats in a single int. KVM will trap-and-execute ECREATE and EINIT as part of SGX virtualization, and on an exception, KVM needs the trapnr so that it can inject the correct fault into the guest.This is missing a bit of a step about how and why ECREATE needs to be run in the host in the first place.
There's (hopefully) good info in the KVM usage patch that can be borrowed:
Quoted 7 lines
Add an ECREATE handler that will be used to intercept ECREATE for the purpose of enforcing and enclave's MISCSELECT, ATTRIBUTES and XFRM, i.e. to allow userspace to restrict SGX features via CPUID. ECREATE will be intercepted when any of the aforementioned masks diverges from hardware in order to enforce the desired CPUID model, i.e. inject #GP if the guest attempts to set a bit that hasn't been enumerated as allowed-1 in CPUID.
Quoted 4 lines
OK, so in plain language: the bare-metal kernel must intercept ECREATE to be able to impose policies on guests. When it does this, the bare-metal kernel runs ECREATE against the userspace mapping of the virtualized EPC.
Thanks. I'll add this to commit message.
Quoted 53 lines
Quoted 45 lines
Quoted 45 lines
Quoted 30 lines
arch/x86/include/asm/sgx.h new file +16
-0@@ -0,0 +1,16 @@ +/* SPDX-License-Identifier: GPL-2.0 */ +#ifndef _ASM_X86_SGX_H +#define _ASM_X86_SGX_H + +#include <linux/types.h> + +#ifdef CONFIG_X86_SGX_VIRTUALIZATION +struct sgx_pageinfo ; + +int sgx_virt_ecreate (struct sgx_pageinfo *pageinfo,void __user *secs, +int *trapnr); +int sgx_virt_einit (void __user *sigstruct,void __user *token, +void __user *secs,u64 *lepubkeyhash,int *trapnr); +#endif + +#endif /* _ASM_X86_SGX_H */ arch/x86/kernel/cpu/sgx/virt.c +0
-0@@ -261,3 +261,58 @@ int __init sgx_virt_epc_init(void)return misc_register (&sgx_virt_epc_dev);arch/x86/kernel/cpu/sgx/virt.c int __init sgx_virt_epc_init(void) · line 272
Quoted 10 lines
arch/x86/kernel/cpu/sgx/virt.c +9
-0} + +int sgx_virt_ecreate (struct sgx_pageinfo *pageinfo,void __user *secs, +int *trapnr) +{ +int ret; + +__uaccess_begin (); + ret =__ecreate (pageinfo, (void *)secs); +__uaccess_end ();The __uaccess_begin/end() worries me. There are *very* few of these in the kernel and it seems like something we want to use as sparingly as possible.
Quoted 4 lines
Quoted 2 lines
Why don't we just use the kernel mapping for 'secs' and not have to deal with stac/clac?
The kernel mapping isn't readily available.
Oh, duh. There's no kernel mapping for EPC... it's not RAM in the first place.
Quoted 8 lines
Quoted 6 lines
At this point, it's not even
guaranteed that @secs points at an EPC page. Unlike the driver code, where the EPC page is allocated on-demand by the kernel, the pointer here is userspace (technically guest) controlled. The caller (KVM) is responsible for ensuring it's a valid userspace address, but the SGX/EPC specific checks are mostly deferred to hardware.Ahh, got it. Kai, could we get some of this into comments or the changelog?
Yes I'll add some into comments.
Re: [RFC PATCH 04/23] x86/cpufeatures: Add SGX1 and SGX2 sub-features
In reply to Borislav PetkovQuoted 8 lines · On Wed, 6 Jan 2021 23:15:27 +0100 Borislav Petkov wrote:
arch/x86/include/asm/cpufeatures.h line 388
Quoted 3 lines · Kai Huang
arch/x86/include/asm/cpufeatures.h +3
-0+/* Intel-defined SGX features, CPUID level 0x00000012:0 (EAX), word 19 */ +#define X86_FEATURE_SGX1 (19*32+ 0)/* SGX1 leaf functions */ +#define X86_FEATURE_SGX2 (19*32+ 1)/* SGX2 leaf functions */ Is anything else from that leaf going to be added later? Bit 5 is "supports ENCLV instruction leaves", 6 is ENCLS insn leaves... are those going to be used in the kernel too eventually?
Bit 5 and Bit 6 are related to reclaiming EPC page from SGX guest, and the mechanism behind the two bits are only supposed to be used by KVM.
There's no urgent request to support them for now (and given basic SGX virtualization is not in upstream), but I don't know whether they need to be supported in the future.
Quoted 9 lines
Rest of them is reserved in the SDM which probably means internal only for now.
Signature · 4 lines
Regards/Gruss, Boris. https://people.kernel.org/tglx/notes-about-netiquette
Sean Christopherson
maintainerRe: [RFC PATCH 04/23] x86/cpufeatures: Add SGX1 and SGX2 sub-features
In reply to Kai HuangQuoted 36 lines · Kai Huang
Quoted 28 lines · On Wed, 6 Jan 2021 14:21:39 -0800 Dave Hansen wrote:
Quoted 21 lines · Kai Huang
Quoted 14 lines · On Wed, 6 Jan 2021 11:39:46 -0800 Dave Hansen wrote:
arch/x86/kernel/cpu/feat_ctl.c static void clear_sgx_caps(void) · line 102
Quoted 9 lines · Kai Huang
arch/x86/kernel/cpu/feat_ctl.c +2
-0@@ -97,6 +97,8 @@ static void clear_sgx_caps(void) {setup_clear_cpu_cap (X86_FEATURE_SGX );setup_clear_cpu_cap (X86_FEATURE_SGX_LC ); +setup_clear_cpu_cap (X86_FEATURE_SGX1 ); +setup_clear_cpu_cap (X86_FEATURE_SGX2 ); }Logically, I think you want this *after* the "Allow SGX virtualization without Launch Control support" patch. As it stands, this will totally disable SGX (including virtualization) if launch control is unavailable.
To me it is better to be here, since
clear_sgx_caps(), which disables SGX totally, should logically clear all SGX feature bits, no matter later patch's behavior. So when new SGX bits are introduced,clear_sgx_caps()should clear them too. Otherwise the logic of this patch (adding new SGX feature bits) is not complete IMHO.Quoted 5 lines
And actually in later patch "Allow SGX virtualization without Launch Control support", a new
clear_sgx_lc()is added, and is called when LC is not available but SGX virtualization is enabled, to make sure onlySGX_LCbit is cleared in this case. I don't quite understand why we need to clear SGX1 and SGX2 inclear_sgx_caps()after the later patch.Quoted 6 lines
I was talking about patch ordering. It could be argued that this goes after the content of patch 05/23. Please _consider_ changing the ordering.
If that doesn't work for some reason, please at least call out in the changelog that it leaves a temporarily funky situation.
Quoted 7 lines
The later patch currently uses SGX1 bit, which is the reason that this patch needs be before later patch.
Sean,
I think it is OK to remove SGX1 bit check in later patch, since I have never seen a machine with SGX bit in CPUID, but w/o SGX1.
The SGX1 check is "needed" to handle the case where SGX is supported but was soft-disabled, e.g. because software disable a machine check bank by writing an MCi_CTL MSR.
Quoted 5 lines
If we remove SGX1 bit check in later, we can put this patch after the later patch.
Do you have comment here? If you are OK, I'll remove SGX1 bit check in later patch and reorder the patch.
Hmm, I'm not sure why the SGX driver was merged without explicitly checking for SGX1 support. I'm pretty sure we had an explicit SGX1 check in the driver path at some point. My guess is that the SGX1 change ended up in the KVM series through a mishandled rebase.
Moving the check later won't break anything that's not already broken. But, arguably checking SGX1 is a bug fix of sorts, e.g. to guard against broken firmware, and should go in as a standalone patch destined for stable. The kernel can't prevent SGX from being soft-disabled after boot, but IMO it should cleanly handle the case where SGX was soft-disabled _before_ boot.
Re: [RFC PATCH 04/23] x86/cpufeatures: Add SGX1 and SGX2 sub-features
Quoted 9 lines · Sean Christopherson
Quoted 5 lines
If we remove SGX1 bit check in later, we can put this patch after the later patch.
Do you have comment here? If you are OK, I'll remove SGX1 bit check in later patch and reorder the patch.
Hmm, I'm not sure why the SGX driver was merged without explicitly checking for SGX1 support. I'm pretty sure we had an explicit SGX1 check in the driver path at some point. My guess is that the SGX1 change ended up in the KVM series through a mishandled rebase.
There was one, but I think it got removed when I asked that the X86_FEATURE_SGX1/2 bits be removed. I actually even mentioned checking the CPUID leaf directly with cpuid...() in initialization. But I missed when that was never done.
It's not a practical problem, but I do agree we should fix it up for 5.10 stable.
Re: [RFC PATCH 04/23] x86/cpufeatures: Add SGX1 and SGX2 sub-features
In reply to Dave HansenQuoted 27 lines · On Wed, 6 Jan 2021 14:21:39 -0800 Dave Hansen wrote:
Quoted 20 lines · Kai Huang
Quoted 13 lines · On Wed, 6 Jan 2021 11:39:46 -0800 Dave Hansen wrote:
arch/x86/kernel/cpu/feat_ctl.c static void clear_sgx_caps(void) · line 102
Quoted 9 lines · Kai Huang
arch/x86/kernel/cpu/feat_ctl.c +2
-0@@ -97,6 +97,8 @@ static void clear_sgx_caps(void) {setup_clear_cpu_cap (X86_FEATURE_SGX );setup_clear_cpu_cap (X86_FEATURE_SGX_LC ); +setup_clear_cpu_cap (X86_FEATURE_SGX1 ); +setup_clear_cpu_cap (X86_FEATURE_SGX2 ); }Logically, I think you want this *after* the "Allow SGX virtualization without Launch Control support" patch. As it stands, this will totally disable SGX (including virtualization) if launch control is unavailable.
To me it is better to be here, since
clear_sgx_caps(), which disables SGX totally, should logically clear all SGX feature bits, no matter later patch's behavior. So when new SGX bits are introduced,clear_sgx_caps()should clear them too. Otherwise the logic of this patch (adding new SGX feature bits) is not complete IMHO.Quoted 5 lines
And actually in later patch "Allow SGX virtualization without Launch Control support", a new
clear_sgx_lc()is added, and is called when LC is not available but SGX virtualization is enabled, to make sure onlySGX_LCbit is cleared in this case. I don't quite understand why we need to clear SGX1 and SGX2 inclear_sgx_caps()after the later patch.
Quoted 7 lines
I was talking about patch ordering. It could be argued that this goes after the content of patch 05/23. Please _consider_ changing the ordering.
If that doesn't work for some reason, please at least call out in the changelog that it leaves a temporarily funky situation.
Hi Dave,
After second thinking, if I understand you correctly, the "funky situation" you are talking about is, w/o patch "Allow SGX virtualization without Launch Control Support", SGX virtualization is disabled too if LC is not available in hardware, but in previous patches (basically patch 3 "Introduce virtual EPC for use by KVM guests"), we have been treating SGX virtualization can be enabled?
In this case, clearing SGX1 and SGX2 bits before or after "Allow SGX virtualization without Launch Control support" patch doesn't make difference, since KVM should always check SGX bit first.
So a better way is to put "Allow SGX virtualization without Launch Control Support" at the beginning of this series? If so, the Kconfig X86_SGX_VIRTUALIZATION needs to be in separate patch at the very beginning.
Does above make sense?
Re: [RFC PATCH 04/23] x86/cpufeatures: Add SGX1 and SGX2 sub-features
Quoted 5 lines · Kai Huang
So a better way is to put "Allow SGX virtualization without Launch Control Support" at the beginning of this series? If so, the
KconfigX86_SGX_VIRTUALIZATIONneeds to be in separate patch at the very beginning.Does above make sense?
I think it's worth trying. No promises that anyone will like the end result, but give it a shot and I'll take a look.
Re: [RFC PATCH 04/23] x86/cpufeatures: Add SGX1 and SGX2 sub-features
In reply to Sean ChristophersonQuoted 60 lines · On Wed, 6 Jan 2021 15:19:41 -0800 Sean Christopherson wrote:
Quoted 36 lines · Kai Huang
Quoted 28 lines · On Wed, 6 Jan 2021 14:21:39 -0800 Dave Hansen wrote:
Quoted 21 lines · Kai Huang
Quoted 14 lines · On Wed, 6 Jan 2021 11:39:46 -0800 Dave Hansen wrote:
arch/x86/kernel/cpu/feat_ctl.c static void clear_sgx_caps(void) · line 102
Quoted 9 lines · Kai Huang
arch/x86/kernel/cpu/feat_ctl.c +2
-0@@ -97,6 +97,8 @@ static void clear_sgx_caps(void) {setup_clear_cpu_cap (X86_FEATURE_SGX );setup_clear_cpu_cap (X86_FEATURE_SGX_LC ); +setup_clear_cpu_cap (X86_FEATURE_SGX1 ); +setup_clear_cpu_cap (X86_FEATURE_SGX2 ); }Logically, I think you want this *after* the "Allow SGX virtualization without Launch Control support" patch. As it stands, this will totally disable SGX (including virtualization) if launch control is unavailable.
To me it is better to be here, since
clear_sgx_caps(), which disables SGX totally, should logically clear all SGX feature bits, no matter later patch's behavior. So when new SGX bits are introduced,clear_sgx_caps()should clear them too. Otherwise the logic of this patch (adding new SGX feature bits) is not complete IMHO.Quoted 5 lines
And actually in later patch "Allow SGX virtualization without Launch Control support", a new
clear_sgx_lc()is added, and is called when LC is not available but SGX virtualization is enabled, to make sure onlySGX_LCbit is cleared in this case. I don't quite understand why we need to clear SGX1 and SGX2 inclear_sgx_caps()after the later patch.Quoted 6 lines
I was talking about patch ordering. It could be argued that this goes after the content of patch 05/23. Please _consider_ changing the ordering.
If that doesn't work for some reason, please at least call out in the changelog that it leaves a temporarily funky situation.
Quoted 7 lines
The later patch currently uses SGX1 bit, which is the reason that this patch needs be before later patch.
Sean,
I think it is OK to remove SGX1 bit check in later patch, since I have never seen a machine with SGX bit in CPUID, but w/o SGX1.
The SGX1 check is "needed" to handle the case where SGX is supported but was soft-disabled, e.g. because software disable a machine check bank by writing an
MCi_CTLMSR.Quoted 5 lines
If we remove SGX1 bit check in later, we can put this patch after the later patch.
Do you have comment here? If you are OK, I'll remove SGX1 bit check in later patch and reorder the patch.
Hmm, I'm not sure why the SGX driver was merged without explicitly checking for SGX1 support. I'm pretty sure we had an explicit SGX1 check in the driver path at some point. My guess is that the SGX1 change ended up in the KVM series through a mishandled rebase.
Quoted 5 lines
Moving the check later won't break anything that's not already broken. But, arguably checking SGX1 is a bug fix of sorts, e.g. to guard against broken firmware, and should go in as a standalone patch destined for stable. The kernel can't prevent SGX from being soft-disabled after boot, but IMO it should cleanly handle the case where SGX was soft-disabled _before_ boot.
It seems I need to dig some history. Thanks Sean for the info!
Re: [RFC PATCH 04/23] x86/cpufeatures: Add SGX1 and SGX2 sub-features
In reply to Dave HansenQuoted 9 lines · On Wed, 6 Jan 2021 15:43:54 -0800 Dave Hansen wrote:
Quoted 5 lines · Kai Huang
So a better way is to put "Allow SGX virtualization without Launch Control Support" at the beginning of this series? If so, the
KconfigX86_SGX_VIRTUALIZATIONneeds to be in separate patch at the very beginning.Does above make sense?
I think it's worth trying. No promises that anyone will like the end result, but give it a shot and I'll take a look.
OK I'll try in next version. Thanks.
Quoted 30 lines · On Wed, 6 Jan 2021 09:07:13 -0800 Dave Hansen wrote:
Quoted 7 lines · Kai Huang
- Virtual EPC
"Virtual EPC" is the EPC section exposed by KVM to guest so SGX software in guest can discover it and use it to create SGX enclaves. KVM exposes SGX to guest via CPUID, and exposes one or more "virtual EPC" sections for guest. The size of "virtual EPC" is passed as Qemu parameter when creating the guest, and the base address is calcualted internally according to guest's
^ calculated
Quoted 1 line
configuration.
This is not a great first paragraph to introduce me to this feature.
Please remind us what EPC *is*, then you can go and talk about why we have to virtualize it, and how "virtual EPC" is different from normal EPC. For instance:
SGX enclave memory is special and is reserved specifically for enclave use. In bare-metal SGX enclaves, the kernel allocates enclave pages, copies data into the pages with privileged instructions, then allows the enclave to start. In this scenario, only initialized pages already assigned to an enclave are mapped to userspace.
In virtualized environments, the hypervisor still needs to do the physical enclave page allocation. The guest kernel is responsible for the data copying (among other things). This means that the job of starting an enclave is now split between hypervisor and guest.
Quoted 7 lines
This series introduces a new misc device:
/dev/sgx_virt_epc. This device allows the host to map *uninitialized* enclave memory into userspace, which can then be passed into a guest.While it might be *possible* to start a host-side enclave with
/dev/sgx_enclaveand pass its memory into a guest, it would be wasteful and convoluted.
Thanks. I'll add this.
Quoted 9 lines
Quoted 5 lines
core/driver to allow userspace (Qemu) to allocate "raw" EPC, and use it as "virtual EPC" for guest. Obviously, unlike EPC allocated for host SGX driver, virtual EPC allocated via
/dev/sgx_virt_epcdoesn't have enclave associated, and how virtual EPC is used by guest is compeletely controlled by guest's SGX^ completely
Please run a spell checker on this thing.
Yeah will do. Thanks for good suggestion.
Quoted 21 lines
Quoted 14 lines
software.
Implement the "raw" EPC allocation in the x86 core-SGX subsystem via
/dev/sgx_virt_epcrather than in KVM. Doing so has two major advantages:- Does not require changes to KVM's uAPI, e.g. EPC gets handled as just another memory backend for guests.
- EPC management is wholly contained in the SGX subsystem, e.g. SGX does not have to export any symbols, changes to reclaim flows don't need to be routed through KVM, SGX's dirty laundry doesn't have to get aired out for the world to see, and so on and so forth.
Quoted 6 lines
The virtual EPC allocated to guests is currently not reclaimable, due to reclaiming EPC from KVM guests is not currently supported. Due to the complications of handling reclaim conflicts between guest and host, KVM EPC oversubscription, which allows total virtual EPC size greater than physical EPC by being able to reclaiming guests' EPC, is significantly more complex than basic support for SGX virtualization.
Quoted 3 lines
It would also help here to remind the reader that enclave pages have a special reclaim mechanism separtae from normal page reclaim, and that mechanism is disabled for these pages.
OK.
Quoted 2 lines
Does the *ABI* here preclude doing oversubscription in the future?
I am Sorry what *ABI* do you mean?
Quoted 6 lines
Quoted 4 lines
- Support SGX virtualization without SGX Launch Control unlocked mode
Although SGX driver requires SGX Launch Control unlocked mode to work, SGX
Although the bare-metal SGX driver requires...
OK.
Quoted 2 lines
Also, didn't we call this "Flexible Launch Control"?
I am actually a little bit confused about all those terms here. I don't think from spec's perspective, there's such thing "Flexible Launch Control", but I think everyone knows what does it mean. But I am not sure whether it is commonly used by community.
I think using FLC is fine if we only want to mention unlocked mode. But if you want to mention both, IMHO it would be better to specifically use LC locked mode and unlocked mode, since technically there's third case that LC is not present at all.
Quoted 6 lines
Quoted 4 lines
virtualization doesn't, since how enclave is created is completely controlled by guest SGX software, which is not necessarily linux. Therefore, this series allows KVM to expose SGX to guest even SGX Launch Control is in locked mode,
... "expose SGX to guests even if" ...
Thanks.
Quoted 7 lines
Quoted 6 lines
or is not present at all. The reason is the goal of SGX virtualization, or virtualization in general, is to expose hardware feature to guest, but not to make assumption how guest will use it. Therefore, KVM should support SGX guest as long as hardware is able to, to have chance to support more potential use cases in cloud environment.
Quoted 5 lines
This is kinda long-winded and misses a lot of important context. How about:
SGX hardware supports two "launch control" modes to limit which enclaves can run. In the "locked" mode, the hardware prevents enclaves from running unless they are blessed by a third party.
or "by Intel".
In the unlocked mode,
Quoted 6 lines
the kernel is in full control of which enclaves can run. The bare-metal SGX code refuses to launch enclaves unless it is in the unlocked mode.
This
sgx_virt_epcdriver does not have such a restriction. This allows guests which are OK with the locked mode to use SGX, even if the host kernel refuses to.
Indeed better. Thanks a lot.
Quoted 8 lines
Quoted 8 lines
- Support exposing SGX2
Due to the same reason above, SGX2 feature detection is added to core SGX code to allow KVM to expose SGX2 to guest, even currently SGX driver doesn't support SGX2, because SGX2 can work just fine in guest w/o any interaction to host SGX driver.
Quoted 6 lines
Quoted 4 lines
- Restricit SGX guest access to provisioning key
To grant guest being able to fully use SGX, guest needs to be able to create provisioning enclave.
"enclave" or "enclaves"?
I think should be "enclave", inside one VM, there should only be one provisioning enclave.
Quoted 4 lines
Quoted 2 lines
However provisioning key is sensitive and is restricted by
^ the
Thanks.
Quoted 7 lines
Quoted 6 lines
/dev/sgx_provisionin host SGX driver, therefore KVM SGX virtualization follows the same role: a newKVM_CAP_SGX_ATTRIBUTEis added to KVM uAPI, and only file descriptor of/dev/sgx_provisionis passed to that CAP by usersppace hypervisor (Qemu) when creating the guest, it can access provisioning bit. This is done by making KVM trape ECREATE instruction from guest, and check the provisioning bit
Quoted 6 lines
^ trap
Quoted 1 line
in ECREATE's attribute.
The grammar in that paragraph is really off to me. Can you give it another go?
I'll refine it. Thanks a lot for input.
Re: [RFC PATCH 03/23] x86/sgx: Introduce virtual EPC for use by KVM guests
In reply to Sean ChristophersonQuoted 10 lines
Quoted 5 lines
arch/x86/kernel/cpu/sgx/virt.c line 25
Quoted 3 lines
arch/x86/kernel/cpu/sgx/virt.c new file +2
-0+static struct mutex virt_epc_lock; +static struct list_head virt_epc_zombie_pages;What does the lock protect?
Effectively, the list of zombie SECS pages. Not sure why I used a generic name.
Quoted 1 line
What are zombie pages?
Quoted 8 lines
My own terminology for SECS pages whose virtual EPC has been destroyed but can't be reclaimed due to them having child EPC pages in other virtual EPCs.
Quoted 2 lines
BTW, if zombies are SECS-only, shouldn't that be in the name rather than "epc"?
I used the
virt_epcprefix/namespace to tag it as a global list. I've no argument against something likezombie_secs_pages.
I'll change to zombie_secs_pages, and lock name to zombie_secs_pages_lock, respectively.
[...]
Quoted 11 lines
Quoted 8 lines
arch/x86/kernel/cpu/sgx/virt.c line 120
Quoted 6 lines
arch/x86/kernel/cpu/sgx/virt.c new file +6
-0+static int sgx_virt_epc_free_page (struct sgx_epc_page *epc_page) +{ +int ret; + +if (!epc_page) +return 0 ;I always worry about these. Why is passing NULL around OK?
I suspect I did it to mimic
kfree()behavior. I don't _think_ the radix (now xarray) usage will ever encounter a NULL entry.
I'll remove the NULL page check.
Quoted 18 lines
Quoted 18 lines
Quoted 4 lines
arch/x86/kernel/cpu/sgx/virt.c new file +2
-0+ ret =__eremove (sgx_get_epc_virt_addr (epc_page)); +if (ret) {arch/x86/kernel/cpu/sgx/virt.c line 143
Quoted 10 lines
arch/x86/kernel/cpu/sgx/virt.c new file +10
-0+/* +* Only SGX_CHILD_PRESENT is expected, which is because of +* EREMOVE-ing an SECS still with child, in which case it can +* be handled by EREMOVE-ing the SECS again after all pages in +* virtual EPC have been EREMOVE-ed. See comments in below in +* sgx_virt_epc_release(). +*/ +WARN_ON_ONCE (ret !=SGX_CHILD_PRESENT ); +return ret; + }I find myself wondering what errors could cause the
WARN_ON_ONCE()to be hit. The SDM indicates that it's only:
Quoted 8 lines
Quoted 4 lines
SGX_ENCLAVE_ACTIf there are still logical processors executing
inside the enclave.Should that be mentioned in the comment?
And faults, which are also spliced into the return value by the ENCLS macros. I do remember hitting this WARN when I broke things, though I can't remember whether it was a fault or the
SGX_ENCLAVE_ACTscenario. Probably the latter?
I'll add a comment saying that there should be no active logical processor still running inside guest's enclave. We cannot handle SGX_ENCLAVE_ACT here anyway.
Quoted 4 lines · Kai Huang
Quoted 1 line · On Wed, 6 Jan 2021 09:07:13 -0800 Dave Hansen wrote:
Does the *ABI* here preclude doing oversubscription in the future?
I am Sorry what *ABI* do you mean?
Oh boy.
https://en.wikipedia.org/wiki/Application_binary_interface
In your patch set that you are posting, /dev/sgx_virt_epc is a new
interface: a new ABI. If we accept your contribution, programs will be build around and expect Linux to support this ABI. An ABI is a contract between software written to use it and the kernel. The kernel tries *really* hard to never break its contracts with applications.
OK, now that we have that out of the way, I'll ask my question in another way:
Your series adds some new interfaces, including /dev/sgx_virt_epc. If the kernel wants to add oversubscription in the future, will old binary application users of /dev/sgx_virt_epc be able to support oversubscription? Or, would users of /dev/sgx_virt_epc need to change to support oversubscription?
Quoted 11 lines
Quoted 1 line
Also, didn't we call this "Flexible Launch Control"?
I am actually a little bit confused about all those terms here. I don't think from spec's perspective, there's such thing "Flexible Launch Control", but I think everyone knows what does it mean. But I am not sure whether it is commonly used by community.
I think using FLC is fine if we only want to mention unlocked mode. But if you want to mention both, IMHO it would be better to specifically use LC locked mode and unlocked mode, since technically there's third case that LC is not present at all.
Could you go over the changelogs from Jarkko's patches and at least make these consistent with those?
Quoted 6 lines
Quoted 6 lines
Quoted 5 lines
or is not present at all. The reason is the goal of SGX virtualization, or virtualization in general, is to expose hardware feature to guest, but not to make assumption how guest will use it. Therefore, KVM should support SGX guest as long as hardware is able to, to have chance to support more potential use cases in cloud environment.
Quoted 7 lines
Quoted 5 lines
This is kinda long-winded and misses a lot of important context. How about:
SGX hardware supports two "launch control" modes to limit which enclaves can run. In the "locked" mode, the hardware prevents enclaves from running unless they are blessed by a third party.
or "by Intel".
From what I understand, Intel had to bless the enclaves but the architecture itself doesn't say "Intel must bless them". But, yeah, in practice, it had to be Intel.
Quoted 9 lines
Quoted 9 lines
Quoted 9 lines
- Support exposing SGX2
Due to the same reason above, SGX2 feature detection is added to core SGX code to allow KVM to expose SGX2 to guest, even currently SGX driver doesn't support SGX2, because SGX2 can work just fine in guest w/o any interaction to host SGX driver.
- Restricit SGX guest access to provisioning key
Quoted 7 lines
Quoted 4 lines
Quoted 2 lines
To grant guest being able to fully use SGX, guest needs to be able to create provisioning enclave.
"enclave" or "enclaves"?
I think should be "enclave", inside one VM, there should only be one provisioning enclave.
This is where the language becomes important. Is the provisioning enclave a one-shot deal? You create one per guest and can never create another? Or, can you restart it? Can you architecturally have more than one active at once? Or, can you only create one once the first one dies?
You'll write that sentence differently based on the answers.
Re: [RFC PATCH 03/23] x86/sgx: Introduce virtual EPC for use by KVM guests
In reply to Kai HuangQuoted 18 lines · Kai Huang
Quoted 18 lines
Quoted 18 lines
arch/x86/kernel/cpu/sgx/virt.c line 143
Quoted 12 lines
arch/x86/kernel/cpu/sgx/virt.c new file +12
-0+ ret =__eremove (sgx_get_epc_virt_addr (epc_page)); +if (ret) { +/* +* Only SGX_CHILD_PRESENT is expected, which is because of +* EREMOVE-ing an SECS still with child, in which case it can +* be handled by EREMOVE-ing the SECS again after all pages in +* virtual EPC have been EREMOVE-ed. See comments in below in +* sgx_virt_epc_release(). +*/ +WARN_ON_ONCE (ret !=SGX_CHILD_PRESENT ); +return ret; + }I find myself wondering what errors could cause the
WARN_ON_ONCE()to be hit. The SDM indicates that it's only:
SGX_ENCLAVE_ACTIf there are still logical processors executing
inside the enclave.
Quoted 7 lines
Quoted 4 lines
Quoted 1 line
Should that be mentioned in the comment?
And faults, which are also spliced into the return value by the ENCLS macros. I do remember hitting this WARN when I broke things, though I can't remember whether it was a fault or the
SGX_ENCLAVE_ACTscenario. Probably the latter?I'll add a comment saying that there should be no active logical processor still running inside guest's enclave. We cannot handle
SGX_ENCLAVE_ACThere anyway.
One more thing...
Could we dump out the *actual* error code with a WARN(), please? If we see a warning, I'd rather not have to disassemble the instructions and check against register values to see whether the error code was sane.
Re: [RFC PATCH 03/23] x86/sgx: Introduce virtual EPC for use by KVM guests
Quoted 27 lines · On Wed, 6 Jan 2021 16:52:49 -0800 Dave Hansen wrote:
Quoted 18 lines · Kai Huang
Quoted 18 lines
Quoted 18 lines
arch/x86/kernel/cpu/sgx/virt.c line 143
Quoted 12 lines
arch/x86/kernel/cpu/sgx/virt.c new file +12
-0+ ret =__eremove (sgx_get_epc_virt_addr (epc_page)); +if (ret) { +/* +* Only SGX_CHILD_PRESENT is expected, which is because of +* EREMOVE-ing an SECS still with child, in which case it can +* be handled by EREMOVE-ing the SECS again after all pages in +* virtual EPC have been EREMOVE-ed. See comments in below in +* sgx_virt_epc_release(). +*/ +WARN_ON_ONCE (ret !=SGX_CHILD_PRESENT ); +return ret; + }I find myself wondering what errors could cause the
WARN_ON_ONCE()to be hit. The SDM indicates that it's only:
SGX_ENCLAVE_ACTIf there are still logical processors executing
inside the enclave.Quoted 7 lines
Quoted 4 lines
Quoted 1 line
Should that be mentioned in the comment?
And faults, which are also spliced into the return value by the ENCLS macros. I do remember hitting this WARN when I broke things, though I can't remember whether it was a fault or the
SGX_ENCLAVE_ACTscenario. Probably the latter?I'll add a comment saying that there should be no active logical processor still running inside guest's enclave. We cannot handle
SGX_ENCLAVE_ACThere anyway.
Quoted 5 lines
One more thing...
Could we dump out the *actual* error code with a
WARN(), please? If we see a warning, I'd rather not have to disassemble the instructions and check against register values to see whether the error code was sane.
Sure. But WARN_ONCE() should be used, right, instead of WARN()?
Re: [RFC PATCH 03/23] x86/sgx: Introduce virtual EPC for use by KVM guests
In reply to Dave HansenQuoted 13 lines · On Wed, 6 Jan 2021 11:35:41 -0800 Dave Hansen wrote:
Quoted 11 lines · Kai Huang
From: Sean Christopherson <[email protected]>
Add a misc device
/dev/sgx_virt_epcto allow userspace to allocate "raw" EPC without an associated enclave. The intended and only known use case for raw EPC allocation is to expose EPC to a KVM guest, hence thevirt_epcmoniker, virt.{c,h} files andX86_SGX_VIRTUALIZATIONKconfig.Modify
sgx_init()to always try to initialize virtual EPC driver, even when SGX driver is disabled due to SGX Launch Control is in locked mode, or not present at all, since SGX virtualization allows to expose SGX to guests that support non-LC configurations.
Quoted 7 lines
The grammar here is a bit off. Here's a rewrite:
Modify
sgx_init()to always try to initialize the virtual EPC driver, even if the bare-metal SGX driver is disabled. The bare-metal driver might be disabled if SGX Launch Control is in locked mode, or not supported in the hardware at all. This allows (non-Linux) guests that support non-LC configurations to use SGX.
Thanks. I'll use yours, except I want to change "bare-metal driver might be disabled.." to "bare-metal driver will be disabled..".
I'll also use all your comments mentioned in your reply to this patch.
[...]
Quoted 8 lines
arch/x86/kernel/cpu/sgx/virt.c line 151
Quoted 4 lines
arch/x86/kernel/cpu/sgx/virt.c new file +4
-0+ +static int sgx_virt_epc_release (struct inode *inode,struct file *file) +{ +struct sgx_virt_epc *epc = file->private_data;FWIW, I hate the "struct
sgx_virt_epc*epc" name. "epc" here is really an instance
How about "struct sgx_virt_epc *vepc" ?
[...]
Quoted 27 lines
Quoted 11 lines
arch/x86/kernel/cpu/sgx/virt.c new file +11
-0+static int sgx_virt_epc_open (struct inode *inode,struct file *file) +{ +struct sgx_virt_epc *epc; + + epc =kzalloc (sizeof (struct sgx_virt_epc ),GFP_KERNEL ); +if (!epc) +return -ENOMEM ; +/* +* Keep the current->mm to virtual EPC. It will be checked in +* sgx_virt_epc_mmap() to prevent, in case of fork, child being +* able to mmap() to the same virtual EPC pages. arch/x86/kernel/cpu/sgx/virt.c line 241
Quoted 10 lines
arch/x86/kernel/cpu/sgx/virt.c new file +10
-0+*/ +mmgrab (current->mm); + epc->mm = current->mm; +mutex_init (&epc->lock); +xa_init (&epc->page_array); + + file->private_data = epc; + +return 0 ; +}I understand why this made sense for regular enclaves, but I'm having a harder time here. If you
mmap(fd, MAP_SHARED),fork(), and then pass that mapping through to two different guests, you get to hold the pieces, just like if you did the same with normal memory.
Quoted 1 line
Why does the kernel need to enforce this policy?
Does Sean's reply in another email satisfy you?
Quoted 8 lines · On Wed, 6 Jan 2021 16:48:58 -0800 Dave Hansen wrote:
Quoted 4 lines · Kai Huang
Quoted 1 line · On Wed, 6 Jan 2021 09:07:13 -0800 Dave Hansen wrote:
Does the *ABI* here preclude doing oversubscription in the future?
I am Sorry what *ABI* do you mean?
Oh boy.
Quoted 7 lines
https://en.wikipedia.org/wiki/Application_binary_interface
In your patch set that you are posting,
/dev/sgx_virt_epcis a new
interface: a new ABI. If we accept your contribution, programs will be build around and expect Linux to support this ABI. An ABI is a contract between software written to use it and the kernel. The kernel tries *really* hard to never break its contracts with applications.
Thanks.
Quoted 9 lines
OK, now that we have that out of the way, I'll ask my question in another way:
Your series adds some new interfaces, including
/dev/sgx_virt_epc. If the kernel wants to add oversubscription in the future, will old binary application users of/dev/sgx_virt_epcbe able to support oversubscription? Or, would users of/dev/sgx_virt_epcneed to change to support oversubscription?
Oversubscription will be completely done in kernel/kvm, and will be transparent to userspace, so it will not impact ABI.
Quoted 8 lines
Quoted 8 lines
Quoted 2 lines
Also, didn't we call this "Flexible Launch Control"?
I am actually a little bit confused about all those terms here. I don't think from spec's perspective, there's such thing "Flexible Launch Control", but I think everyone knows what does it mean. But I am not sure whether it is commonly used by community.
Quoted 7 lines
Quoted 4 lines
I think using FLC is fine if we only want to mention unlocked mode. But if you want to mention both, IMHO it would be better to specifically use LC locked mode and unlocked mode, since technically there's third case that LC is not present at all.
Could you go over the changelogs from Jarkko's patches and at least make these consistent with those?
I'll dig into them.
[...]
Quoted 10 lines
Quoted 9 lines
Quoted 6 lines
Quoted 4 lines
- Restricit SGX guest access to provisioning key
To grant guest being able to fully use SGX, guest needs to be able to create provisioning enclave.
"enclave" or "enclaves"?
I think should be "enclave", inside one VM, there should only be one provisioning enclave.
Quoted 8 lines
This is where the language becomes important. Is the provisioning enclave a one-shot deal? You create one per guest and can never create another? Or, can you restart it? Can you architecturally have more than one active at once? Or, can you only create one once the first one dies?
You'll write that sentence differently based on the answers.
I think I can just change to "guest needs to be able to access provisioning key". :)
Re: [RFC PATCH 03/23] x86/sgx: Introduce virtual EPC for use by KVM guests
In reply to Kai HuangQuoted 4 lines · Kai Huang
Quoted 3 lines
Could we dump out the *actual* error code with a
WARN(), please? If we see a warning, I'd rather not have to disassemble the instructions and check against register values to see whether the error code was sane.Sure. But
WARN_ONCE()should be used, right, instead ofWARN()?
Whatever will let you get a printf-format string out and only happens once.
Re: [RFC PATCH 03/23] x86/sgx: Introduce virtual EPC for use by KVM guests
In reply to Kai HuangQuoted 7 lines · Kai Huang
Quoted 6 lines
I understand why this made sense for regular enclaves, but I'm having a harder time here. If you
mmap(fd, MAP_SHARED),fork(), and then pass that mapping through to two different guests, you get to hold the pieces, just like if you did the same with normal memory.Why does the kernel need to enforce this policy?
Does Sean's reply in another email satisfy you?
I'm not totally convinced.
Please give it a go in the changelog for the next one and try to convince me that this is a good idea. Focus on what the downsides will be if the kernel does not enforce this policy. What will break, and why will it be bad? Why is the kernel in the best position to thwart the badness?
Borislav Petkov
maintainerRe: [RFC PATCH 04/23] x86/cpufeatures: Add SGX1 and SGX2 sub-features
In reply to Kai HuangQuoted 3 lines · Kai Huang
There's no urgent request to support them for now (and given basic SGX virtualization is not in upstream), but I don't know whether they need to be supported in the future.
If that is the case, then wasting a whole leaf for two bits doesn't make too much sense. And it looks like the kvm reverse lookup can be taught to deal with composing that leaf dynamically when needed instead.
Thx.
Signature · 4 lines
Regards/Gruss,
Boris.
https://people.kernel.org/tglx/notes-about-netiquetteSean Christopherson
maintainerQuoted 9 lines · Kai Huang
Quoted 5 lines · On Wed, 6 Jan 2021 16:48:58 -0800 Dave Hansen wrote:
Your series adds some new interfaces, including
/dev/sgx_virt_epc. If the kernel wants to add oversubscription in the future, will old binary application users of/dev/sgx_virt_epcbe able to support oversubscription? Or, would users of/dev/sgx_virt_epcneed to change to support oversubscription?Oversubscription will be completely done in kernel/kvm, and will be transparent to userspace, so it will not impact ABI.
It's not transparent to userpsace, odds are very good that userspace would want to opt in/out of EPC reclaim for its VMs. E.g. for cases where it would be preferable to fail to launch a VM than degrade performance.
That being said, there are no anticipated /dev/sgx_virt_epc ABI changes to support reclaim, as the ABI changes will be in KVM. In the KVM oversubscription POC, I added a KVM ioctl to allow enabling EPC reclaim/oversubscription. That ioctl took a fd for a /dev/sgx_virt_epc instance.
The reason for routing through KVM was to solve two dependencies issues:
- KVM needs a reference to the virt_epc instance to handle SGX_CONFLICT VM-Exits
- The SGX subsystem needs to be able to translate GPAs to HVAs to retrieve the SECS for a page it is reclaiming. That requires a KVM instance and helper function.
Routing the ioctl through KVM allows KVM to hand over a pointer of itself along with a GPA->HVA helper, and the SGX subsystem in turn can hand back the virt_epc instance resolved from the fd.
It would be possible to invert the flow, e.g. pass in a KVM fd to a new /dev/sgx_virt_epc ioctl, but I suspect that would be kludgier, and in any case it would be a new ioctl and so would not break existing users.
Re: [RFC PATCH 04/23] x86/cpufeatures: Add SGX1 and SGX2 sub-features
In reply to Borislav PetkovQuoted 8 lines · On Thu, 7 Jan 2021 07:41:25 +0100 Borislav Petkov wrote:
Quoted 3 lines · Kai Huang
There's no urgent request to support them for now (and given basic SGX virtualization is not in upstream), but I don't know whether they need to be supported in the future.
If that is the case, then wasting a whole leaf for two bits doesn't make too much sense. And it looks like the kvm reverse lookup can be taught to deal with composing that leaf dynamically when needed instead.
I am not sure changing reverse lookup to handle dynamic would be acceptable. To me it is ugly, and I don't have a first glance on how to do it. KVM can query host CPUID when dealing with SGX w/o X86_FEATURE_SGX1/2, but it is not as straightforward as having X86_FEATURE_SGX1/2.
And as Sean pointed out, SGX1 bit is also needed by both SGX driver and init_ia32_feat_ctl():
https://www.spinics.net/lists/kvm/msg231973.html
So having it would make things easier.
And regarding to other bits of this leaf, to me: 1) we cannot rule out possibility that bit 5 and bit 6 will be supported in the future; 2) I cannot talk more but we cannot rule out the possibility that there will be other bits introduced in the future.
Sean, what do you think?
Quoted 8 lines
Thx.
Signature · 4 lines
Regards/Gruss, Boris. https://people.kernel.org/tglx/notes-about-netiquette
Quoted 8 lines · On Thu, 7 Jan 2021 08:14:44 -0800 Sean Christopherson wrote:
Quoted 7 lines · Kai Huang
Quoted 5 lines · On Wed, 6 Jan 2021 16:48:58 -0800 Dave Hansen wrote:
Your series adds some new interfaces, including
/dev/sgx_virt_epc. If the kernel wants to add oversubscription in the future, will old binary application users of/dev/sgx_virt_epcbe able to support oversubscription? Or, would users of/dev/sgx_virt_epcneed to change to support oversubscription?
Quoted 6 lines
Quoted 2 lines
Oversubscription will be completely done in kernel/kvm, and will be transparent to userspace, so it will not impact ABI.
It's not transparent to userpsace, odds are very good that userspace would want to opt in/out of EPC reclaim for its VMs. E.g. for cases where it would be preferable to fail to launch a VM than degrade performance.
It seems reasonable use case, but I don't have immediate picture how it requires new ABI related to virtualization. For instance, SGX driver should expose sysfs saying how frequent the EPC swapping is (with KVM oversubscription, host SGX code should provide such info in whole I think), and cloud admin can determine whether to launch new VM.
Another argument is, theoretically, cloud admin may not know how EPC will be used in guest, so potentially guest will only use very little EPC, thus creating new VM won't hurt a lot, so I am not sure that, if we want to upstream KVM oversubscription one day, do we need to consider such case.
Quoted 5 lines
That being said, there are no anticipated
/dev/sgx_virt_epcABI changes to support reclaim, as the ABI changes will be in KVM. In the KVM oversubscription POC, I added a KVM ioctl to allow enabling EPC reclaim/oversubscription. That ioctl took a fd for a/dev/sgx_virt_epcinstance.
Adding IOCTL to enable/disable oversubscription for particular VM seems user-case dependent, and I am not sure whether we need to support that if we want to upstream oversubscription one day. To me, it makes sense to upstream *basic* oversubscription (which just supports reclaiming EPC from VM) first, and then we can extend if needed according to use cases.
Anyway, oversubscription won't break existing ABI as you mentioned.
Quoted 16 lines
The reason for routing through KVM was to solve two dependencies issues:
- KVM needs a reference to the
virt_epcinstance to handleSGX_CONFLICTVM-Exits- The SGX subsystem needs to be able to translate GPAs to HVAs to retrieve the SECS for a page it is reclaiming. That requires a KVM instance and helper function.
Routing the ioctl through KVM allows KVM to hand over a pointer of itself along with a
GPA->HVAhelper, and the SGX subsystem in turn can hand back thevirt_epcinstance resolved from the fd.It would be possible to invert the flow, e.g. pass in a KVM fd to a new
/dev/sgx_virt_epcioctl, but I suspect that would be kludgier, and in any case it would be a new ioctl and so would not break existing users.
Re: [RFC PATCH 04/23] x86/cpufeatures: Add SGX1 and SGX2 sub-features
In reply to Kai HuangQuoted 6 lines · Kai Huang
Quoted 5 lines · On Thu, 7 Jan 2021 07:41:25 +0100 Borislav Petkov wrote:
Quoted 3 lines · Kai Huang
There's no urgent request to support them for now (and given basic SGX virtualization is not in upstream), but I don't know whether they need to be supported in the future.
Quoted 8 lines
Quoted 3 lines
If that is the case, then wasting a whole leaf for two bits doesn't make too much sense. And it looks like the kvm reverse lookup can be taught to deal with composing that leaf dynamically when needed instead.
I am not sure changing reverse lookup to handle dynamic would be acceptable. To me it is ugly, and I don't have a first glance on how to do it. KVM can query host CPUID when dealing with SGX w/o X86_FEATURE_SGX1/2, but it is not as straightforward as having X86_FEATURE_SGX1/2.
So, Boris was pretty direct here. Could you please go spend a bit of time to see what it would take to make these dynamic? You can check what our (Intel) plans are for this leaf, but if it's going to remain sparsely-used, we need to look into making the leaves a bit more dynamic.
Quoted 4 lines
And regarding to other bits of this leaf, to me: 1) we cannot rule out possibility that bit 5 and bit 6 will be supported in the future; 2) I cannot talk more but we cannot rule out the possibility that there will be other bits introduced in the future.
From the Intel side, let's go look at the features that are coming. We have a list of CPUID bits that have been dedicated to future CPU features. Let's look if 1 of these bits or 30 is coming. I don't think it's exactly a state secret approximately how many CPUID bits we *think* will get used in this leaf.
We can't exactly put together a roadmap of bits, microarchitectures and chip release dates. But, we can at least say, "we have immediate plans for most of the leaf" or "we don't plan to fill up the leaf any time soon."
Re: [RFC PATCH 04/23] x86/cpufeatures: Add SGX1 and SGX2 sub-features
Quoted 16 lines · On Thu, 7 Jan 2021 21:10:29 -0800 Dave Hansen wrote:
Quoted 6 lines · Kai Huang
Quoted 5 lines · On Thu, 7 Jan 2021 07:41:25 +0100 Borislav Petkov wrote:
Quoted 3 lines · Kai Huang
There's no urgent request to support them for now (and given basic SGX virtualization is not in upstream), but I don't know whether they need to be supported in the future.
Quoted 8 lines
Quoted 3 lines
If that is the case, then wasting a whole leaf for two bits doesn't make too much sense. And it looks like the kvm reverse lookup can be taught to deal with composing that leaf dynamically when needed instead.
I am not sure changing reverse lookup to handle dynamic would be acceptable. To me it is ugly, and I don't have a first glance on how to do it. KVM can query host CPUID when dealing with SGX w/o X86_FEATURE_SGX1/2, but it is not as straightforward as having X86_FEATURE_SGX1/2.
Quoted 4 lines
So, Boris was pretty direct here. Could you please go spend a bit of time to see what it would take to make these dynamic? You can check what our (Intel) plans are for this leaf, but if it's going to remain sparsely-used, we need to look into making the leaves a bit more dynamic.
I don't think reverse lookup can be made dyanmic, but like I said if we don't have X86_FEATURE_SGX1/2, KVM needs to query raw CPUID when dealing with SGX.
The purpose of reverse lookup is to simplify KVM to have one common helper to check whether guest's CPUID has particular hardware feature bit or not. For instance, it changes guest_cpuid_has_xxx(cpuid) to guest_cpuid_has(cpuid, X86_FEATURE_xxx), so KVM can get rid of bunch of dedicated guest_cpuid_has_xxx() for each feature, but just use X86_FEATURE_xxx with one function. W/o X86_FEATURE_SGX1/2, when dealing with them, KVM needs to have dedicated functions but cannot use common one. That is a drawback for KVM.
Btw, one thing I forgot to say is with X86_FEATURE_SGX1/2, "sgx1" and "sgx2" will be in /proc/cpuinfo auatomatically. I think showing "sgx2" (and other future SGX features) in /proc/cpuinfo is helpful. W/o X86_FEATURE_SGX1/2, we need specific handling, if we want to show them in /proc/cpuinfo.
That being said, if all those doesn't convince Boris and you guys, and Sean has no say here, I'll remove X86_FEATURE_SGX1/2 in next version.
Quoted 15 lines
Quoted 5 lines
And regarding to other bits of this leaf, to me: 1) we cannot rule out possibility that bit 5 and bit 6 will be supported in the future; 2) I cannot talk more but we cannot rule out the possibility that there will be other bits introduced in the future.
From the Intel side, let's go look at the features that are coming. We have a list of CPUID bits that have been dedicated to future CPU features. Let's look if 1 of these bits or 30 is coming. I don't think it's exactly a state secret approximately how many CPUID bits we *think* will get used in this leaf.
We can't exactly put together a roadmap of bits, microarchitectures and chip release dates. But, we can at least say, "we have immediate plans for most of the leaf" or "we don't plan to fill up the leaf any time soon."
Borislav Petkov
maintainerRe: [RFC PATCH 04/23] x86/cpufeatures: Add SGX1 and SGX2 sub-features
Quoted 5 lines · Kai Huang
Quoted 5 lines
Quoted 4 lines
I am not sure changing reverse lookup to handle dynamic would be acceptable. To me it is ugly, and I don't have a first glance on how to do it. KVM can query host CPUID when dealing with SGX w/o X86_FEATURE_SGX1/2, but it is not as straightforward as having X86_FEATURE_SGX1/2.
Quoted 7 lines
Quoted 4 lines
So, Boris was pretty direct here. Could you please go spend a bit of time to see what it would take to make these dynamic? You can check what our (Intel) plans are for this leaf, but if it's going to remain sparsely-used, we need to look into making the leaves a bit more dynamic.
I don't think reverse lookup can be made dyanmic, but like I said if we don't have X86_FEATURE_SGX1/2, KVM needs to query raw CPUID when dealing with SGX.
How about before you go and say that "it is ugly" and "don't think can be made" you actually go and *really* try it first? Because actually trying is sometimes faster than trying to find arguments against it. :)
Because I just did it and unless I'm missing something obvious - I haven't actually tested it - this is not ugly at all and in the long run it will become one big switch-case, which is perfectly fine.
---
arch/x86/include/asm/cpufeature.h +1 -0
@@ -30,6 +30,7 @@ enum cpuid_leafsCPUID_7_ECX ,CPUID_8000_0007_EBX ,CPUID_7_EDX , + CPUID_12_EAX, /* used only by KVM for now */ };#ifdef CONFIG_X86_FEATURE_NAMES
arch/x86/include/asm/cpufeatures.h +2 -0
@@ -292,6 +292,8 @@#define X86_FEATURE_FENCE_SWAPGS_KERNEL (11*32+ 5)/* "" LFENCE in kernel entry SWAPGS path */ #define X86_FEATURE_SPLIT_LOCK_DETECT (11*32+ 6)/* #AC for split lock */ #define X86_FEATURE_PER_THREAD_MBA (11*32+ 7)/* "" Per-thread Memory Bandwidth Allocation */ +#define X86_FEATURE_SGX1 (11*32+ 8)/* SGX1 leaf functions */ +#define X86_FEATURE_SGX2 (11*32+ 9)/* SGX2 leaf functions */ /* Intel-defined CPU features, CPUID level 0x00000007:1 (EAX), word 12 */ #define X86_FEATURE_AVX512_BF16 (12*32+ 5)/* AVX512 BFLOAT16 instructions */
arch/x86/kvm/cpuid.h +27 -3
@@ -63,8 +63,27 @@ static const struct cpuid_reg reverse_cpuid[] = { [CPUID_8000_0007_EBX] = {0x80000007, 0, CPUID_EBX}, [CPUID_7_EDX] = { 7, 0, CPUID_EDX}, [CPUID_7_1_EAX] = { 7, 1, CPUID_EAX}, + [CPUID_12_EAX] = { 0x12, 0, CPUID_EAX}, }; +/* +* Map a synthetic X86_FEATURE bit definition to the corresponding bit in the +* hardware CPUID leaf. +*/ +static int map_synthetic_leaf (int x86_feature) +{ +switch (x86_feature) { +case X86_FEATURE_SGX1 :return BIT (0 ); +case X86_FEATURE_SGX2 :return BIT (1 ); +default : +break ; + } + +WARN_ON_ONCE (1 ); + +return 0 ; +} +/* * Reverse CPUID and its derivatives can only be used for hardware-defined * feature words, i.e. words whose bits directly correspond to a CPUID leaf.
@@ -78,7 +97,6 @@ static __always_inline void reverse_cpuid_check(unsigned int x86_leaf)BUILD_BUG_ON (x86_leaf ==CPUID_LNX_1 );BUILD_BUG_ON (x86_leaf ==CPUID_LNX_2 );BUILD_BUG_ON (x86_leaf ==CPUID_LNX_3 ); -BUILD_BUG_ON (x86_leaf ==CPUID_LNX_4 );BUILD_BUG_ON (x86_leaf >=ARRAY_SIZE (reverse_cpuid));BUILD_BUG_ON (reverse_cpuid[x86_leaf].function ==0 ); }
@@ -91,8 +109,14 @@ static __always_inline void reverse_cpuid_check(unsigned int x86_leaf)*/ static __always_inline u32 __feature_bit(int x86_feature) { -reverse_cpuid_check (x86_feature /32 ); -return 1 << (x86_feature &31 ); +int leaf = x86_feature /32 ; + +reverse_cpuid_check (leaf); + +if (leaf ==CPUID_LNX_4 ) +return map_synthetic_leaf (x86_feature); + +return BIT (x86_feature &31 ); }#define feature_bit (name) __feature_bit(X86_FEATURE_##name)
Signature · 4 lines
Regards/Gruss,
Boris.
https://people.kernel.org/tglx/notes-about-netiquetteRe: [RFC PATCH 04/23] x86/cpufeatures: Add SGX1 and SGX2 sub-features
Quoted 11 lines · On Fri, 8 Jan 2021 08:17:22 +0100 Borislav Petkov wrote:
Quoted 10 lines · Kai Huang
Quoted 9 lines
Quoted 4 lines
I am not sure changing reverse lookup to handle dynamic would be acceptable. To me it is ugly, and I don't have a first glance on how to do it. KVM can query host CPUID when dealing with SGX w/o X86_FEATURE_SGX1/2, but it is not as straightforward as having X86_FEATURE_SGX1/2.
So, Boris was pretty direct here. Could you please go spend a bit of time to see what it would take to make these dynamic? You can check what our (Intel) plans are for this leaf, but if it's going to remain sparsely-used, we need to look into making the leaves a bit more dynamic.
Quoted 6 lines
Quoted 2 lines
I don't think reverse lookup can be made dyanmic, but like I said if we don't have X86_FEATURE_SGX1/2, KVM needs to query raw CPUID when dealing with SGX.
How about before you go and say that "it is ugly" and "don't think can be made" you actually go and *really* try it first? Because actually trying is sometimes faster than trying to find arguments against it. :)
THanks. Lesson learned :)
Quoted 4 lines
Because I just did it and unless I'm missing something obvious - I haven't actually tested it - this is not ugly at all and in the long run it will become one big switch-case, which is perfectly fine.
No offence, but using synthetic bits is a little bit hack to me,given they are actually hardware feature bits. And using synthetic leaf in reverse lookup is against current KVM code.
I'll try my own way in next version, but thank you for the insight! :)
Quoted 90 lines
---
arch/x86/include/asm/cpufeature.h +1
-0@@ -30,6 +30,7 @@ enum cpuid_leafsCPUID_7_ECX ,CPUID_8000_0007_EBX ,CPUID_7_EDX , + CPUID_12_EAX, /* used only by KVM for now */ };#ifdef CONFIG_X86_FEATURE_NAMES arch/x86/include/asm/cpufeatures.h +2
-0@@ -292,6 +292,8 @@#define X86_FEATURE_FENCE_SWAPGS_KERNEL (11*32+ 5)/* "" LFENCE in kernel entry SWAPGS path */ #define X86_FEATURE_SPLIT_LOCK_DETECT (11*32+ 6)/* #AC for split lock */ #define X86_FEATURE_PER_THREAD_MBA (11*32+ 7)/* "" Per-thread Memory Bandwidth Allocation */ +#define X86_FEATURE_SGX1 (11*32+ 8)/* SGX1 leaf functions */ +#define X86_FEATURE_SGX2 (11*32+ 9)/* SGX2 leaf functions */ /* Intel-defined CPU features, CPUID level 0x00000007:1 (EAX), word 12 */ #define X86_FEATURE_AVX512_BF16 (12*32+ 5)/* AVX512 BFLOAT16 instructions */ arch/x86/kvm/cpuid.h +27
-3@@ -63,8 +63,27 @@ static const struct cpuid_reg reverse_cpuid[] = { [CPUID_8000_0007_EBX] = {0x80000007, 0, CPUID_EBX}, [CPUID_7_EDX] = { 7, 0, CPUID_EDX}, [CPUID_7_1_EAX] = { 7, 1, CPUID_EAX}, + [CPUID_12_EAX] = { 0x12, 0, CPUID_EAX}, }; +/* +* Map a synthetic X86_FEATURE bit definition to the corresponding bit in the +* hardware CPUID leaf. +*/ +static int map_synthetic_leaf (int x86_feature) +{ +switch (x86_feature) { +case X86_FEATURE_SGX1 :return BIT (0 ); +case X86_FEATURE_SGX2 :return BIT (1 ); +default : +break ; + } + +WARN_ON_ONCE (1 ); + +return 0 ; +} +/* * Reverse CPUID and its derivatives can only be used for hardware-defined * feature words, i.e. words whose bits directly correspond to a CPUID leaf. @@ -78,7 +97,6 @@ static __always_inline void reverse_cpuid_check(unsigned int x86_leaf)BUILD_BUG_ON (x86_leaf ==CPUID_LNX_1 );BUILD_BUG_ON (x86_leaf ==CPUID_LNX_2 );BUILD_BUG_ON (x86_leaf ==CPUID_LNX_3 ); -BUILD_BUG_ON (x86_leaf ==CPUID_LNX_4 );BUILD_BUG_ON (x86_leaf >=ARRAY_SIZE (reverse_cpuid));BUILD_BUG_ON (reverse_cpuid[x86_leaf].function ==0 ); }@@ -91,8 +109,14 @@ static __always_inline void reverse_cpuid_check(unsigned int x86_leaf)*/ static __always_inline u32 __feature_bit(int x86_feature) { -reverse_cpuid_check (x86_feature /32 ); -return 1 << (x86_feature &31 ); +int leaf = x86_feature /32 ; + +reverse_cpuid_check (leaf); + +if (leaf ==CPUID_LNX_4 ) +return map_synthetic_leaf (x86_feature); + +return BIT (x86_feature &31 ); }#define feature_bit (name) __feature_bit(X86_FEATURE_##name)Signature · 5 lines
-- Regards/Gruss, Boris. https://people.kernel.org/tglx/notes-about-netiquette
Borislav Petkov
maintainerRe: [RFC PATCH 04/23] x86/cpufeatures: Add SGX1 and SGX2 sub-features
Quoted 2 lines · Kai Huang
No offence, but using synthetic bits is a little bit hack to me,given they are actually hardware feature bits.
Why?
Perhaps you need to have a look at Documentation/x86/cpuinfo.rst first.
Quoted 2 lines
And using synthetic leaf in reverse lookup is against current KVM code.
You know how the kernel gets improved each day and old limitations are not valid anymore?
Quoted 1 line
I'll try my own way in next version, but thank you for the insight! :)
Feel free but remember to keep it simple. You can use mine too, if you want to, as long as you attribute it with a Suggested-by or so.
Signature · 4 lines
Regards/Gruss,
Boris.
https://people.kernel.org/tglx/notes-about-netiquetteRe: [RFC PATCH 04/23] x86/cpufeatures: Add SGX1 and SGX2 sub-features
Quoted 7 lines · On Fri, 8 Jan 2021 09:13:14 +0100 Borislav Petkov wrote:
Quoted 2 lines · Kai Huang
No offence, but using synthetic bits is a little bit hack to me,given they are actually hardware feature bits.
Why?
Perhaps you need to have a look at
Documentation/x86/cpuinfo.rstfirst.
Will take a look. Thanks.
Quoted 11 lines
Quoted 3 lines
And using synthetic leaf in reverse lookup is against current KVM code.
You know how the kernel gets improved each day and old limitations are not valid anymore?
Quoted 1 line
I'll try my own way in next version, but thank you for the insight! :)
Feel free but remember to keep it simple. You can use mine too, if you want to, as long as you attribute it with a Suggested-by or so.
OK. Thanks.
Quoted 6 lines
Signature · 5 lines
-- Regards/Gruss, Boris. https://people.kernel.org/tglx/notes-about-netiquette
Sean Christopherson
maintainerRe: [RFC PATCH 04/23] x86/cpufeatures: Add SGX1 and SGX2 sub-features
In reply to Borislav PetkovQuoted 10 lines · Borislav Petkov
Quoted 9 lines · Kai Huang
Quoted 9 lines
Quoted 4 lines
I am not sure changing reverse lookup to handle dynamic would be acceptable. To me it is ugly, and I don't have a first glance on how to do it. KVM can query host CPUID when dealing with SGX w/o X86_FEATURE_SGX1/2, but it is not as straightforward as having X86_FEATURE_SGX1/2.
So, Boris was pretty direct here. Could you please go spend a bit of time to see what it would take to make these dynamic? You can check what our (Intel) plans are for this leaf, but if it's going to remain sparsely-used, we need to look into making the leaves a bit more dynamic.
To be fair, this is the third time we've got conflicting, direct feedback on this exact issue. I do agree that it doesn't make sense to burn a whole word for just two features, I guess I just feel like whining.
[*] https://lore.kernel.org/kvm/[email protected]/
[*] https://lore.kernel.org/linux-sgx/[email protected]/
Quoted 42 lines
Quoted 2 lines
I don't think reverse lookup can be made dyanmic, but like I said if we don't have X86_FEATURE_SGX1/2, KVM needs to query raw CPUID when dealing with SGX.
How about before you go and say that "it is ugly" and "don't think can be made" you actually go and *really* try it first? Because actually trying is sometimes faster than trying to find arguments against it. :)
Because I just did it and unless I'm missing something obvious - I haven't actually tested it - this is not ugly at all and in the long run it will become one big switch-case, which is perfectly fine.
---
arch/x86/include/asm/cpufeature.h +1
-0@@ -30,6 +30,7 @@ enum cpuid_leafsCPUID_7_ECX ,CPUID_8000_0007_EBX ,CPUID_7_EDX , + CPUID_12_EAX, /* used only by KVM for now */ };#ifdef CONFIG_X86_FEATURE_NAMES arch/x86/include/asm/cpufeatures.h +2
-0@@ -292,6 +292,8 @@#define X86_FEATURE_FENCE_SWAPGS_KERNEL (11*32+ 5)/* "" LFENCE in kernel entry SWAPGS path */ #define X86_FEATURE_SPLIT_LOCK_DETECT (11*32+ 6)/* #AC for split lock */ #define X86_FEATURE_PER_THREAD_MBA (11*32+ 7)/* "" Per-thread Memory Bandwidth Allocation */ +#define X86_FEATURE_SGX1 (11*32+ 8)/* SGX1 leaf functions */ +#define X86_FEATURE_SGX2 (11*32+ 9)/* SGX2 leaf functions */ /* Intel-defined CPU features, CPUID level 0x00000007:1 (EAX), word 12 */ #define X86_FEATURE_AVX512_BF16 (12*32+ 5)/* AVX512 BFLOAT16 instructions */ arch/x86/kvm/cpuid.h +0
-0
arch/x86/kvm/cpuid.h static const struct cpuid_reg reverse_cpuid[] = { · line 67
Quoted 5 lines
arch/x86/kvm/cpuid.h +1
-0@@ -63,8 +63,27 @@ static const struct cpuid_reg reverse_cpuid[] = { [CPUID_8000_0007_EBX] = {0x80000007, 0, CPUID_EBX}, [CPUID_7_EDX] = { 7, 0, CPUID_EDX}, [CPUID_7_1_EAX] = { 7, 1, CPUID_EAX}, + [CPUID_12_EAX] = { 0x12, 0, CPUID_EAX}, };
As is, this won't build (if KVM uses the features) because KVM cares about where the feature actually lives in Linux's words. The addition of CPUID_12_EAX is unnecessary, and the new entry in reverse_cpuid would need to be s/CPUID_12_EAX/CPUID_LNX_4.
That being said, I dislike this approach as it introduces fragility into KVM's CPUID shenanigans. E.g. fixing the above will make guest_cpuid_has() functional, but kvm_cpu_cap_mask() and cpuid_entry_override() will not work as expected.
Here's a more involved approach that I believe will work (compile tested only) and retains KVM's build magic. Idea is to allocate a word kvm_cpu_caps for the hardware-defined, Linux-scattered features, and use boot_cpu_has() to bridge the gap when populating kvm_cpu_caps.
Another alternative would be to have KVM use boot_cpu_has() for everything, and omit the memcpy from boot_cpu_data.x86_capability -> kvm_cpu_caps. That would eliminate some of the special logic for scattered features, but it adds nearly 3k bytes to kvm_set_cpu_caps(), which is hard to stomach even though it's effectively one-and-done code.
From 6bdd61e23f1c0bd7519a3a6391c95cde5456f79d Mon Sep 17 00:00:00 2001
From: Sean Christopherson <[email protected]>
Date: Fri, 8 Jan 2021 15:46:11 -0800
Subject: [PATCH] KVM: x86: Add support for reverse CPUID lookup of scattered
features
--- arch/x86/include/asm/cpufeatures.h | 2 ++ arch/x86/kvm/cpuid.c | 36 ++++++++++++++++++--- arch/x86/kvm/cpuid.h | 50 +++++++++++++++++++++++++++--- 3 files changed, 78 insertions(+), 10 deletions(-)
arch/x86/include/asm/cpufeatures.h +2 -0
@@ -291,6 +291,8 @@#define X86_FEATURE_FENCE_SWAPGS_KERNEL (11*32+ 5)/* "" LFENCE in kernel entry SWAPGS path */ #define X86_FEATURE_SPLIT_LOCK_DETECT (11*32+ 6)/* #AC for split lock */ #define X86_FEATURE_PER_THREAD_MBA (11*32+ 7)/* "" Per-thread Memory Bandwidth Allocation */ +#define X86_FEATURE_SGX1 (11*32+ 8)/* SGX1 leafs */ +#define X86_FEATURE_SGX2 (11*32+ 9)/* SGX2 leafs */ /* Intel-defined CPU features, CPUID level 0x00000007:1 (EAX), word 12 */ #define X86_FEATURE_AVX512_BF16 (12*32+ 5)/* AVX512 BFLOAT16 instructions */
arch/x86/kvm/cpuid.c +31 -5
@@ -28,7 +28,7 @@* Unlike "struct cpuinfo_x86.x86_capability", kvm_cpu_caps doesn't need to be * aligned to sizeof(unsigned long) because it's not accessed via bitops. */ -u32 kvm_cpu_caps[NCAPINTS ]__read_mostly ; +u32 kvm_cpu_caps[NR_KVM_CPU_CAPS ]__read_mostly ;EXPORT_SYMBOL_GPL (kvm_cpu_caps);static u32 xstate_required_size (u64 xstate_bv,bool compacted)
@@ -53,6 +53,7 @@ static u32 xstate_required_size(u64 xstate_bv, bool compacted) }#define F feature_bit +#define SF (name) (boot_cpu_has(X86_FEATURE_##name) ? F(name) : 0)static inline struct kvm_cpuid_entry2 *cpuid_entry2_find (struct kvm_cpuid_entry2 *entries,int nent,u32 function,u32 index)
@@ -331,13 +332,13 @@ int kvm_vcpu_ioctl_get_cpuid2(struct kvm_vcpu *vcpu,return r; } -static __always_inline void kvm_cpu_cap_mask(enum cpuid_leafs leaf, u32 mask) +/* Mask kvm_cpu_caps for @leaf with the raw CPUID capabilities of this CPU. */ +static __always_inline void __kvm_cpu_cap_mask(enum cpuid_leafs leaf) {const struct cpuid_reg cpuid =x86_feature_cpuid (leaf *32 );struct kvm_cpuid_entry2 entry;reverse_cpuid_check (leaf); - kvm_cpu_caps[leaf] &= mask;cpuid_count (cpuid.function, cpuid.index, &entry.eax, &entry.ebx, &entry.ecx, &entry.edx);
@@ -345,6 +346,26 @@ static __always_inline void kvm_cpu_cap_mask(enum cpuid_leafs leaf, u32 mask) kvm_cpu_caps[leaf] &= *__cpuid_entry_get_reg (&entry, cpuid.reg); } +static __always_inline void kvm_cpu_cap_mask(enum cpuid_leafs leaf, u32 mask) +{ +/* Use the "init" variant for scattered leafs. */ +BUILD_BUG_ON (leaf >=NCAPINTS ); + + kvm_cpu_caps[leaf] &= mask; + +__kvm_cpu_cap_mask (leaf); +} + +static __always_inline void kvm_cpu_cap_init(enum cpuid_leafs leaf, u32 mask) +{ +/* Use the "mask" variant for hardwared-defined leafs. */ +BUILD_BUG_ON (leaf <NCAPINTS ); + + kvm_cpu_caps[leaf] = mask; + +__kvm_cpu_cap_mask (leaf); +} +void kvm_set_cpu_caps (void ) {unsigned int f_nx =is_efer_nx () ?F (NX ) :0 ;
@@ -355,12 +376,13 @@ void kvm_set_cpu_caps(void)unsigned int f_gbpages =0 ;unsigned int f_lm =0 ;#endif +memset (kvm_cpu_caps,0 ,sizeof (kvm_cpu_caps)); -BUILD_BUG_ON (sizeof (kvm_cpu_caps) > +BUILD_BUG_ON (sizeof (kvm_cpu_caps) - (NKVMCAPINTS *sizeof (*kvm_cpu_caps)) >sizeof (boot_cpu_data.x86_capability));memcpy (&kvm_cpu_caps, &boot_cpu_data.x86_capability, -sizeof (kvm_cpu_caps)); +sizeof (kvm_cpu_caps) - (NKVMCAPINTS *sizeof (*kvm_cpu_caps))); kvm_cpu_cap_mask(CPUID_1_ECX,/*
@@ -503,6 +525,10 @@ void kvm_set_cpu_caps(void)F (ACE2 ) |F (ACE2_EN ) |F (PHE ) |F (PHE_EN ) |F (PMM ) |F (PMM_EN ) ); + +kvm_cpu_cap_init (CPUID_12_EAX , +SF (SGX1 ) |SF (SGX2 ) + ); }EXPORT_SYMBOL_GPL (kvm_set_cpu_caps);
arch/x86/kvm/cpuid.h +45 -5
98 lines not on this page · raw message
--
2.30.0.284.gd98b1dd5eaa7-goog
Borislav Petkov
maintainerRe: [RFC PATCH 04/23] x86/cpufeatures: Add SGX1 and SGX2 sub-features
Quoted 6 lines · Sean Christopherson
To be fair, this is the third time we've got conflicting, direct feedback on this exact issue. I do agree that it doesn't make sense to burn a whole word for just two features, I guess I just feel like whining.
[*] https://lore.kernel.org/kvm/[email protected]/
[*] https://lore.kernel.org/linux-sgx/[email protected]/
Well, sorry that I confused you guys but in hindsight we probably should have stopped you right then and there from imposing kvm requirements on the machinery behind *_cpu_has() and kvm should have been a regular user of those interfaces like the rest of the kernel code - nothing more.
And if you'd like to do your own X86_FEATURE_* querying but then extend it with its own functionality, then that should have been decoupled.
And I will look at your patch later when brain is actually awake but I strongly feel that in order to avoid such situations in the future, *_cpu_has() internal functionality should be separate from kvm's respective CPUID leafs representation. For obvious reasons.
And if there should be some partial sharing - if that makes sense at all
- then that should be first agreed upon.
Thx.
Signature · 4 lines
Regards/Gruss,
Boris.
https://people.kernel.org/tglx/notes-about-netiquetteSean Christopherson
maintainerRe: [RFC PATCH 04/23] x86/cpufeatures: Add SGX1 and SGX2 sub-features
Quoted 13 lines · Borislav Petkov
Quoted 6 lines · Sean Christopherson
To be fair, this is the third time we've got conflicting, direct feedback on this exact issue. I do agree that it doesn't make sense to burn a whole word for just two features, I guess I just feel like whining.
[*] https://lore.kernel.org/kvm/[email protected]/
[*] https://lore.kernel.org/linux-sgx/[email protected]/Well, sorry that I confused you guys but in hindsight we probably should have stopped you right then and there from imposing kvm requirements on the machinery behind
*_cpu_has()and kvm should have been a regular user of those interfaces like the rest of the kernel code - nothing more.
Quoted 7 lines
And if you'd like to do your own X86_FEATURE_* querying but then extend it with its own functionality, then that should have been decoupled.
And I will look at your patch later when brain is actually awake but I strongly feel that in order to avoid such situations in the future,
*_cpu_has()internal functionality should be separate from kvm's respective CPUID leafs representation. For obvious reasons.
I kinda agree, but I'd prefer not to fully decouple KVM's CPUID stuff. The more manual definitions/translations we have to create, the more likely it is that we'll screw something up.
Quoted 2 lines
And if there should be some partial sharing - if that makes sense at all
- then that should be first agreed upon.
Assuming the code I wrote actually works, I think that gets KVM to the point where handling scattered features isn't awful, which should eliminate most of the friction. KVM would still be relying on the internals of *_cpu_has(), but there are quite a few build-time assertions that help keep things aligned. And, what's best for the kernel will be what's best for KVM the vast majority of the time, e.g. I don't anticipate the kernel scattering densely populated words just for giggles.
Borislav Petkov
maintainerRe: [RFC PATCH 04/23] x86/cpufeatures: Add SGX1 and SGX2 sub-features
In reply to Sean Christophersonarch/x86/kvm/cpuid.h line 13
Quoted 13 lines · Sean Christopherson
arch/x86/kvm/cpuid.h +4
-1@@ -7,7 +7,25 @@#include <asm/processor.h> #include <uapi/asm/kvm_para.h> -extern u32 kvm_cpu_caps[NCAPINTS ]__read_mostly ; +/* +* Hardware-defined CPUID leafs that are scattered in the kernel, but need to +* be directly by KVM. Note, these word values conflict with the kernel's +* "bug" caps, but KVM doesn't use those.
This feels like another conflict waiting to happen if KVM decides to use them at some point...
So let me get this straight: KVM wants to use X86_FEATURE_* which means, those numbers must map to the respective words in its CPUID caps representation kvm_cpu_caps, AFAICT.
Then, it wants the leafs to correspond to the hardware leafs layout so that it can do:
kvm_cpu_caps[leaf] &= *__cpuid_entry_get_reg(&entry, cpuid.reg);
which comes straight from CPUID.
So lemme look at one word:
kvm_cpu_cap_mask(CPUID_1_EDX,
F(FPU) | F(VME) | F(DE) | F(PSE) |
F(TSC) | F(MSR) | F(PAE) | F(MCE) |
...it would build the bitmask of the CPUID leaf using X86_FEATURE_* bits and then mask it out with the hardware leaf read from CPUID.
But why?
Why doesn't it simply build those leafs in kvm_cpu_caps from the leafs we've already queried?
Oh it does so a bit earlier:
memcpy(&kvm_cpu_caps, &boot_cpu_data.x86_capability,
sizeof(kvm_cpu_caps));and that kvm_cpu_cap_mask() call is to clear some bits in kvm_cpu_caps which is kvm-specific thing (not supported stuff etc).
But then why does kvm_cpu_cap_mask() does cpuid_count()? Didn't it just read the bits from boot_cpu_data.x86_capability? And those bits we do query and massage extensively during boot. So why does KVM needs to query CPUID again instead of using what we've already queried?
Maybe I'm missing something kvm-specific.
In any case, this feels somewhat weird: you have *_cpu_has() on baremetal abstracting almost completely from CPUID by collecting all feature bits it needs into its own structure - x86_capability[] along with accessors for it - and then you want to "abstract back" to CPUID leafs from that interface. I wonder why.
Anyway, more questions tomorrow.
Gnight and good luck. :)
Signature · 4 lines
Regards/Gruss,
Boris.
https://people.kernel.org/tglx/notes-about-netiquetteJarkko Sakkinen
maintainerQuoted 40 lines · Kai Huang
--- Disclaimer ---
These patches were originally written by Sean Christopherson while at Intel. Now that Sean has left Intel, I (Kai) have taken over getting them upstream. This series needs more review before it can be merged. It is being posted publicly and under RFC so Sean and others can review it. Maintainers are safe ignoring it for now.
------------------
Hi all,
This series adds KVM SGX virtualization support. The first 12 patches starting with x86/sgx or x86/cpu.. are necessary changes to x86 and SGX core/driver to support KVM SGX virtualization, while the rest are patches to KVM subsystem.
Please help to review this series. Also I'd like to hear what is the proper way to merge this series, since it contains change to both x86/SGX and KVM subsystem. Any feedback is highly appreciated. And please let me know if I forgot to CC anyone, or anyone wants to be removed from CC. Thanks in advance!
This series is based against latest tip tree's x86/sgx branch. You can also get the code from tip branch of kvm-sgx repo on github:
https://github.com/intel/kvm-sgx.git tip
It also requires Qemu changes to create VM with SGX support. You can find Qemu repo here:
https://github.com/intel/qemu-sgx.git next
Please refer to README.md of above qemu-sgx repo for detail on how to create guest with SGX support. At meantime, for your quick reference you can use below command to create SGX guest:
#qemu-system-x86_64 -smp 4 -m 2G -drive file=<your_vm_image>,if=virtio \ -cpu host,+sgx_provisionkey \ -sgx-epc id=epc1,memdev=mem1 \ -object memory-backend-epc,id=mem1,size=64M,prealloc
Quoted 7 lines
Please note that the SGX relevant part is:
-cpu host,+sgx_provisionkey \ -sgx-epc id=epc1,memdev=mem1 \ -object memory-backend-epc,id=mem1,size=64M,preallocAnd you can change other parameters of your qemu command based on your needs.
Thanks a lot documenting these snippets to the cover letter. I dig these up from lore once my environment is working.
I'm setting up Arch based test environment with the eye on this patch set and generic Linux keyring patches:
https://git.kernel.org/pub/scm/linux/kernel/git/jarkko/arch.git/
Still have some minor bits to adjust before I can start deploying it for SGX testing. For this patch set I'll use two instances of it.
Quoted 7 lines
=========
KVM SGX virtualization Overview- Virtual EPC
"Virtual EPC" is the EPC section exposed by KVM to guest so SGX software in guest can discover it and use it to create SGX enclaves. KVM exposes SGX to
Virtual EPC is a representation of an EPC section. And there is no "the".
Quoted 11 lines
guest via CPUID, and exposes one or more "virtual EPC" sections for guest. The size of "virtual EPC" is passed as Qemu parameter when creating the guest, and the base address is calcualted internally according to guest's configuration.
To support virtual EPC, add a new misc device
/dev/sgx_virt_epcto SGX core/driver to allow userspace (Qemu) to allocate "raw" EPC, and use it as "virtual EPC" for guest. Obviously, unlike EPC allocated for host SGX driver, virtual EPC allocated via/dev/sgx_virt_epcdoesn't have enclave associated, and how virtual EPC is used by guest is compeletely controlled by guest's SGX software.
I think that /dev/sgx_vepc would be a clear enough name for the device. This text has now a bit confusing "terminology" related to this.
In some places virtual EPC is quotes, and in other places it is not. I think that you could consistently an abbervation vEPC (without quotations):
"
vEPC
====
Virtual EPC, shortened as vEPC, is a representation of ... "
Quoted 2 lines
Implement the "raw" EPC allocation in the x86 core-SGX subsystem via
/dev/sgx_virt_epcrather than in KVM. Doing so has two major advantages:
Maybe you could remove "raw" from the text.
Quoted 2 lines
- Does not require changes to KVM's uAPI, e.g. EPC gets handled as just another memory backend for guests.
Why this an advantage? No objection, just a question.
Quoted 4 lines
- EPC management is wholly contained in the SGX subsystem, e.g. SGX does not have to export any symbols, changes to reclaim flows don't need to be routed through KVM, SGX's dirty laundry doesn't have to get aired out for the world to see, and so on and so forth.
No comments to this before understanding code changes better.
Quoted 6 lines
The virtual EPC allocated to guests is currently not reclaimable, due to reclaiming EPC from KVM guests is not currently supported. Due to the complications of handling reclaim conflicts between guest and host, KVM EPC oversubscription, which allows total virtual EPC size greater than physical EPC by being able to reclaiming guests' EPC, is significantly more complex than basic support for SGX virtualization.
I think it should be really in the center of the patch set description that this patch set implements segmentation of EPC, not oversubscription. It should be clear immediately. It's a core part of knowing "what I'm looking at".
Quoted 11 lines
- Support SGX virtualization without SGX Launch Control unlocked mode
Although SGX driver requires SGX Launch Control unlocked mode to work, SGX virtualization doesn't, since how enclave is created is completely controlled by guest SGX software, which is not necessarily linux. Therefore, this series allows KVM to expose SGX to guest even SGX Launch Control is in locked mode, or is not present at all. The reason is the goal of SGX virtualization, or virtualization in general, is to expose hardware feature to guest, but not to make assumption how guest will use it. Therefore, KVM should support SGX guest as long as hardware is able to, to have chance to support more potential use cases in cloud environment.
AFAIK the convergence point with the FLC was, and is that Linux never enables SGX with locked MSRs.
And I don't understand, if it is not fine to allow locked SGX for a *process*, why is it fine for a *virtual machine*? They have a lot same.
I cannot remember out of top of my head, could the Intel SHA256 be read when booted with unlocked MSRs. If that is the case, then you can still support guests with that configuration.
Context-dependent guidelines tend to also trash code big time. Also, for the sake of a sane kernel code base, I would consider only supporting unlocked MSRs.
/Jarkko
Sean Christopherson
maintainerRe: [RFC PATCH 04/23] x86/cpufeatures: Add SGX1 and SGX2 sub-features
In reply to Borislav PetkovQuoted 9 lines · Borislav Petkov
Quoted 8 lines · Sean Christopherson
arch/x86/kvm/cpuid.h +0
-0@@ -7,7 +7,25 @@#include <asm/processor.h> #include <uapi/asm/kvm_para.h>
Quoted 8 lines
arch/x86/kvm/cpuid.h line 13
Quoted 5 lines
arch/x86/kvm/cpuid.h +4
-1-extern u32 kvm_cpu_caps[NCAPINTS ]__read_mostly ; +/* +* Hardware-defined CPUID leafs that are scattered in the kernel, but need to +* be directly by KVM. Note, these word values conflict with the kernel's +* "bug" caps, but KVM doesn't use those. This feels like another conflict waiting to happen if KVM decides to use them at some point...
Yes, but KVM including the bug caps in kvm_cpu_caps is extremely unlikely, and arguably flat out wrong. Currently, kvm_cpu_caps includes only CPUID-based features that can be exposed direcly to the guest. I could see a scenario where KVM exposed "bug" capabilities to the guest via a paravirt interface, but I would expect that KVM would either filter and expose the kernel's bug caps without userspace input, or would add a KVM-defined paravirt CPUID leaf to enumerate the caps and track _that_ in kvm_cpu_caps.
Anyways, I agree that overlapping the bug caps it's a bit of unnecessary cleverness. I'm not opposed to incorporating NBUGINTS into KVM, but that would mean explicitly pulling in even more x86_capability implementation details.
Quoted 3 lines
So let me get this straight: KVM wants to use X86_FEATURE_* which means, those numbers must map to the respective words in its CPUID caps representation
kvm_cpu_caps, AFAICT.
That part is deliberate and isn't a dependency so much as how things are implemented. The true dependency is on the bit offsets within each word. The kernel could completely rescramble the word numbering and KVM would chug along happily. What KVM won't play nice with is if the kernel broke up a hardware- defined, gathered CPUID leaf/word into scattered features spread out amongst multiple Linux-defined words.
Quoted 28 lines
Then, it wants the leafs to correspond to the hardware leafs layout so that it can do:
kvm_cpu_caps[leaf] &= *__cpuid_entry_get_reg(&entry, cpuid.reg);which comes straight from CPUID.
So lemme look at one word:
kvm_cpu_cap_mask(CPUID_1_EDX, F(FPU) | F(VME) | F(DE) | F(PSE) | F(TSC) | F(MSR) | F(PAE) | F(MCE) | ...it would build the bitmask of the CPUID leaf using X86_FEATURE_* bits and then mask it out with the hardware leaf read from CPUID.
But why?
Why doesn't it simply build those leafs in
kvm_cpu_capsfrom the leafs we've already queried?Oh it does so a bit earlier:
memcpy(&kvm_cpu_caps, &boot_cpu_data.x86_capability, sizeof(kvm_cpu_caps));
Quoted 7 lines
and that
kvm_cpu_cap_mask()call is to clear some bits inkvm_cpu_capswhich is kvm-specific thing (not supported stuff etc).But then why does
kvm_cpu_cap_mask()doescpuid_count()? Didn't it just read the bits from boot_cpu_data.x86_capability? And those bits we do query and massage extensively during boot. So why does KVM needs to query CPUID again instead of using what we've already queried?
It's mostly historical; before the kvm_cpu_caps concept was introduced, the code had grown organically to include both boot_cpu_data and raw CPUID info. The vast, vast majority of the time, doing CPUID is likely redundant. But, as noted in commit d8577a4c238f ("KVM: x86: Do host CPUID at load time to mask KVM cpu caps"), the code is quite cheap and runs once at KVM load. My argument back then was, and still is, that an extra bit of paranoia is justified since the code and operations are quite nearly free.
This particular dependency can be broken, and quite easily at that. Rather than memcpy() boot_cpu_data.x86_capability, it's trivially easy to redefine the F() macro to invoke boot_cpu_has(), which would allow dropping the memcpy(). The big downside, and why I didn't post the code, is that doing so means every feature routed through F() requires some form of BT+Jcc (or CMOVcc) sequence, whereas the mempcy() approach allows the F() features to be encoded as a single literal by the compiler.
From a latency perspective, the extra code is negligible. The big issue is that all those extra checks add 2k+ bytes of code. Eliminating the mempcy() doesn't actually break KVM's dependency on the bit offsets, so we'd be bloating kvm.ko by a noticeable amount without providing substantial value.
And, this behavior is mostly opportunistic; the true justification/motiviation for taking a dependency on the X86_FEATURE_* bit offsets is for communication with userspace, querying the guest CPU model, and runtime checks.
Quoted 7 lines
Maybe I'm missing something kvm-specific.
In any case, this feels somewhat weird: you have
*_cpu_has()on baremetal abstracting almost completely from CPUID by collecting all feature bits it needs into its own structure - x86_capability[] along with accessors for it - and then you want to "abstract back" to CPUID leafs from that interface. I wonder why.
It's effectively for communication with userspace. Userspace, via ioctl(), dictates the vCPU model to KVM, including the exact CPUID results. to properly virtualize/emulate the defined vCPU model, KVM must query the dictated CPUID results to determine what features are supported, what guest operations should fault, etc... E.g. if the vCPU model, via CPUID, states that SMEP isn't supported then KVM needs to inject a #GP if the guest attempts to set CR4.SMEP.
KVM also uses the hardware-defined CPUID ABI to advertise which features are supported by both hardware and KVM. This is the kvm_cpu_cap stuff, where KVM reads boot_cpu_data to see what features were enabled by the kernel.
It would be possible for KVM to break the dependency on X86_FEATURE_* bit offsets by defining a translation layer, but I strongly feel that adding manual translations will do more harm than good as it increases the odds of us botching a translation or using the wrong feature flag, creates potential namespace conflicts, etc...
Sean Christopherson
maintainerQuoted 5 lines · Jarkko Sakkinen
Quoted 2 lines · Kai Huang
- Does not require changes to KVM's uAPI, e.g. EPC gets handled as just another memory backend for guests.
Why this an advantage? No objection, just a question.
There are zero KVM changes required to support exposing EPC to a guest. KVM's MMU is completely ignorant of what physical backing is used for any given host virtual address. KVM has to be aware of various VM_* flags, e.g. VM_PFNMAP and VM_IO, but that code is arch agnostic and is quite isolated.
Quoted 14 lines
Quoted 4 lines
- EPC management is wholly contained in the SGX subsystem, e.g. SGX does not have to export any symbols, changes to reclaim flows don't need to be routed through KVM, SGX's dirty laundry doesn't have to get aired out for the world to see, and so on and so forth.
No comments to this before understanding code changes better.
Quoted 6 lines
The virtual EPC allocated to guests is currently not reclaimable, due to reclaiming EPC from KVM guests is not currently supported. Due to the complications of handling reclaim conflicts between guest and host, KVM EPC oversubscription, which allows total virtual EPC size greater than physical EPC by being able to reclaiming guests' EPC, is significantly more complex than basic support for SGX virtualization.
Quoted 3 lines
I think it should be really in the center of the patch set description that this patch set implements segmentation of EPC, not oversubscription. It should be clear immediately. It's a core part of knowing "what I'm looking at".
Technically, it doesn't implement EPC segmentation of EPC. It implements non-reclaimable EPC allocation. Even that is somewhat untrue as the EPC can be forcefully reclaimed, but doing so will destroy the guest contents.
Userspace can oversubscribe the EPC to KVM guests, but it would need to kill, migrate, or pause one or more VMs if the pool of physical EPC were exhausted.
Quoted 12 lines
Quoted 11 lines
- Support SGX virtualization without SGX Launch Control unlocked mode
Although SGX driver requires SGX Launch Control unlocked mode to work, SGX virtualization doesn't, since how enclave is created is completely controlled by guest SGX software, which is not necessarily linux. Therefore, this series allows KVM to expose SGX to guest even SGX Launch Control is in locked mode, or is not present at all. The reason is the goal of SGX virtualization, or virtualization in general, is to expose hardware feature to guest, but not to make assumption how guest will use it. Therefore, KVM should support SGX guest as long as hardware is able to, to have chance to support more potential use cases in cloud environment.
Quoted 5 lines
AFAIK the convergence point with the FLC was, and is that Linux never enables SGX with locked MSRs.
And I don't understand, if it is not fine to allow locked SGX for a *process*, why is it fine for a *virtual machine*? They have a lot same.
Because it's a completely different OS/kernel. If the user has a kernel that supports locked SGX, then so be it. There's no novel circumvention of the kernel policy, e.g. the user could simply boot the non-upstream kernel directly, and running an upstream kernel in the guest will not cause the kernel to support SGX.
There are any number of things that are allowed in a KVM guest that are not allowed in a bare metal process.
Quoted 3 lines
I cannot remember out of top of my head, could the Intel SHA256 be read when booted with unlocked MSRs. If that is the case, then you can still support guests with that configuration.
No, it's not guaranteed to be readable as firmware could have already changed the values in the MSRs.
Quoted 3 lines
Context-dependent guidelines tend to also trash code big time. Also, for the sake of a sane kernel code base, I would consider only supporting unlocked MSRs.
It's one line of a code to teach the kernel driver not to load if the MSRs are locked. And IMO, that one line of code is a net positive as it makes it clear in the driver itself that it chooses not support locked MSRs, even if SGX itself is fully enabled.
Borislav Petkov
maintainerRe: [RFC PATCH 04/23] x86/cpufeatures: Add SGX1 and SGX2 sub-features
In reply to Sean ChristophersonQuoted 11 lines · Sean Christopherson
Yes, but KVM including the bug caps in
kvm_cpu_capsis extremely unlikely, and arguably flat out wrong. Currently,kvm_cpu_capsincludes only CPUID-based features that can be exposed direcly to the guest. I could see a scenario where KVM exposed "bug" capabilities to the guest via a paravirt interface, but I would expect that KVM would either filter and expose the kernel's bug caps without userspace input, or would add a KVM-defined paravirt CPUID leaf to enumerate the caps and track _that_ inkvm_cpu_caps.Anyways, I agree that overlapping the bug caps it's a bit of unnecessary cleverness. I'm not opposed to incorporating NBUGINTS into KVM, but that would mean explicitly pulling in even more
x86_capabilityimplementation details.
Also, the kernel and kvm being part of it :) kinda tries to fix those bugs and not expose them to the guest so exposing a bug would probably be only for testing purposes...
Quoted 2 lines
That part is deliberate and isn't a dependency so much as how things are implemented. The true dependency is on the bit offsets within each word.
Right.
Quoted 4 lines
The kernel could completely rescramble the word numbering and KVM would chug along happily. What KVM won't play nice with is if the kernel broke up a hardware- defined, gathered CPUID leaf/word into scattered features spread out amongst multiple Linux-defined words.
Yes, kvm wants the bits just as they are in the CPUID leafs from the hw.
Quoted 7 lines
It's mostly historical; before the
kvm_cpu_capsconcept was introduced, the code had grown organically to include bothboot_cpu_dataand raw CPUID info. The vast, vast majority of the time, doing CPUID is likely redundant. But, as noted in commitd8577a4c238f("KVM: x86: Do host CPUID at load time to mask KVM cpu caps"), the code is quite cheap and runs once at KVM load. My argument back then was, and still is, that an extra bit of paranoia is justified since the code and operations are quite nearly free.
Ok.
Quoted 8 lines
This particular dependency can be broken, and quite easily at that. Rather than
memcpy()boot_cpu_data.x86_capability, it's trivially easy to redefine theF()macro to invokeboot_cpu_has(), which would allow dropping thememcpy(). The big downside, and why I didn't post the code, is that doing so means every feature routed throughF()requires some form of BT+Jcc (or CMOVcc) sequence, whereas themempcy()approach allows theF()features to be encoded as a single literal by the compiler.
Quoted 8 lines
From a latency perspective, the extra code is negligible. The big issue is that all those extra checks add 2k+ bytes of code. Eliminating the
mempcy()doesn't actually break KVM's dependency on the bit offsets, so we'd be bloating kvm.ko by a noticeable amount without providing substantial value.And, this behavior is mostly opportunistic; the true justification/motiviation for taking a dependency on the X86_FEATURE_* bit offsets is for communication with userspace, querying the guest CPU model, and runtime checks.
Ok, I guess we'll try to find a middle ground here and not let stuff grow too ugly to live.
Quoted 2 lines
It's effectively for communication with userspace. Userspace, via
ioctl(), dictates the vCPU model to KVM, including the exact CPUID results.
And using the CPUID leafs with the exact bit positions is sort of an "interface" there, I see.
Quoted 9 lines
to properly
virtualize/emulate the defined vCPU model, KVM must query the dictated CPUID results to determine what features are supported, what guest operations should fault, etc... E.g. if the vCPU model, via CPUID, states that SMEP isn't supported then KVM needs to inject a #GP if the guest attempts to set CR4.SMEP.KVM also uses the hardware-defined CPUID ABI to advertise which features are supported by both hardware and KVM. This is the
kvm_cpu_capstuff, where KVM readsboot_cpu_datato see what features were enabled by the kernel.
Right.
Quoted 5 lines
It would be possible for KVM to break the dependency on X86_FEATURE_* bit offsets by defining a translation layer, but I strongly feel that adding manual translations will do more harm than good as it increases the odds of us botching a translation or using the wrong feature flag, creates potential namespace conflicts, etc...
Ok, lemme see if we might encounter more issues down the road...
+enum kvm_only_cpuid_leafs {
+ CPUID_12_EAX = NCAPINTS,
+ NR_KVM_CPU_CAPS,
+
+ NKVMCAPINTS = NR_KVM_CPU_CAPS - NCAPINTS,
+};
+
What happens when we decide to allocate a separate leaf for CPUID_12_EAX down the road?
You do it already here
Subject: [PATCH 04/13] x86/cpufeatures: Assign dedicated feature word for AMD mem encryption
for the AMD leaf.
I'm thinking this way around - from scattered to a hw one - should be ok because that should work easily. The other way around, taking a hw leaf and scattering it around x86_capability[] array elems would probably be nasty but with your change that should work too.
Yah, I'm just hypothesizing here - I don't think this "other way around" will ever happen...
Hmm, yap, I can cautiously say that with your change we should be ok...
Thx.
Signature · 4 lines
Regards/Gruss,
Boris.
https://people.kernel.org/tglx/notes-about-netiquetteSean Christopherson
maintainerRe: [RFC PATCH 04/23] x86/cpufeatures: Add SGX1 and SGX2 sub-features
Quoted 17 lines · Borislav Petkov
Quoted 5 lines · Sean Christopherson
It would be possible for KVM to break the dependency on X86_FEATURE_* bit offsets by defining a translation layer, but I strongly feel that adding manual translations will do more harm than good as it increases the odds of us botching a translation or using the wrong feature flag, creates potential namespace conflicts, etc...
Ok, lemme see if we might encounter more issues down the road...
arch/x86/kvm/cpuid.h +7
-0+enum kvm_only_cpuid_leafs { +CPUID_12_EAX =NCAPINTS , +NR_KVM_CPU_CAPS , + +NKVMCAPINTS =NR_KVM_CPU_CAPS -NCAPINTS , +}; +
Quoted 2 lines
What happens when we decide to allocate a separate leaf for
CPUID_12_EAXdown the road?
Well, mechanically, that would generate a build failure if the kernel does the obvious things and names the 'enum cpuid_leafs' entry CPUID_12_EAX. That would be an obvious clue that KVM should be updated.
If the kernel named the enum entry something different, and we botched the code review, KVM would continue to work, but would unnecessarily copy the bits it
cares about to its own word. E.g. the boot_cpu_has() checks and translation to
__X86_FEATURE_* would still be valid. As far as failure modes go, that's not terrible.
Quoted 23 lines
You do it already here
Subject: [PATCH 04/13] x86/cpufeatures: Assign dedicated feature word for AMD mem encryption
for the AMD leaf.
I'm thinking this way around - from scattered to a hw one - should be ok because that should work easily. The other way around, taking a hw leaf and scattering it around x86_capability[] array elems would probably be nasty but with your change that should work too.
Yah, I'm just hypothesizing here - I don't think this "other way around" will ever happen...
Hmm, yap, I can cautiously say that with your change we should be ok...
Thx.
Signature · 4 lines
Regards/Gruss, Boris. https://people.kernel.org/tglx/notes-about-netiquette
Jarkko Sakkinen
maintainerRe: [RFC PATCH 01/23] x86/sgx: Split out adding EPC page to free list to separate helper
In reply to Kai HuangQuoted 8 lines · Kai Huang
From: Sean Christopherson <[email protected]>
SGX virtualization requires to allocate "raw" EPC and use it as virtual EPC for SGX guest. Unlike EPC used by SGX driver, virtual EPC doesn't track how EPC pages are used in VM, e.g. (de)construction of enclaves, so it cannot guarantee EREMOVE success, e.g. it doesn't have a priori knowledge of which pages are SECS with non-zero child counts.
Quoted 6 lines
Split
sgx_free_page()into two parts so that the "add to free list" part can be used by virtual EPC without having to modify the EREMOVE logic insgx_free_page().Signed-off-by Sean Christopherson <[email protected]>Signed-off-by Kai Huang <[email protected]>
I have a better idea with the same outcome for KVM.
https://lore.kernel.org/linux-sgx/[email protected]/T/#t
/Jarkko
Re: [RFC PATCH 01/23] x86/sgx: Split out adding EPC page to free list to separate helper
Quoted 13 lines · On Tue, 12 Jan 2021 00:38:40 +0200 Jarkko Sakkinen wrote:
Quoted 12 lines · Kai Huang
From: Sean Christopherson <[email protected]>
SGX virtualization requires to allocate "raw" EPC and use it as virtual EPC for SGX guest. Unlike EPC used by SGX driver, virtual EPC doesn't track how EPC pages are used in VM, e.g. (de)construction of enclaves, so it cannot guarantee EREMOVE success, e.g. it doesn't have a priori knowledge of which pages are SECS with non-zero child counts.
Split
sgx_free_page()into two parts so that the "add to free list" part can be used by virtual EPC without having to modify the EREMOVE logic insgx_free_page().
Quoted 6 lines
Quoted 2 lines
Signed-off-by Sean Christopherson <[email protected]>Signed-off-by Kai Huang <[email protected]>I have a better idea with the same outcome for KVM.
I agree with your patch this one can be replaced. I'll include your patch in next version, and once it is upstreamed, it can be removed in my series.
Sean, please let me know if you have objection.
Re: [RFC PATCH 02/23] x86/sgx: Add enum for SGX_CHILD_PRESENT error code
In reply to Kai HuangQuoted 8 lines · Kai Huang
From: Sean Christopherson <[email protected]>
SGX virtualization requires to allocate "raw" EPC and use it as "virtual EPC" for SGX guest. Unlike EPC used by SGX driver, virtual EPC doesn't track how EPC pages are used in VM, e.g. (de)construction of enclaves, so it cannot guarantee EREMOVE success, e.g. it doesn't have a priori knowledge of which pages are SECS with non-zero child counts.
Quoted 5 lines
Add
SGX_CHILD_PRESENTfor use by SGX virtualization to assert EREMOVE failures are expected, but only due toSGX_CHILD_PRESENT.Signed-off-by Sean Christopherson <[email protected]>Signed-off-by Kai Huang <[email protected]>
/Jarkko
Re: [RFC PATCH 02/23] x86/sgx: Add enum for SGX_CHILD_PRESENT error code
Quoted 9 lines · On Tue, 12 Jan 2021 01:32:19 +0200 Jarkko Sakkinen wrote:
Quoted 8 lines · Kai Huang
From: Sean Christopherson <[email protected]>
SGX virtualization requires to allocate "raw" EPC and use it as "virtual EPC" for SGX guest. Unlike EPC used by SGX driver, virtual EPC doesn't track how EPC pages are used in VM, e.g. (de)construction of enclaves, so it cannot guarantee EREMOVE success, e.g. it doesn't have a priori knowledge of which pages are SECS with non-zero child counts.
Quoted 7 lines
Quoted 5 lines
Add
SGX_CHILD_PRESENTfor use by SGX virtualization to assert EREMOVE failures are expected, but only due toSGX_CHILD_PRESENT.Signed-off-by Sean Christopherson <[email protected]>Signed-off-by Kai Huang <[email protected]>Acked-by Jarkko Sakkinen <[email protected]>
Thanks Jarkko.
Dave suggested to change patch subject to explicitly call out hardware error code:
Add SGX_CHILD_PRESENT hardware error code
I suppose this also works for you, and I can have your Acked-by after I changed that in v2?
Thanks,
-Kai
Jarkko Sakkinen
maintainerRe: [RFC PATCH 03/23] x86/sgx: Introduce virtual EPC for use by KVM guests
In reply to Kai HuangQuoted 29 lines · Kai Huang
From: Sean Christopherson <[email protected]>
Add a misc device
/dev/sgx_virt_epcto allow userspace to allocate "raw" EPC without an associated enclave. The intended and only known use case for raw EPC allocation is to expose EPC to a KVM guest, hence thevirt_epcmoniker, virt.{c,h} files andX86_SGX_VIRTUALIZATIONKconfig.Modify
sgx_init()to always try to initialize virtual EPC driver, even when SGX driver is disabled due to SGX Launch Control is in locked mode, or not present at all, since SGX virtualization allows to expose SGX to guests that support non-LC configurations.Implement the "raw" EPC allocation in the x86 core-SGX subsystem via
/dev/sgx_virt_epcrather than in KVM. Doing so has two major advantages:- Does not require changes to KVM's uAPI, e.g. EPC gets handled as just another memory backend for guests.
- EPC management is wholly contained in the SGX subsystem, e.g. SGX does not have to export any symbols, changes to reclaim flows don't need to be routed through KVM, SGX's dirty laundry doesn't have to get aired out for the world to see, and so on and so forth.
The virtual EPC allocated to guests is currently not reclaimable, due to oversubscription of EPC for KVM guests is not currently supported. Due to the complications of handling reclaim conflicts between guest and host, KVM EPC oversubscription is significantly more complex than basic support for SGX virtualization.
Quoted 3 lines
Signed-off-by Sean Christopherson <[email protected]>Co-developed-by Kai Huang <[email protected]>Signed-off-by Kai Huang <[email protected]>
The commit message does not describe the code changes. It should have an understandable explanation of fops. There is nothing about the implementation right now.
/Jarkko
Quoted 361 lines
--- arch/x86/Kconfig | 12 ++ arch/x86/kernel/cpu/sgx/Makefile | 1 + arch/x86/kernel/cpu/sgx/main.c | 5 +- arch/x86/kernel/cpu/sgx/virt.c | 263 +++++++++++++++++++++++++++++++ arch/x86/kernel/cpu/sgx/virt.h | 14 ++ 5 files changed, 294 insertions(+), 1 deletion(-) create mode 100644 arch/x86/kernel/cpu/sgx/virt.c create mode 100644 arch/x86/kernel/cpu/sgx/virt.harch/x86/Kconfig +12
-0@@ -1947,6 +1947,18 @@ config X86_SGX If unsure, say N. +config X86_SGX_VIRTUALIZATION + bool "Software Guard eXtensions (SGX) Virtualization" + depends on X86_SGX && KVM_INTEL + help + + Enables KVM guests to create SGX enclaves. + + This includes support to expose "raw" unreclaimable enclave memory to + guests via a device node, e.g. /dev/sgx_virt_epc. + + If unsure, say N. + config EFI bool "EFI runtime service support" depends on ACPIarch/x86/kernel/cpu/sgx/Makefile +1
-0@@ -3,3 +3,4 @@ obj-y += \ encl.o \ ioctl.o \ main.o +obj-$(CONFIG_X86_SGX_VIRTUALIZATION) += virt.oarch/x86/kernel/cpu/sgx/main.c +4
-1@@ -9,9 +9,11 @@#include <linux/sched/mm.h> #include <linux/sched/signal.h> #include <linux/slab.h> +#include "arch.h" #include "driver.h" #include "encl.h" #include "encls.h" +#include "virt.h" struct sgx_epc_section sgx_epc_sections[SGX_MAX_EPC_SECTIONS ];static int sgx_nr_epc_sections;@@ -726,7 +728,8 @@ static void __init sgx_init(void)if (!sgx_page_reclaimer_init ()) goto err_page_cache; - ret =sgx_drv_init (); +/* Success if the native *or* virtual EPC driver initialized cleanly. */ + ret = !!sgx_drv_init () & !!sgx_virt_epc_init ();if (ret) goto err_kthread;arch/x86/kernel/cpu/sgx/virt.c new file +263
-0@@ -0,0 +1,263 @@ +// SPDX-License-Identifier: GPL-2.0 +/* Copyright(c) 2016-20 Intel Corporation. */ + +#include <linux/miscdevice.h> +#include <linux/mm.h> +#include <linux/mman.h> +#include <linux/sched/mm.h> +#include <linux/sched/signal.h> +#include <linux/slab.h> +#include <linux/xarray.h> +#include <asm/sgx.h> +#include <uapi/asm/sgx.h> + +#include "encls.h" +#include "sgx.h" +#include "virt.h" + +struct sgx_virt_epc { +struct xarray page_array; +struct mutex lock; +struct mm_struct *mm; +}; + +static struct mutex virt_epc_lock; +static struct list_head virt_epc_zombie_pages; + +static int __sgx_virt_epc_fault (struct sgx_virt_epc *epc, +struct vm_area_struct *vma,unsigned long addr) +{ +struct sgx_epc_page *epc_page; +unsigned long index, pfn; +int ret; + +/* epc->lock must already have been hold */ + +/* Calculate index of EPC page in virtual EPC's page_array */ + index = vma->vm_pgoff +PFN_DOWN (addr - vma->vm_start); + + epc_page =xa_load (&epc->page_array, index); +if (epc_page) +return 0 ; + + epc_page =sgx_alloc_epc_page (epc, false); +if (IS_ERR (epc_page)) +return PTR_ERR (epc_page); + + ret =xa_err (xa_store (&epc->page_array, index, epc_page,GFP_KERNEL )); +if (ret) + goto err_free; + + pfn =PFN_DOWN (sgx_get_epc_phys_addr (epc_page)); + + ret =vmf_insert_pfn (vma, addr, pfn); +if (ret !=VM_FAULT_NOPAGE ) { + ret = -EFAULT ; + goto err_delete; + } + +return 0 ; + +err_delete: +xa_erase (&epc->page_array, index); +err_free: +sgx_free_epc_page (epc_page); +return ret; +} + +static vm_fault_t sgx_virt_epc_fault (struct vm_fault *vmf) +{ +struct vm_area_struct *vma = vmf->vma; +struct sgx_virt_epc *epc = vma->vm_private_data; +int ret; + +mutex_lock (&epc->lock); + ret =__sgx_virt_epc_fault (epc, vma, vmf->address); +mutex_unlock (&epc->lock); + +if (!ret) +return VM_FAULT_NOPAGE ; + +if (ret == -EBUSY && (vmf->flags &FAULT_FLAG_ALLOW_RETRY )) { +mmap_read_unlock (vma->vm_mm); +return VM_FAULT_RETRY ; + } + +return VM_FAULT_SIGBUS ; +} + +const struct vm_operations_struct sgx_virt_epc_vm_ops = { + .fault = sgx_virt_epc_fault, +}; + +static int sgx_virt_epc_mmap (struct file *file,struct vm_area_struct *vma) +{ +struct sgx_virt_epc *epc = file->private_data; + +if (!(vma->vm_flags &VM_SHARED )) +return -EINVAL ; + +/* +* Don't allow mmap() from child after fork(), since child and parent +* cannot map to the same EPC. +*/ +if (vma->vm_mm != epc->mm) +return -EINVAL ; + + vma->vm_ops = &sgx_virt_epc_vm_ops; +/* Don't copy VMA in fork() */ + vma->vm_flags |=VM_PFNMAP |VM_IO |VM_DONTDUMP |VM_DONTCOPY ; + vma->vm_private_data = file->private_data; + +return 0 ; +} + +static int sgx_virt_epc_free_page (struct sgx_epc_page *epc_page) +{ +int ret; + +if (!epc_page) +return 0 ; + +/* +* Explicitly EREMOVE virtual EPC page. Virtual EPC is only used by +* guest, and in normal condition guest should have done EREMOVE for +* all EPC pages before they are freed here. But it's possible guest +* is killed or crashed unnormally in which case EREMOVE has not been +* done. Do EREMOVE unconditionally here to cover both cases, because +* it's not possible to tell whether guest has done EREMOVE, since +* virtual EPC page status is not tracked. And it is fine to EREMOVE +* EPC page multiple times. +*/ + ret =__eremove (sgx_get_epc_virt_addr (epc_page)); +if (ret) { +/* +* Only SGX_CHILD_PRESENT is expected, which is because of +* EREMOVE-ing an SECS still with child, in which case it can +* be handled by EREMOVE-ing the SECS again after all pages in +* virtual EPC have been EREMOVE-ed. See comments in below in +* sgx_virt_epc_release(). +*/ +WARN_ON_ONCE (ret !=SGX_CHILD_PRESENT ); +return ret; + } + +__sgx_free_epc_page (epc_page); +return 0 ; +} + +static int sgx_virt_epc_release (struct inode *inode,struct file *file) +{ +struct sgx_virt_epc *epc = file->private_data; +struct sgx_epc_page *epc_page, *tmp, *entry; +unsigned long index; + +LIST_HEAD (secs_pages); + +mmdrop (epc->mm); + +xa_for_each (&epc->page_array, index, entry) { +/* +* Virtual EPC pages are not tracked, so it's possible for +* EREMOVE to fail due to, e.g. a SECS page still has children +* if guest was shutdown unexpectedly. If it is the case, leave +* it in the xarray and retry EREMOVE below later. +*/ +if (sgx_virt_epc_free_page (entry)) +continue ; + +xa_erase (&epc->page_array, index); + } + +/* +* Retry all failed pages after iterating through the entire tree, at +* which point all children should be removed and the SECS pages can be +* nuked as well...unless userspace has exposed multiple instance of +* virtual EPC to a single VM. +*/ +xa_for_each (&epc->page_array, index, entry) { + epc_page = entry; +/* +* Error here means that EREMOVE failed due to a SECS page +* still has child on *another* EPC instance. Put it to a +* temporary SECS list which will be spliced to 'zombie page +* list' and will be EREMOVE-ed again when freeing another +* virtual EPC instance. +*/ +if (sgx_virt_epc_free_page (epc_page)) +list_add_tail (&epc_page->list, &secs_pages); + +xa_erase (&epc->page_array, index); + } + +/* +* Third time's a charm. Try to EREMOVE zombie SECS pages from virtual +* EPC instances that were previously released, i.e. free SECS pages +* that were in limbo due to having children in *this* EPC instance. +*/ +mutex_lock (&virt_epc_lock); +list_for_each_entry_safe (epc_page, tmp, &virt_epc_zombie_pages, list) { +/* +* Speculatively remove the page from the list of zombies, if +* the page is successfully EREMOVE it will be added to the +* list of free pages. If EREMOVE fails, throw the page on the +* local list, which will be spliced on at the end. +*/ +list_del (&epc_page->list); + +if (sgx_virt_epc_free_page (epc_page)) +list_add_tail (&epc_page->list, &secs_pages); + } + +if (!list_empty (&secs_pages)) +list_splice_tail (&secs_pages, &virt_epc_zombie_pages); +mutex_unlock (&virt_epc_lock); + +kfree (epc); + +return 0 ; +} + +static int sgx_virt_epc_open (struct inode *inode,struct file *file) +{ +struct sgx_virt_epc *epc; + + epc =kzalloc (sizeof (struct sgx_virt_epc ),GFP_KERNEL ); +if (!epc) +return -ENOMEM ; +/* +* Keep the current->mm to virtual EPC. It will be checked in +* sgx_virt_epc_mmap() to prevent, in case of fork, child being +* able to mmap() to the same virtual EPC pages. +*/ +mmgrab (current->mm); + epc->mm = current->mm; +mutex_init (&epc->lock); +xa_init (&epc->page_array); + + file->private_data = epc; + +return 0 ; +} + +static const struct file_operations sgx_virt_epc_fops = { + .owner =THIS_MODULE , + .open = sgx_virt_epc_open, + .release = sgx_virt_epc_release, + .mmap = sgx_virt_epc_mmap, +}; + +static struct miscdevice sgx_virt_epc_dev = { + .minor =MISC_DYNAMIC_MINOR , + .name ="sgx_virt_epc" , + .nodename ="sgx_virt_epc" , + .fops = &sgx_virt_epc_fops, +}; + +int __init sgx_virt_epc_init (void ) +{ +INIT_LIST_HEAD (&virt_epc_zombie_pages); +mutex_init (&virt_epc_lock); + +return misc_register (&sgx_virt_epc_dev); +}arch/x86/kernel/cpu/sgx/virt.h new file +14
-0@@ -0,0 +1,14 @@ +/* SPDX-License-Identifier: (GPL-2.0 OR BSD-3-Clause) */ +#ifndef _ASM_X86_SGX_VIRT_H +#define _ASM_X86_SGX_VIRT_H + +#ifdef CONFIG_X86_SGX_VIRTUALIZATION +int __init sgx_virt_epc_init (void ); +#else +static inline int __init sgx_virt_epc_init (void ) +{ +return -ENODEV ; +} +#endif + +#endif /* _ASM_X86_SGX_VIRT_H */ Signature · 1 line
2.29.2
Re: [RFC PATCH 04/23] x86/cpufeatures: Add SGX1 and SGX2 sub-features
In reply to Kai HuangQuoted 11 lines · Kai Huang
From: Sean Christopherson <[email protected]>
Add a feature word to hold SGX features enumerated via CPUID.0x12.0x0, along with flags for SGX1 and SGX2. As part of virtualizing SGX, KVM needs to expose the SGX CPUID leafs to its guest. SGX1 and SGX2 need to be in a dedicated feature word so that they can be queried via KVM's reverse CPUID lookup to properly emulate the expected guest behavior.
Signed-off-by Sean Christopherson <[email protected]>[Kai: Also clear SGX1 and SGX2 bits in clear_sgx_caps().]Signed-off-by Kai Huang <[email protected]>
/Jarkko
Jarkko Sakkinen
maintainerRe: [RFC PATCH 02/23] x86/sgx: Add enum for SGX_CHILD_PRESENT error code
In reply to Dave HansenQuoted 8 lines · Dave Hansen
Quoted 2 lines · Kai Huang
Add
SGX_CHILD_PRESENTfor use by SGX virtualization to assert EREMOVE failures are expected, but only due toSGX_CHILD_PRESENT.This dances around the fact that this is an architectural error-code. Could that be explicit? Maybe the subject should be:
Add
SGX_CHILD_PRESENThardware error code
Yeah, a valid point. Please, change this.
/Jarkko
Re: [RFC PATCH 03/23] x86/sgx: Introduce virtual EPC for use by KVM guests
In reply to Jarkko SakkinenQuoted 30 lines · On Tue, 12 Jan 2021 01:38:23 +0200 Jarkko Sakkinen wrote:
Quoted 29 lines · Kai Huang
From: Sean Christopherson <[email protected]>
Add a misc device
/dev/sgx_virt_epcto allow userspace to allocate "raw" EPC without an associated enclave. The intended and only known use case for raw EPC allocation is to expose EPC to a KVM guest, hence thevirt_epcmoniker, virt.{c,h} files andX86_SGX_VIRTUALIZATIONKconfig.Modify
sgx_init()to always try to initialize virtual EPC driver, even when SGX driver is disabled due to SGX Launch Control is in locked mode, or not present at all, since SGX virtualization allows to expose SGX to guests that support non-LC configurations.Implement the "raw" EPC allocation in the x86 core-SGX subsystem via
/dev/sgx_virt_epcrather than in KVM. Doing so has two major advantages:- Does not require changes to KVM's uAPI, e.g. EPC gets handled as just another memory backend for guests.
- EPC management is wholly contained in the SGX subsystem, e.g. SGX does not have to export any symbols, changes to reclaim flows don't need to be routed through KVM, SGX's dirty laundry doesn't have to get aired out for the world to see, and so on and so forth.
The virtual EPC allocated to guests is currently not reclaimable, due to oversubscription of EPC for KVM guests is not currently supported. Due to the complications of handling reclaim conflicts between guest and host, KVM EPC oversubscription is significantly more complex than basic support for SGX virtualization.
Quoted 7 lines
Quoted 3 lines
Signed-off-by Sean Christopherson <[email protected]>Co-developed-by Kai Huang <[email protected]>Signed-off-by Kai Huang <[email protected]>The commit message does not describe the code changes. It should have an understandable explanation of fops. There is nothing about the implementation right now.
Thanks for feedback. Does "understabdable explanation of fops" mean I should add one sentence to say, for instance: "userspace hypervisor should open the /dev/sgx_virt_epc, use mmap() to get a valid address range, and then use that address range to create KVM memory region"?
Or should I include an example of how to use /dev/sgx_virt_epc in userspace, for instance, below?
fd = open("/dev/sgx_virt_epc", O_RDWR);
void *addr = mmap(NULL, size, ..., fd);
/* userspace hypervisor uses addr, size to create KVM memory slot */
...I dug the SGX driver side to understand what should I add, but in below commit I don't see description of fops either:
commit 3fe0778edac8628637e2fd23835996523b1a3372 Author: Jarkko Sakkinen <[email protected]> Date: Fri Nov 13 00:01:22 2020 +0200 x86/sgx: Add an SGX misc driver interface
Quoted 364 lines
/Jarkko
Quoted 361 lines
--- arch/x86/Kconfig | 12 ++ arch/x86/kernel/cpu/sgx/Makefile | 1 + arch/x86/kernel/cpu/sgx/main.c | 5 +- arch/x86/kernel/cpu/sgx/virt.c | 263 +++++++++++++++++++++++++++++++ arch/x86/kernel/cpu/sgx/virt.h | 14 ++ 5 files changed, 294 insertions(+), 1 deletion(-) create mode 100644 arch/x86/kernel/cpu/sgx/virt.c create mode 100644 arch/x86/kernel/cpu/sgx/virt.harch/x86/Kconfig +12
-0@@ -1947,6 +1947,18 @@ config X86_SGX If unsure, say N. +config X86_SGX_VIRTUALIZATION + bool "Software Guard eXtensions (SGX) Virtualization" + depends on X86_SGX && KVM_INTEL + help + + Enables KVM guests to create SGX enclaves. + + This includes support to expose "raw" unreclaimable enclave memory to + guests via a device node, e.g. /dev/sgx_virt_epc. + + If unsure, say N. + config EFI bool "EFI runtime service support" depends on ACPIarch/x86/kernel/cpu/sgx/Makefile +1
-0@@ -3,3 +3,4 @@ obj-y += \ encl.o \ ioctl.o \ main.o +obj-$(CONFIG_X86_SGX_VIRTUALIZATION) += virt.oarch/x86/kernel/cpu/sgx/main.c +4
-1@@ -9,9 +9,11 @@#include <linux/sched/mm.h> #include <linux/sched/signal.h> #include <linux/slab.h> +#include "arch.h" #include "driver.h" #include "encl.h" #include "encls.h" +#include "virt.h" struct sgx_epc_section sgx_epc_sections[SGX_MAX_EPC_SECTIONS ];static int sgx_nr_epc_sections;@@ -726,7 +728,8 @@ static void __init sgx_init(void)if (!sgx_page_reclaimer_init ()) goto err_page_cache; - ret =sgx_drv_init (); +/* Success if the native *or* virtual EPC driver initialized cleanly. */ + ret = !!sgx_drv_init () & !!sgx_virt_epc_init ();if (ret) goto err_kthread;arch/x86/kernel/cpu/sgx/virt.c new file +263
-0@@ -0,0 +1,263 @@ +// SPDX-License-Identifier: GPL-2.0 +/* Copyright(c) 2016-20 Intel Corporation. */ + +#include <linux/miscdevice.h> +#include <linux/mm.h> +#include <linux/mman.h> +#include <linux/sched/mm.h> +#include <linux/sched/signal.h> +#include <linux/slab.h> +#include <linux/xarray.h> +#include <asm/sgx.h> +#include <uapi/asm/sgx.h> + +#include "encls.h" +#include "sgx.h" +#include "virt.h" + +struct sgx_virt_epc { +struct xarray page_array; +struct mutex lock; +struct mm_struct *mm; +}; + +static struct mutex virt_epc_lock; +static struct list_head virt_epc_zombie_pages; + +static int __sgx_virt_epc_fault (struct sgx_virt_epc *epc, +struct vm_area_struct *vma,unsigned long addr) +{ +struct sgx_epc_page *epc_page; +unsigned long index, pfn; +int ret; + +/* epc->lock must already have been hold */ + +/* Calculate index of EPC page in virtual EPC's page_array */ + index = vma->vm_pgoff +PFN_DOWN (addr - vma->vm_start); + + epc_page =xa_load (&epc->page_array, index); +if (epc_page) +return 0 ; + + epc_page =sgx_alloc_epc_page (epc, false); +if (IS_ERR (epc_page)) +return PTR_ERR (epc_page); + + ret =xa_err (xa_store (&epc->page_array, index, epc_page,GFP_KERNEL )); +if (ret) + goto err_free; + + pfn =PFN_DOWN (sgx_get_epc_phys_addr (epc_page)); + + ret =vmf_insert_pfn (vma, addr, pfn); +if (ret !=VM_FAULT_NOPAGE ) { + ret = -EFAULT ; + goto err_delete; + } + +return 0 ; + +err_delete: +xa_erase (&epc->page_array, index); +err_free: +sgx_free_epc_page (epc_page); +return ret; +} + +static vm_fault_t sgx_virt_epc_fault (struct vm_fault *vmf) +{ +struct vm_area_struct *vma = vmf->vma; +struct sgx_virt_epc *epc = vma->vm_private_data; +int ret; + +mutex_lock (&epc->lock); + ret =__sgx_virt_epc_fault (epc, vma, vmf->address); +mutex_unlock (&epc->lock); + +if (!ret) +return VM_FAULT_NOPAGE ; + +if (ret == -EBUSY && (vmf->flags &FAULT_FLAG_ALLOW_RETRY )) { +mmap_read_unlock (vma->vm_mm); +return VM_FAULT_RETRY ; + } + +return VM_FAULT_SIGBUS ; +} + +const struct vm_operations_struct sgx_virt_epc_vm_ops = { + .fault = sgx_virt_epc_fault, +}; + +static int sgx_virt_epc_mmap (struct file *file,struct vm_area_struct *vma) +{ +struct sgx_virt_epc *epc = file->private_data; + +if (!(vma->vm_flags &VM_SHARED )) +return -EINVAL ; + +/* +* Don't allow mmap() from child after fork(), since child and parent +* cannot map to the same EPC. +*/ +if (vma->vm_mm != epc->mm) +return -EINVAL ; + + vma->vm_ops = &sgx_virt_epc_vm_ops; +/* Don't copy VMA in fork() */ + vma->vm_flags |=VM_PFNMAP |VM_IO |VM_DONTDUMP |VM_DONTCOPY ; + vma->vm_private_data = file->private_data; + +return 0 ; +} + +static int sgx_virt_epc_free_page (struct sgx_epc_page *epc_page) +{ +int ret; + +if (!epc_page) +return 0 ; + +/* +* Explicitly EREMOVE virtual EPC page. Virtual EPC is only used by +* guest, and in normal condition guest should have done EREMOVE for +* all EPC pages before they are freed here. But it's possible guest +* is killed or crashed unnormally in which case EREMOVE has not been +* done. Do EREMOVE unconditionally here to cover both cases, because +* it's not possible to tell whether guest has done EREMOVE, since +* virtual EPC page status is not tracked. And it is fine to EREMOVE +* EPC page multiple times. +*/ + ret =__eremove (sgx_get_epc_virt_addr (epc_page)); +if (ret) { +/* +* Only SGX_CHILD_PRESENT is expected, which is because of +* EREMOVE-ing an SECS still with child, in which case it can +* be handled by EREMOVE-ing the SECS again after all pages in +* virtual EPC have been EREMOVE-ed. See comments in below in +* sgx_virt_epc_release(). +*/ +WARN_ON_ONCE (ret !=SGX_CHILD_PRESENT ); +return ret; + } + +__sgx_free_epc_page (epc_page); +return 0 ; +} + +static int sgx_virt_epc_release (struct inode *inode,struct file *file) +{ +struct sgx_virt_epc *epc = file->private_data; +struct sgx_epc_page *epc_page, *tmp, *entry; +unsigned long index; + +LIST_HEAD (secs_pages); + +mmdrop (epc->mm); + +xa_for_each (&epc->page_array, index, entry) { +/* +* Virtual EPC pages are not tracked, so it's possible for +* EREMOVE to fail due to, e.g. a SECS page still has children +* if guest was shutdown unexpectedly. If it is the case, leave +* it in the xarray and retry EREMOVE below later. +*/ +if (sgx_virt_epc_free_page (entry)) +continue ; + +xa_erase (&epc->page_array, index); + } + +/* +* Retry all failed pages after iterating through the entire tree, at +* which point all children should be removed and the SECS pages can be +* nuked as well...unless userspace has exposed multiple instance of +* virtual EPC to a single VM. +*/ +xa_for_each (&epc->page_array, index, entry) { + epc_page = entry; +/* +* Error here means that EREMOVE failed due to a SECS page +* still has child on *another* EPC instance. Put it to a +* temporary SECS list which will be spliced to 'zombie page +* list' and will be EREMOVE-ed again when freeing another +* virtual EPC instance. +*/ +if (sgx_virt_epc_free_page (epc_page)) +list_add_tail (&epc_page->list, &secs_pages); + +xa_erase (&epc->page_array, index); + } + +/* +* Third time's a charm. Try to EREMOVE zombie SECS pages from virtual +* EPC instances that were previously released, i.e. free SECS pages +* that were in limbo due to having children in *this* EPC instance. +*/ +mutex_lock (&virt_epc_lock); +list_for_each_entry_safe (epc_page, tmp, &virt_epc_zombie_pages, list) { +/* +* Speculatively remove the page from the list of zombies, if +* the page is successfully EREMOVE it will be added to the +* list of free pages. If EREMOVE fails, throw the page on the +* local list, which will be spliced on at the end. +*/ +list_del (&epc_page->list); + +if (sgx_virt_epc_free_page (epc_page)) +list_add_tail (&epc_page->list, &secs_pages); + } + +if (!list_empty (&secs_pages)) +list_splice_tail (&secs_pages, &virt_epc_zombie_pages); +mutex_unlock (&virt_epc_lock); + +kfree (epc); + +return 0 ; +} + +static int sgx_virt_epc_open (struct inode *inode,struct file *file) +{ +struct sgx_virt_epc *epc; + + epc =kzalloc (sizeof (struct sgx_virt_epc ),GFP_KERNEL ); +if (!epc) +return -ENOMEM ; +/* +* Keep the current->mm to virtual EPC. It will be checked in +* sgx_virt_epc_mmap() to prevent, in case of fork, child being +* able to mmap() to the same virtual EPC pages. +*/ +mmgrab (current->mm); + epc->mm = current->mm; +mutex_init (&epc->lock); +xa_init (&epc->page_array); + + file->private_data = epc; + +return 0 ; +} + +static const struct file_operations sgx_virt_epc_fops = { + .owner =THIS_MODULE , + .open = sgx_virt_epc_open, + .release = sgx_virt_epc_release, + .mmap = sgx_virt_epc_mmap, +}; + +static struct miscdevice sgx_virt_epc_dev = { + .minor =MISC_DYNAMIC_MINOR , + .name ="sgx_virt_epc" , + .nodename ="sgx_virt_epc" , + .fops = &sgx_virt_epc_fops, +}; + +int __init sgx_virt_epc_init (void ) +{ +INIT_LIST_HEAD (&virt_epc_zombie_pages); +mutex_init (&virt_epc_lock); + +return misc_register (&sgx_virt_epc_dev); +}arch/x86/kernel/cpu/sgx/virt.h new file +14
-0@@ -0,0 +1,14 @@ +/* SPDX-License-Identifier: (GPL-2.0 OR BSD-3-Clause) */ +#ifndef _ASM_X86_SGX_VIRT_H +#define _ASM_X86_SGX_VIRT_H + +#ifdef CONFIG_X86_SGX_VIRTUALIZATION +int __init sgx_virt_epc_init (void ); +#else +static inline int __init sgx_virt_epc_init (void ) +{ +return -ENODEV ; +} +#endif + +#endif /* _ASM_X86_SGX_VIRT_H */ Signature · 1 line
2.29.2
Quoted 52 lines · On Mon, 11 Jan 2021 19:20:48 +0200 Jarkko Sakkinen wrote:
Quoted 40 lines · Kai Huang
--- Disclaimer ---
These patches were originally written by Sean Christopherson while at Intel. Now that Sean has left Intel, I (Kai) have taken over getting them upstream. This series needs more review before it can be merged. It is being posted publicly and under RFC so Sean and others can review it. Maintainers are safe ignoring it for now.
------------------
Hi all,
This series adds KVM SGX virtualization support. The first 12 patches starting with x86/sgx or x86/cpu.. are necessary changes to x86 and SGX core/driver to support KVM SGX virtualization, while the rest are patches to KVM subsystem.
Please help to review this series. Also I'd like to hear what is the proper way to merge this series, since it contains change to both x86/SGX and KVM subsystem. Any feedback is highly appreciated. And please let me know if I forgot to CC anyone, or anyone wants to be removed from CC. Thanks in advance!
This series is based against latest tip tree's x86/sgx branch. You can also get the code from tip branch of kvm-sgx repo on github:
https://github.com/intel/kvm-sgx.git tip
It also requires Qemu changes to create VM with SGX support. You can find Qemu repo here:
https://github.com/intel/qemu-sgx.git next
Please refer to README.md of above qemu-sgx repo for detail on how to create guest with SGX support. At meantime, for your quick reference you can use below command to create SGX guest:
#qemu-system-x86_64 -smp 4 -m 2G -drive file=<your_vm_image>,if=virtio \ -cpu host,+sgx_provisionkey \ -sgx-epc id=epc1,memdev=mem1 \ -object memory-backend-epc,id=mem1,size=64M,preallocQuoted 7 lines
Please note that the SGX relevant part is:
-cpu host,+sgx_provisionkey \ -sgx-epc id=epc1,memdev=mem1 \ -object memory-backend-epc,id=mem1,size=64M,preallocAnd you can change other parameters of your qemu command based on your needs.
Thanks a lot documenting these snippets to the cover letter. I dig these up from lore once my environment is working.
Quoted 7 lines
I'm setting up Arch based test environment with the eye on this patch set and generic Linux keyring patches:
https://git.kernel.org/pub/scm/linux/kernel/git/jarkko/arch.git/
Still have some minor bits to adjust before I can start deploying it for SGX testing. For this patch set I'll use two instances of it.
Thanks. Please let me know if you need anything more.
Quoted 23 lines
Quoted 8 lines
=========
KVM SGX virtualization Overview- Virtual EPC
"Virtual EPC" is the EPC section exposed by KVM to guest so SGX software in guest can discover it and use it to create SGX enclaves. KVM exposes SGX to
Virtual EPC is a representation of an EPC section. And there is no "the".
Quoted 11 lines
guest via CPUID, and exposes one or more "virtual EPC" sections for guest. The size of "virtual EPC" is passed as Qemu parameter when creating the guest, and the base address is calcualted internally according to guest's configuration.
To support virtual EPC, add a new misc device
/dev/sgx_virt_epcto SGX core/driver to allow userspace (Qemu) to allocate "raw" EPC, and use it as "virtual EPC" for guest. Obviously, unlike EPC allocated for host SGX driver, virtual EPC allocated via/dev/sgx_virt_epcdoesn't have enclave associated, and how virtual EPC is used by guest is compeletely controlled by guest's SGX software.
Quoted 2 lines
I think that
/dev/sgx_vepcwould be a clear enough name for the device. This text has now a bit confusing "terminology" related to this.
/dev/sgx_virt_epc may be clearer from userspace's perspective, for instance, if people see /dev/sgx_vepc, they may have to think about what it is, while /dev/sgx_virt_epc they may not.
But I don't have strong objection here. Does anyone has anything to say here?
Jarkko Sakkinen
maintainerRe: [RFC PATCH 02/23] x86/sgx: Add enum for SGX_CHILD_PRESENT error code
In reply to Kai HuangQuoted 18 lines · Kai Huang
Quoted 9 lines · On Tue, 12 Jan 2021 01:32:19 +0200 Jarkko Sakkinen wrote:
Quoted 8 lines · Kai Huang
From: Sean Christopherson <[email protected]>
SGX virtualization requires to allocate "raw" EPC and use it as "virtual EPC" for SGX guest. Unlike EPC used by SGX driver, virtual EPC doesn't track how EPC pages are used in VM, e.g. (de)construction of enclaves, so it cannot guarantee EREMOVE success, e.g. it doesn't have a priori knowledge of which pages are SECS with non-zero child counts.
Quoted 7 lines
Quoted 5 lines
Add
SGX_CHILD_PRESENTfor use by SGX virtualization to assert EREMOVE failures are expected, but only due toSGX_CHILD_PRESENT.Signed-off-by Sean Christopherson <[email protected]>Signed-off-by Kai Huang <[email protected]>Acked-by Jarkko Sakkinen <[email protected]>
Quoted 8 lines
Thanks Jarkko.
Dave suggested to change patch subject to explicitly call out hardware error code:
AddSGX_CHILD_PRESENThardware error codeI suppose this also works for you, and I can have your Acked-by after I changed that in v2?
Yeah, I agree with that.
/Jarkko
Jarkko Sakkinen
maintainerRe: [RFC PATCH 03/23] x86/sgx: Introduce virtual EPC for use by KVM guests
In reply to Kai HuangQuoted 44 lines · Kai Huang
Quoted 30 lines · On Tue, 12 Jan 2021 01:38:23 +0200 Jarkko Sakkinen wrote:
Quoted 29 lines · Kai Huang
From: Sean Christopherson <[email protected]>
Add a misc device
/dev/sgx_virt_epcto allow userspace to allocate "raw" EPC without an associated enclave. The intended and only known use case for raw EPC allocation is to expose EPC to a KVM guest, hence thevirt_epcmoniker, virt.{c,h} files andX86_SGX_VIRTUALIZATIONKconfig.Modify
sgx_init()to always try to initialize virtual EPC driver, even when SGX driver is disabled due to SGX Launch Control is in locked mode, or not present at all, since SGX virtualization allows to expose SGX to guests that support non-LC configurations.Implement the "raw" EPC allocation in the x86 core-SGX subsystem via
/dev/sgx_virt_epcrather than in KVM. Doing so has two major advantages:- Does not require changes to KVM's uAPI, e.g. EPC gets handled as just another memory backend for guests.
- EPC management is wholly contained in the SGX subsystem, e.g. SGX does not have to export any symbols, changes to reclaim flows don't need to be routed through KVM, SGX's dirty laundry doesn't have to get aired out for the world to see, and so on and so forth.
The virtual EPC allocated to guests is currently not reclaimable, due to oversubscription of EPC for KVM guests is not currently supported. Due to the complications of handling reclaim conflicts between guest and host, KVM EPC oversubscription is significantly more complex than basic support for SGX virtualization.
Quoted 7 lines
Quoted 3 lines
Signed-off-by Sean Christopherson <[email protected]>Co-developed-by Kai Huang <[email protected]>Signed-off-by Kai Huang <[email protected]>The commit message does not describe the code changes. It should have an understandable explanation of fops. There is nothing about the implementation right now.
Thanks for feedback. Does "understabdable explanation of fops" mean I should add one sentence to say, for instance: "userspace hypervisor should open the
/dev/sgx_virt_epc, usemmap()to get a valid address range, and then use that address range to create KVM memory region"?
Quoted 7 lines
Or should I include an example of how to use
/dev/sgx_virt_epcin userspace, for instance, below?fd = open("/dev/sgx_virt_epc", O_RDWR); void *addr = mmap(NULL, size, ..., fd); /* userspace hypervisor uses addr, size to create KVM memory slot */ ...
I would suggest just to describe them in few sentences. Just write how you understand them in one paragraph.
/Jarkko
Jarkko Sakkinen
maintainerQuoted 11 lines · Sean Christopherson
Quoted 5 lines · Jarkko Sakkinen
Quoted 2 lines · Kai Huang
- Does not require changes to KVM's uAPI, e.g. EPC gets handled as just another memory backend for guests.
Why this an advantage? No objection, just a question.
There are zero KVM changes required to support exposing EPC to a guest. KVM's MMU is completely ignorant of what physical backing is used for any given host virtual address. KVM has to be aware of various VM_* flags, e.g.
VM_PFNMAPandVM_IO, but that code is arch agnostic and is quite isolated.
Right, thanks for explanation.
Quoted 14 lines
Quoted 14 lines
Quoted 4 lines
- EPC management is wholly contained in the SGX subsystem, e.g. SGX does not have to export any symbols, changes to reclaim flows don't need to be routed through KVM, SGX's dirty laundry doesn't have to get aired out for the world to see, and so on and so forth.
No comments to this before understanding code changes better.
Quoted 6 lines
The virtual EPC allocated to guests is currently not reclaimable, due to reclaiming EPC from KVM guests is not currently supported. Due to the complications of handling reclaim conflicts between guest and host, KVM EPC oversubscription, which allows total virtual EPC size greater than physical EPC by being able to reclaiming guests' EPC, is significantly more complex than basic support for SGX virtualization.
Quoted 7 lines
Quoted 3 lines
I think it should be really in the center of the patch set description that this patch set implements segmentation of EPC, not oversubscription. It should be clear immediately. It's a core part of knowing "what I'm looking at".
Technically, it doesn't implement EPC segmentation of EPC. It implements non-reclaimable EPC allocation. Even that is somewhat untrue as the EPC can be forcefully reclaimed, but doing so will destroy the guest contents.
In SGX case, that isn't actually as a bad as a policy in high stress situations as with "normal" applications. Runtimes must expect dissappearance of the enclave at any point of time anyway...
Quoted 2 lines
Userspace can oversubscribe the EPC to KVM guests, but it would need to kill, migrate, or pause one or more VMs if the pool of physical EPC were exhausted.
Right.
Quoted 19 lines
Quoted 13 lines
Quoted 12 lines
- Support SGX virtualization without SGX Launch Control unlocked mode
Although SGX driver requires SGX Launch Control unlocked mode to work, SGX virtualization doesn't, since how enclave is created is completely controlled by guest SGX software, which is not necessarily linux. Therefore, this series allows KVM to expose SGX to guest even SGX Launch Control is in locked mode, or is not present at all. The reason is the goal of SGX virtualization, or virtualization in general, is to expose hardware feature to guest, but not to make assumption how guest will use it. Therefore, KVM should support SGX guest as long as hardware is able to, to have chance to support more potential use cases in cloud environment.
Quoted 5 lines
AFAIK the convergence point with the FLC was, and is that Linux never enables SGX with locked MSRs.
And I don't understand, if it is not fine to allow locked SGX for a *process*, why is it fine for a *virtual machine*? They have a lot same.
Quoted 8 lines
Because it's a completely different OS/kernel. If the user has a kernel that supports locked SGX, then so be it. There's no novel circumvention of the kernel policy, e.g. the user could simply boot the non-upstream kernel directly, and running an upstream kernel in the guest will not cause the kernel to support SGX.
There are any number of things that are allowed in a KVM guest that are not allowed in a bare metal process.
I buy this.
Quoted 6 lines
Quoted 3 lines
I cannot remember out of top of my head, could the Intel SHA256 be read when booted with unlocked MSRs. If that is the case, then you can still support guests with that configuration.
No, it's not guaranteed to be readable as firmware could have already changed the values in the MSRs.
Right.
Quoted 8 lines
Quoted 3 lines
Context-dependent guidelines tend to also trash code big time. Also, for the sake of a sane kernel code base, I would consider only supporting unlocked MSRs.
It's one line of a code to teach the kernel driver not to load if the MSRs are locked. And IMO, that one line of code is a net positive as it makes it clear in the driver itself that it chooses not support locked MSRs, even if SGX itself is fully enabled.
Yup, I think this clears my concerns, thank you.
/Jarkko
Re: [RFC PATCH 04/23] x86/cpufeatures: Add SGX1 and SGX2 sub-features
In reply to Sean ChristophersonQuoted 25 lines · On Mon, 11 Jan 2021 11:20:11 -0800 Sean Christopherson wrote:
Quoted 17 lines · Borislav Petkov
Quoted 5 lines · Sean Christopherson
It would be possible for KVM to break the dependency on X86_FEATURE_* bit offsets by defining a translation layer, but I strongly feel that adding manual translations will do more harm than good as it increases the odds of us botching a translation or using the wrong feature flag, creates potential namespace conflicts, etc...
Ok, lemme see if we might encounter more issues down the road...
arch/x86/kvm/cpuid.h +7
-0+enum kvm_only_cpuid_leafs { +CPUID_12_EAX =NCAPINTS , +NR_KVM_CPU_CAPS , + +NKVMCAPINTS =NR_KVM_CPU_CAPS -NCAPINTS , +}; +Quoted 2 lines
What happens when we decide to allocate a separate leaf for
CPUID_12_EAXdown the road?Well, mechanically, that would generate a build failure if the kernel does the obvious things and names the 'enum
cpuid_leafs' entryCPUID_12_EAX. That would be an obvious clue that KVM should be updated.
Quoted 5 lines
If the kernel named the enum entry something different, and we botched the code review, KVM would continue to work, but would unnecessarily copy the bits it
cares about to its own word. E.g. theboot_cpu_has()checks and translation to
__X86_FEATURE_* would still be valid. As far as failure modes go, that's not terrible.
Should we add a dedicated, i.e. kvm_scattered_cpu_caps[], instead of using existing kvm_cpu_cap[NCAPINTS]? If so this issue can be avoided??
Quoted 24 lines
Quoted 24 lines
You do it already here
Subject: [PATCH 04/13] x86/cpufeatures: Assign dedicated feature word for AMD mem encryption
for the AMD leaf.
I'm thinking this way around - from scattered to a hw one - should be ok because that should work easily. The other way around, taking a hw leaf and scattering it around x86_capability[] array elems would probably be nasty but with your change that should work too.
Yah, I'm just hypothesizing here - I don't think this "other way around" will ever happen...
Hmm, yap, I can cautiously say that with your change we should be ok...
Thx.
Signature · 4 lines
Regards/Gruss, Boris. https://people.kernel.org/tglx/notes-about-netiquette
Jarkko Sakkinen
maintainerQuoted 86 lines · Kai Huang
Quoted 52 lines · On Mon, 11 Jan 2021 19:20:48 +0200 Jarkko Sakkinen wrote:
Quoted 40 lines · Kai Huang
--- Disclaimer ---
These patches were originally written by Sean Christopherson while at Intel. Now that Sean has left Intel, I (Kai) have taken over getting them upstream. This series needs more review before it can be merged. It is being posted publicly and under RFC so Sean and others can review it. Maintainers are safe ignoring it for now.
------------------
Hi all,
This series adds KVM SGX virtualization support. The first 12 patches starting with x86/sgx or x86/cpu.. are necessary changes to x86 and SGX core/driver to support KVM SGX virtualization, while the rest are patches to KVM subsystem.
Please help to review this series. Also I'd like to hear what is the proper way to merge this series, since it contains change to both x86/SGX and KVM subsystem. Any feedback is highly appreciated. And please let me know if I forgot to CC anyone, or anyone wants to be removed from CC. Thanks in advance!
This series is based against latest tip tree's x86/sgx branch. You can also get the code from tip branch of kvm-sgx repo on github:
https://github.com/intel/kvm-sgx.git tip
It also requires Qemu changes to create VM with SGX support. You can find Qemu repo here:
https://github.com/intel/qemu-sgx.git next
Please refer to README.md of above qemu-sgx repo for detail on how to create guest with SGX support. At meantime, for your quick reference you can use below command to create SGX guest:
#qemu-system-x86_64 -smp 4 -m 2G -drive file=<your_vm_image>,if=virtio \ -cpu host,+sgx_provisionkey \ -sgx-epc id=epc1,memdev=mem1 \ -object memory-backend-epc,id=mem1,size=64M,preallocQuoted 7 lines
Please note that the SGX relevant part is:
-cpu host,+sgx_provisionkey \ -sgx-epc id=epc1,memdev=mem1 \ -object memory-backend-epc,id=mem1,size=64M,preallocAnd you can change other parameters of your qemu command based on your needs.
Thanks a lot documenting these snippets to the cover letter. I dig these up from lore once my environment is working.
Quoted 7 lines
I'm setting up Arch based test environment with the eye on this patch set and generic Linux keyring patches:
https://git.kernel.org/pub/scm/linux/kernel/git/jarkko/arch.git/
Still have some minor bits to adjust before I can start deploying it for SGX testing. For this patch set I'll use two instances of it.
Thanks. Please let me know if you need anything more.
Quoted 23 lines
Quoted 8 lines
=========
KVM SGX virtualization Overview- Virtual EPC
"Virtual EPC" is the EPC section exposed by KVM to guest so SGX software in guest can discover it and use it to create SGX enclaves. KVM exposes SGX to
Virtual EPC is a representation of an EPC section. And there is no "the".
Quoted 11 lines
guest via CPUID, and exposes one or more "virtual EPC" sections for guest. The size of "virtual EPC" is passed as Qemu parameter when creating the guest, and the base address is calcualted internally according to guest's configuration.
To support virtual EPC, add a new misc device
/dev/sgx_virt_epcto SGX core/driver to allow userspace (Qemu) to allocate "raw" EPC, and use it as "virtual EPC" for guest. Obviously, unlike EPC allocated for host SGX driver, virtual EPC allocated via/dev/sgx_virt_epcdoesn't have enclave associated, and how virtual EPC is used by guest is compeletely controlled by guest's SGX software.
Quoted 8 lines
Quoted 2 lines
I think that
/dev/sgx_vepcwould be a clear enough name for the device. This text has now a bit confusing "terminology" related to this.
/dev/sgx_virt_epcmay be clearer from userspace's perspective, for instance, if people see/dev/sgx_vepc, they may have to think about what it is, while/dev/sgx_virt_epcthey may not.But I don't have strong objection here. Does anyone has anything to say here?
It's already an abberevation to start with, why leave it halfways?
Especially when three remaining words have been shrunk to single characters ('E', 'P' and 'C').
/Jarkko
Re: [RFC PATCH 03/23] x86/sgx: Introduce virtual EPC for use by KVM guests
In reply to Jarkko SakkinenQuoted 48 lines · On Tue, 12 Jan 2021 03:50:23 +0200 Jarkko Sakkinen wrote:
Quoted 47 lines · Kai Huang
Quoted 30 lines · On Tue, 12 Jan 2021 01:38:23 +0200 Jarkko Sakkinen wrote:
Quoted 29 lines · Kai Huang
From: Sean Christopherson <[email protected]>
Add a misc device
/dev/sgx_virt_epcto allow userspace to allocate "raw" EPC without an associated enclave. The intended and only known use case for raw EPC allocation is to expose EPC to a KVM guest, hence thevirt_epcmoniker, virt.{c,h} files andX86_SGX_VIRTUALIZATIONKconfig.Modify
sgx_init()to always try to initialize virtual EPC driver, even when SGX driver is disabled due to SGX Launch Control is in locked mode, or not present at all, since SGX virtualization allows to expose SGX to guests that support non-LC configurations.Implement the "raw" EPC allocation in the x86 core-SGX subsystem via
/dev/sgx_virt_epcrather than in KVM. Doing so has two major advantages:- Does not require changes to KVM's uAPI, e.g. EPC gets handled as just another memory backend for guests.
- EPC management is wholly contained in the SGX subsystem, e.g. SGX does not have to export any symbols, changes to reclaim flows don't need to be routed through KVM, SGX's dirty laundry doesn't have to get aired out for the world to see, and so on and so forth.
The virtual EPC allocated to guests is currently not reclaimable, due to oversubscription of EPC for KVM guests is not currently supported. Due to the complications of handling reclaim conflicts between guest and host, KVM EPC oversubscription is significantly more complex than basic support for SGX virtualization.
Quoted 7 lines
Quoted 3 lines
Signed-off-by Sean Christopherson <[email protected]>Co-developed-by Kai Huang <[email protected]>Signed-off-by Kai Huang <[email protected]>The commit message does not describe the code changes. It should have an understandable explanation of fops. There is nothing about the implementation right now.
Thanks for feedback. Does "understabdable explanation of fops" mean I should add one sentence to say, for instance: "userspace hypervisor should open the
/dev/sgx_virt_epc, usemmap()to get a valid address range, and then use that address range to create KVM memory region"?Or should I include an example of how to use
/dev/sgx_virt_epcin userspace, for instance, below?
Quoted 7 lines
Quoted 4 lines
fd = open("/dev/sgx_virt_epc", O_RDWR); void *addr = mmap(NULL, size, ..., fd); /* userspace hypervisor uses addr, size to create KVM memory slot */ ...I would suggest just to describe them in few sentences. Just write how you understand them in one paragraph.
Will do. Thanks.
Quoted 16 lines
Quoted 16 lines
Quoted 11 lines
Quoted 8 lines
To support virtual EPC, add a new misc device
/dev/sgx_virt_epcto SGX core/driver to allow userspace (Qemu) to allocate "raw" EPC, and use it as "virtual EPC" for guest. Obviously, unlike EPC allocated for host SGX driver, virtual EPC allocated via/dev/sgx_virt_epcdoesn't have enclave associated, and how virtual EPC is used by guest is compeletely controlled by guest's SGX software.I think that
/dev/sgx_vepcwould be a clear enough name for the device. This text has now a bit confusing "terminology" related to this.
/dev/sgx_virt_epcmay be clearer from userspace's perspective, for instance, if people see/dev/sgx_vepc, they may have to think about what it is, while/dev/sgx_virt_epcthey may not.
Quoted 7 lines
Quoted 1 line
But I don't have strong objection here. Does anyone has anything to say here?
It's already an abberevation to start with, why leave it halfways?
Especially when three remaining words have been shrunk to single characters ('E', 'P' and 'C').
I have expressed my opinion above. And as I said I don't have strong objection here. I'll change to /dev/sgx_vepc if no one opposes.
Borislav Petkov
maintainerRe: [RFC PATCH 04/23] x86/cpufeatures: Add SGX1 and SGX2 sub-features
In reply to Sean ChristophersonQuoted 3 lines · Sean Christopherson
Well, mechanically, that would generate a build failure if the kernel does the obvious things and names the 'enum
cpuid_leafs' entryCPUID_12_EAX. That would be an obvious clue that KVM should be updated.
Then we need to properly document that whenever someone does that change, someone needs to touch the proper places.
Quoted 5 lines
If the kernel named the enum entry something different, and we botched the code review, KVM would continue to work, but would unnecessarily copy the bits it
cares about to its own word. E.g. theboot_cpu_has()checks and translation to
__X86_FEATURE_* would still be valid. As far as failure modes go, that's not terrible.
Right, which reminds me: with your prototype patch, we would have:
static __always_inline void __kvm_cpu_cap_mask(enum cpuid_leafs leaf)
{
const struct cpuid_reg cpuid = x86_feature_cpuid(leaf * 32);
struct kvm_cpuid_entry2 entry;
reverse_cpuid_check(leaf);
cpuid_count(cpuid.function, cpuid.index,
&entry.eax, &entry.ebx, &entry.ecx, &entry.edx);
kvm_cpu_caps[leaf] &= *__cpuid_entry_get_reg(&entry, cpuid.reg);
}which does read CPUID from the hw and kvm_cpu_caps[] has already the copied bits from boot_cpu_data.x86_capability.
Now you said that reading the CPUID is mostly redundant but we're paranoid so we do it anyway, just in case, so how about we remove the copying of boot_cpu_data.x86_capability? That's one less dependency on the baremetal implementation.
Practically, nothing changes for kvm because it will read CPUID which is the canonical thing anyway. And this should simplify the deal more and keep it simple(r).
Hmmm.
Signature · 4 lines
Regards/Gruss,
Boris.
https://people.kernel.org/tglx/notes-about-netiquetteSean Christopherson
maintainerRe: [RFC PATCH 04/23] x86/cpufeatures: Add SGX1 and SGX2 sub-features
Quoted 32 lines · Borislav Petkov
Quoted 3 lines · Sean Christopherson
Well, mechanically, that would generate a build failure if the kernel does the obvious things and names the 'enum
cpuid_leafs' entryCPUID_12_EAX. That would be an obvious clue that KVM should be updated.Then we need to properly document that whenever someone does that change, someone needs to touch the proper places.
Quoted 5 lines
If the kernel named the enum entry something different, and we botched the code review, KVM would continue to work, but would unnecessarily copy the bits it
cares about to its own word. E.g. theboot_cpu_has()checks and translation to
__X86_FEATURE_* would still be valid. As far as failure modes go, that's not terrible.Right, which reminds me: with your prototype patch, we would have:
static
__always_inlinevoid__kvm_cpu_cap_mask(enum cpuid_leafs leaf){ const struct cpuid_reg cpuid = x86_feature_cpuid(leaf * 32); struct kvm_cpuid_entry2 entry; reverse_cpuid_check(leaf); cpuid_count(cpuid.function, cpuid.index, &entry.eax, &entry.ebx, &entry.ecx, &entry.edx); kvm_cpu_caps[leaf] &= *__cpuid_entry_get_reg(&entry, cpuid.reg); }which does read CPUID from the hw and kvm_cpu_caps[] has already the copied bits from boot_cpu_data.x86_capability.
Quoted 8 lines
Now you said that reading the CPUID is mostly redundant but we're paranoid so we do it anyway, just in case, so how about we remove the copying of boot_cpu_data.x86_capability? That's one less dependency on the baremetal implementation.
Practically, nothing changes for kvm because it will read CPUID which is the canonical thing anyway. And this should simplify the deal more and keep it simple(r).
We want the boot_cpu_data.x86_capability memcpy() so that KVM doesn't advertise support for features that are intentionally disabled in the kernel, e.g. via kernel params. Except for a few special cases, e.g. LA57, KVM doesn't enable features in the guest if they're disabled in the host, even if the features are supported in hardware.
For some features, e.g. SMEP and SMAP, honoring boot_cpu_data is mostly about respecting the kernel's wishes, i.e. barring hardware bugs, enabling such features in the guest won't break anything. But for other features, e.g. XSAVE based features, enabling them in the guest without proper support in the host will corrupt guest and/or host state.
So it's really the CPUID read that is (mostly) superfluous.
Borislav Petkov
maintainerRe: [RFC PATCH 04/23] x86/cpufeatures: Add SGX1 and SGX2 sub-features
Quoted 11 lines · Sean Christopherson
We want the boot_cpu_data.x86_capability
memcpy()so that KVM doesn't advertise support for features that are intentionally disabled in the kernel, e.g. via kernel params. Except for a few special cases, e.g. LA57, KVM doesn't enable features in the guest if they're disabled in the host, even if the features are supported in hardware.For some features, e.g. SMEP and SMAP, honoring
boot_cpu_datais mostly about respecting the kernel's wishes, i.e. barring hardware bugs, enabling such features in the guest won't break anything. But for other features, e.g. XSAVE based features, enabling them in the guest without proper support in the host will corrupt guest and/or host state.
Ah ok, that is an important point.
Quoted 1 line
So it's really the CPUID read that is (mostly) superfluous.
Yeah, but that is cheap, as we established.
Ok then, I don't see anything that might be a problem and I guess we can try that handling of scattered bits in kvm and see how far we'll get.
Thx.
Signature · 4 lines
Regards/Gruss,
Boris.
https://people.kernel.org/tglx/notes-about-netiquetteRe: [RFC PATCH 04/23] x86/cpufeatures: Add SGX1 and SGX2 sub-features
Quoted 18 lines · Borislav Petkov
Quoted 11 lines · Sean Christopherson
We want the boot_cpu_data.x86_capability
memcpy()so that KVM doesn't advertise support for features that are intentionally disabled in the kernel, e.g. via kernel params. Except for a few special cases, e.g. LA57, KVM doesn't enable features in the guest if they're disabled in the host, even if the features are supported in hardware.For some features, e.g. SMEP and SMAP, honoring
boot_cpu_datais mostly about respecting the kernel's wishes, i.e. barring hardware bugs, enabling such features in the guest won't break anything. But for other features, e.g. XSAVE based features, enabling them in the guest without proper support in the host will corrupt guest and/or host state.Ah ok, that is an important point.
Quoted 6 lines
Quoted 1 line
So it's really the CPUID read that is (mostly) superfluous.
Yeah, but that is cheap, as we established.
Ok then, I don't see anything that might be a problem and I guess we can try that handling of scattered bits in kvm and see how far we'll get.
Hi Sean, Boris,
Thanks for all your feedback.
Sean,
Do you want to send me your patch (so that with your SoB), or do you want me to copy & paste the code you posted in this series, plus Suggested-by you? Or how do you want to proceed?
Also to me it is better to separate X86_FEATURE_SGX1/2 with rest of KVM changes?
And do you think adding a dedicated, i.e. kvm_scattered_cpu_caps[], instead of using existing kvm_cpu_cap[NCAPINTS] would be helpful to solve the problem caused by adding new leaf to x86 core (see my another reply in this thread)?
Quoted 3 lines
Thx.
Sean Christopherson
maintainerRe: [RFC PATCH 01/23] x86/sgx: Split out adding EPC page to free list to separate helper
In reply to Kai HuangQuoted 17 lines · Kai Huang
Quoted 16 lines · On Tue, 12 Jan 2021 00:38:40 +0200 Jarkko Sakkinen wrote:
Quoted 14 lines · Kai Huang
From: Sean Christopherson <[email protected]>
SGX virtualization requires to allocate "raw" EPC and use it as virtual EPC for SGX guest. Unlike EPC used by SGX driver, virtual EPC doesn't track how EPC pages are used in VM, e.g. (de)construction of enclaves, so it cannot guarantee EREMOVE success, e.g. it doesn't have a priori knowledge of which pages are SECS with non-zero child counts.
Split
sgx_free_page()into two parts so that the "add to free list" part can be used by virtual EPC without having to modify the EREMOVE logic insgx_free_page().Signed-off-by Sean Christopherson <[email protected]>Signed-off-by Kai Huang <[email protected]>
Quoted 8 lines
Quoted 3 lines
I have a better idea with the same outcome for KVM.
I agree with your patch this one can be replaced. I'll include your patch in next version, and once it is upstreamed, it can be removed in my series.
Sean, please let me know if you have objection.
6 of one, half dozen of the other. I liked not having to modify the existing call sites, but it's your code.
Though on that topic, this snippet is wrong:
@@ -431,7 +443,8 @@ void sgx_encl_release(struct kref *ref)
va_page = list_first_entry(&encl->va_pages, struct sgx_va_page,
list);
list_del(&va_page->list);
- sgx_free_epc_page(va_page->epc_page);
+ sgx_reset_epc_page(entry->epc_page);
+ sgx_free_epc_page(entry->epc_page);
s/entry/va_page in the new code.
P.S. I apparently hadn't been subscribed linux-sgx and so didn't see those
patches. I'm now subscribed and can chime-in as needed.
Sean Christopherson
maintainerRe: [RFC PATCH 04/23] x86/cpufeatures: Add SGX1 and SGX2 sub-features
In reply to Kai HuangQuoted 27 lines · Kai Huang
Quoted 13 lines · Borislav Petkov
Quoted 11 lines · Sean Christopherson
We want the boot_cpu_data.x86_capability
memcpy()so that KVM doesn't advertise support for features that are intentionally disabled in the kernel, e.g. via kernel params. Except for a few special cases, e.g. LA57, KVM doesn't enable features in the guest if they're disabled in the host, even if the features are supported in hardware.For some features, e.g. SMEP and SMAP, honoring
boot_cpu_datais mostly about respecting the kernel's wishes, i.e. barring hardware bugs, enabling such features in the guest won't break anything. But for other features, e.g. XSAVE based features, enabling them in the guest without proper support in the host will corrupt guest and/or host state.Quoted 8 lines
Ah ok, that is an important point.
Quoted 1 line
So it's really the CPUID read that is (mostly) superfluous.
Yeah, but that is cheap, as we established.
Ok then, I don't see anything that might be a problem and I guess we can try that handling of scattered bits in kvm and see how far we'll get.
Hi Sean, Boris,
Thanks for all your feedback.
Quoted 7 lines
Sean,
Do you want to send me your patch (so that with your SoB), or do you want me to copy & paste the code you posted in this series, plus Suggested-by you? Or how do you want to proceed?
Also to me it is better to separate X86_FEATURE_SGX1/2 with rest of KVM changes?
Hmm, I'll split the changes into two proper patches and send them to you off list.
Quoted 3 lines
And do you think adding a dedicated, i.e. kvm_scattered_cpu_caps[], instead of using existing kvm_cpu_cap[NCAPINTS] would be helpful to solve the problem caused by adding new leaf to x86 core (see my another reply in this thread)?
Probably not, because then we'd have to add new helpers to deal with the new array, or change all the helpers to take the array as a pointer. Blasting past NCAPINTS is a little evil, but it does slot in nicely to the existing code.
Re: [RFC PATCH 04/23] x86/cpufeatures: Add SGX1 and SGX2 sub-features
Quoted 30 lines · Sean Christopherson
Quoted 29 lines · Kai Huang
Quoted 13 lines · Borislav Petkov
Quoted 11 lines · Sean Christopherson
We want the boot_cpu_data.x86_capability
memcpy()so that KVM doesn't advertise support for features that are intentionally disabled in the kernel, e.g. via kernel params. Except for a few special cases, e.g. LA57, KVM doesn't enable features in the guest if they're disabled in the host, even if the features are supported in hardware.For some features, e.g. SMEP and SMAP, honoring
boot_cpu_datais mostly about respecting the kernel's wishes, i.e. barring hardware bugs, enabling such features in the guest won't break anything. But for other features, e.g. XSAVE based features, enabling them in the guest without proper support in the host will corrupt guest and/or host state.Quoted 8 lines
Ah ok, that is an important point.
Quoted 1 line
So it's really the CPUID read that is (mostly) superfluous.
Yeah, but that is cheap, as we established.
Ok then, I don't see anything that might be a problem and I guess we can try that handling of scattered bits in kvm and see how far we'll get.
Hi Sean, Boris,
Thanks for all your feedback.
Sean,
Quoted 7 lines
Quoted 5 lines
Do you want to send me your patch (so that with your SoB), or do you want me to copy & paste the code you posted in this series, plus Suggested-by you? Or how do you want to proceed?
Also to me it is better to separate X86_FEATURE_SGX1/2 with rest of KVM changes?
Hmm, I'll split the changes into two proper patches and send them to you off list.
Thanks.
Quoted 8 lines
Quoted 4 lines
And do you think adding a dedicated, i.e. kvm_scattered_cpu_caps[], instead of using existing kvm_cpu_cap[NCAPINTS] would be helpful to solve the problem caused by adding new leaf to x86 core (see my another reply in this thread)?
Probably not, because then we'd have to add new helpers to deal with the new array, or change all the helpers to take the array as a pointer. Blasting past NCAPINTS is a little evil, but it does slot in nicely to the existing code.
Sure. Thanks.
Re: [RFC PATCH 01/23] x86/sgx: Split out adding EPC page to free list to separate helper
In reply to Sean ChristophersonQuoted 42 lines · On Tue, 12 Jan 2021 13:45:24 -0800 Sean Christopherson wrote:
Quoted 17 lines · Kai Huang
Quoted 16 lines · On Tue, 12 Jan 2021 00:38:40 +0200 Jarkko Sakkinen wrote:
Quoted 14 lines · Kai Huang
From: Sean Christopherson <[email protected]>
SGX virtualization requires to allocate "raw" EPC and use it as virtual EPC for SGX guest. Unlike EPC used by SGX driver, virtual EPC doesn't track how EPC pages are used in VM, e.g. (de)construction of enclaves, so it cannot guarantee EREMOVE success, e.g. it doesn't have a priori knowledge of which pages are SECS with non-zero child counts.
Split
sgx_free_page()into two parts so that the "add to free list" part can be used by virtual EPC without having to modify the EREMOVE logic insgx_free_page().Signed-off-by Sean Christopherson <[email protected]>Signed-off-by Kai Huang <[email protected]>Quoted 8 lines
Quoted 3 lines
I have a better idea with the same outcome for KVM.
I agree with your patch this one can be replaced. I'll include your patch in next version, and once it is upstreamed, it can be removed in my series.
Sean, please let me know if you have objection.
6 of one, half dozen of the other. I liked not having to modify the existing call sites, but it's your code.
Though on that topic, this snippet is wrong:
(quoted diff) +2
-1@@ -431,7 +443,8 @@ void sgx_encl_release(struct kref *ref) va_page = list_first_entry(&encl->va_pages, struct sgx_va_page, list); list_del(&va_page->list); - sgx_free_epc_page(va_page->epc_page); + sgx_reset_epc_page(entry->epc_page); + sgx_free_epc_page(entry->epc_page);s/entry/va_page in the new code.
Quoted 2 lines
P.S. I apparently hadn't been subscribed linux-sgx and so didn't see those
patches. I'm now subscribed and can chime-in as needed.
Thanks. I also have replied to Jarkko's v2 patch, and I think you can see it now.
I think if Jarkko's patch is eventually merged to upstream, we can drop this patch. So please help to comment if Jarkko's patch is reasonable, since I don't have history with SGX driver and cannot immediately tell if it is reasonable.
Jarkko Sakkinen
maintainerRe: [RFC PATCH 01/23] x86/sgx: Split out adding EPC page to free list to separate helper
In reply to Sean ChristophersonQuoted 22 lines · Sean Christopherson
Quoted 21 lines · Kai Huang
Quoted 19 lines · On Tue, 12 Jan 2021 00:38:40 +0200 Jarkko Sakkinen wrote:
Quoted 8 lines · Kai Huang
From: Sean Christopherson <[email protected]>
SGX virtualization requires to allocate "raw" EPC and use it as virtual EPC for SGX guest. Unlike EPC used by SGX driver, virtual EPC doesn't track how EPC pages are used in VM, e.g. (de)construction of enclaves, so it cannot guarantee EREMOVE success, e.g. it doesn't have a priori knowledge of which pages are SECS with non-zero child counts.
Quoted 6 lines
Split
sgx_free_page()into two parts so that the "add to free list" part can be used by virtual EPC without having to modify the EREMOVE logic insgx_free_page().Signed-off-by Sean Christopherson <[email protected]>Signed-off-by Kai Huang <[email protected]>I have a better idea with the same outcome for KVM.
Quoted 7 lines
Quoted 4 lines
I agree with your patch this one can be replaced. I'll include your patch in next version, and once it is upstreamed, it can be removed in my series.
Sean, please let me know if you have objection.
6 of one, half dozen of the other. I liked not having to modify the existing call sites, but it's your code.
Only the call sites contained sgx_encl_release() require a call to sgx_reset_epc_page(). That's three in total.
Review void sgx_encl_release(struct kref *ref) · line 447
Quoted 9 lines
Though on that topic, this snippet is wrong:
(quoted diff) +2
-1@@ -431,7 +443,8 @@ void sgx_encl_release(struct kref *ref) va_page = list_first_entry(&encl->va_pages, struct sgx_va_page, list); list_del(&va_page->list); - sgx_free_epc_page(va_page->epc_page); + sgx_reset_epc_page(entry->epc_page); + sgx_free_epc_page(entry->epc_page);
Thanks for the remark.
I checked why I did not see this when running test_sgx. I had not specified local tree for BuildRoot (LINUX_OVERRIDE_SRCDIR) when building test image.
Quoted 4 lines
s/entry/va_page in the new code.
P.S. I apparently hadn't been subscribed linux-sgx and so didn't see those
patches. I'm now subscribed and can chime-in as needed.
OK, great, I can also CC directly to the next version.
/Jarkko
Re: [RFC PATCH 03/23] x86/sgx: Introduce virtual EPC for use by KVM guests
In reply to Dave HansenQuoted 9 lines · On Wed, 6 Jan 2021 21:02:54 -0800 Dave Hansen wrote:
Quoted 7 lines · Kai Huang
Quoted 6 lines
I understand why this made sense for regular enclaves, but I'm having a harder time here. If you
mmap(fd, MAP_SHARED),fork(), and then pass that mapping through to two different guests, you get to hold the pieces, just like if you did the same with normal memory.Why does the kernel need to enforce this policy?
Does Sean's reply in another email satisfy you?
Quoted 7 lines
I'm not totally convinced.
Please give it a go in the changelog for the next one and try to convince me that this is a good idea. Focus on what the downsides will be if the kernel does not enforce this policy. What will break, and why will it be bad? Why is the kernel in the best position to thwart the badness?
Hi Dave, Sean,
Sorry for late reply of this. I feel I should try again to get consensus here.
From virtual EPC's perspective, if we don't force this in kernel, then *theoretically*, userspace can use fork() to make multiple VMs map to the same physical EPC, which will potentially cause enclaves in all VMs to behave abnormally. So to me, from this perspective, it's better to enforce in kernel so that only first VM can use this virtual EPC instance, because EPC by architectural design cannot be shared.
But as Sean said, KVM doesn't support VM across multiple mm structs. And if I read code correctly, KVM doesn't support userspace to use fork() to create new VM. For instance, when creating VM, KVM grabs current->mm and keeps it in 'struct kvm' for bookkeeping, and kvm_vcpu_ioctl() and kvm_device_ioctl() will refuse to work if kvm->mm doesn't equal to current->mm. So in practice, I believe w/o enforcing this in kernel, we should also have no problem here.
Sean, please correct me if I am wrong.
Dave, if above stands, do you think it is reasonable to keep current->mm in epc->mm and enforce in sgx_virt_epc_mmap()?
Quoted 20 lines · Kai Huang
Quoted 20 lines
Quoted 19 lines
Quoted 11 lines
Quoted 10 lines
To support virtual EPC, add a new misc device
/dev/sgx_virt_epcto SGX core/driver to allow userspace (Qemu) to allocate "raw" EPC, and use it as "virtual EPC" for guest. Obviously, unlike EPC allocated for host SGX driver,
virtual EPC allocated via/dev/sgx_virt_epcdoesn't have enclave associated,
and how virtual EPC is used by guest is compeletely controlled by guest's SGX
software.Quoted 2 lines
I think that
/dev/sgx_vepcwould be a clear enough name for the device. This text has now a bit confusing "terminology" related to this.
/dev/sgx_virt_epcmay be clearer from userspace's perspective, for instance, if people see/dev/sgx_vepc, they may have to think about what it is, while/dev/sgx_virt_epcthey may not.But I don't have strong objection here. Does anyone has anything to say here?
Quoted 8 lines
Quoted 5 lines
It's already an abberevation to start with, why leave it halfways?
Especially when three remaining words have been shrunk to single characters ('E', 'P' and 'C').
I have expressed my opinion above. And as I said I don't have strong objection here. I'll change to
/dev/sgx_vepcif no one opposes.
Hi Jarkko,
I am reluctant to change to /dev/sgx_vepc now, because there are lots of 'sgx_virt_epc' in the code. For instance, 'struct sgx_virt_epc', and function names in sgx/virt.c are all sgx_virt_epc_xxx(), which has 'sgx_virt_epc' as prefix. I feel changing to /dev/sgx_vepc only is kinda incomplete, but I really don't want to change so many 'sgx_virt_epc' to 'sgx_vepc'.
(Plus I still think 'virt_epc' is more obvious than 'vepc' from userspace's perspective.)
Does it make sense?
Re: [RFC PATCH 03/23] x86/sgx: Introduce virtual EPC for use by KVM guests
In reply to Kai HuangQuoted 14 lines · Kai Huang
Quoted 1 line
From virtual EPC's perspective, if we don't force this in kernel, then
*theoretically*, userspace can use
fork()to make multiple VMs map to the same physical EPC, which will potentially cause enclaves in all VMs to behave abnormally. So to me, from this perspective, it's better to enforce in kernel so that only first VM can use this virtual EPC instance, because EPC by architectural design cannot be shared.But as Sean said, KVM doesn't support VM across multiple mm structs. And if I read code correctly, KVM doesn't support userspace to use
fork()to create new VM. For instance, when creating VM, KVM grabscurrent->mmand keeps it in 'struct kvm' for bookkeeping, andkvm_vcpu_ioctl()andkvm_device_ioctl()will refuse to work ifkvm->mmdoesn't equal tocurrent->mm. So in practice, I believe w/o enforcing this in kernel, we should also have no problem here.
Quoted 4 lines
Sean, please correct me if I am wrong.
Dave, if above stands, do you think it is reasonable to keep
current->mminepc->mmand enforce insgx_virt_epc_mmap()?
Everything you wrote above tells me the kernel should not be enforcing the behavior. You basically said that it's only a theoretical problem, and old if someone goes and does something with KVM that's nobody can do today.
You've 100% convinced me that having the kernel enforce this is *un*reasonable.
Re: [RFC PATCH 03/23] x86/sgx: Introduce virtual EPC for use by KVM guests
Quoted 20 lines · On Fri, 15 Jan 2021 07:39:44 -0800 Dave Hansen wrote:
Quoted 14 lines · Kai Huang
Quoted 1 line
From virtual EPC's perspective, if we don't force this in kernel, then
*theoretically*, userspace can use
fork()to make multiple VMs map to the same physical EPC, which will potentially cause enclaves in all VMs to behave abnormally. So to me, from this perspective, it's better to enforce in kernel so that only first VM can use this virtual EPC instance, because EPC by architectural design cannot be shared.But as Sean said, KVM doesn't support VM across multiple mm structs. And if I read code correctly, KVM doesn't support userspace to use
fork()to create new VM. For instance, when creating VM, KVM grabscurrent->mmand keeps it in 'struct kvm' for bookkeeping, andkvm_vcpu_ioctl()andkvm_device_ioctl()will refuse to work ifkvm->mmdoesn't equal tocurrent->mm. So in practice, I believe w/o enforcing this in kernel, we should also have no problem here.Quoted 4 lines
Sean, please correct me if I am wrong.
Dave, if above stands, do you think it is reasonable to keep
current->mminepc->mmand enforce insgx_virt_epc_mmap()?
Quoted 7 lines
Everything you wrote above tells me the kernel should not be enforcing the behavior. You basically said that it's only a theoretical problem, and old if someone goes and does something with KVM that's nobody can do today.
You've 100% convinced me that having the kernel enforce this is *un*reasonable.
Sean, I'll remove epc->mm, unless I see your further objection.
Thanks to you both.
Sean Christopherson
maintainerRe: [RFC PATCH 03/23] x86/sgx: Introduce virtual EPC for use by KVM guests
Quoted 26 lines · Kai Huang
Quoted 25 lines · On Fri, 15 Jan 2021 07:39:44 -0800 Dave Hansen wrote:
Quoted 14 lines · Kai Huang
Quoted 1 line
From virtual EPC's perspective, if we don't force this in kernel, then
*theoretically*, userspace can use
fork()to make multiple VMs map to the same physical EPC, which will potentially cause enclaves in all VMs to behave abnormally. So to me, from this perspective, it's better to enforce in kernel so that only first VM can use this virtual EPC instance, because EPC by architectural design cannot be shared.But as Sean said, KVM doesn't support VM across multiple mm structs. And if I read code correctly, KVM doesn't support userspace to use
fork()to create new VM. For instance, when creating VM, KVM grabscurrent->mmand keeps it in 'struct kvm' for bookkeeping, andkvm_vcpu_ioctl()andkvm_device_ioctl()will refuse to work ifkvm->mmdoesn't equal tocurrent->mm. So in practice, I believe w/o enforcing this in kernel, we should also have no problem here.Quoted 4 lines
Sean, please correct me if I am wrong.
Dave, if above stands, do you think it is reasonable to keep
current->mminepc->mmand enforce insgx_virt_epc_mmap()?Everything you wrote above tells me the kernel should not be enforcing the behavior. You basically said that it's only a theoretical problem, and old if someone goes and does something with KVM that's nobody can do today.
Quoted 4 lines
Quoted 2 lines
You've 100% convinced me that having the kernel enforce this is *un*reasonable.
Sean, I'll remove
epc->mm, unless I see your further objection.
It's probably ok. I guess worst case scenario, to avoid the mm tracking nightmare for oversubscription, you could prevent attaching KVM to a virtual EPC if there is already a mm associated with the EPC, or if there are already EPC pages "in" the virt EPC.
Re: [RFC PATCH 03/23] x86/sgx: Introduce virtual EPC for use by KVM guests
Quoted 30 lines · On Fri, 15 Jan 2021 13:45:21 -0800 Sean Christopherson wrote:
Quoted 29 lines · Kai Huang
Quoted 27 lines · On Fri, 15 Jan 2021 07:39:44 -0800 Dave Hansen wrote:
Quoted 14 lines · Kai Huang
Quoted 1 line
From virtual EPC's perspective, if we don't force this in kernel, then
*theoretically*, userspace can use
fork()to make multiple VMs map to the same physical EPC, which will potentially cause enclaves in all VMs to behave abnormally. So to me, from this perspective, it's better to enforce in kernel so that only first VM can use this virtual EPC instance, because EPC by architectural design cannot be shared.But as Sean said, KVM doesn't support VM across multiple mm structs. And if I read code correctly, KVM doesn't support userspace to use
fork()to create new VM. For instance, when creating VM, KVM grabscurrent->mmand keeps it in 'struct kvm' for bookkeeping, andkvm_vcpu_ioctl()andkvm_device_ioctl()will refuse to work ifkvm->mmdoesn't equal tocurrent->mm. So in practice, I believe w/o enforcing this in kernel, we should also have no problem here.Quoted 4 lines
Sean, please correct me if I am wrong.
Dave, if above stands, do you think it is reasonable to keep
current->mminepc->mmand enforce insgx_virt_epc_mmap()?Everything you wrote above tells me the kernel should not be enforcing the behavior. You basically said that it's only a theoretical problem, and old if someone goes and does something with KVM that's nobody can do today.
You've 100% convinced me that having the kernel enforce this is *un*reasonable.
Quoted 6 lines
Quoted 1 line
Sean, I'll remove
epc->mm, unless I see your further objection.It's probably ok. I guess worst case scenario, to avoid the mm tracking nightmare for oversubscription, you could prevent attaching KVM to a virtual EPC if there is already a mm associated with the EPC, or if there are already EPC pages "in" the virt EPC.
Since we are not 100% certain oversubscription will be upstreamed, I think it makes sense to address when we do it. For now, let us just drop it.
Makes sense? Thanks.
Jarkko Sakkinen
maintainerQuoted 38 lines · Kai Huang
Quoted 20 lines · Kai Huang
Quoted 20 lines
Quoted 19 lines
Quoted 11 lines
Quoted 10 lines
To support virtual EPC, add a new misc device
/dev/sgx_virt_epcto SGX core/driver to allow userspace (Qemu) to allocate "raw" EPC, and use it as "virtual EPC" for guest. Obviously, unlike EPC allocated for host SGX driver,
virtual EPC allocated via/dev/sgx_virt_epcdoesn't have enclave associated,
and how virtual EPC is used by guest is compeletely controlled by guest's SGX
software.Quoted 2 lines
I think that
/dev/sgx_vepcwould be a clear enough name for the device. This text has now a bit confusing "terminology" related to this.
/dev/sgx_virt_epcmay be clearer from userspace's perspective, for instance, if people see/dev/sgx_vepc, they may have to think about what it is, while/dev/sgx_virt_epcthey may not.But I don't have strong objection here. Does anyone has anything to say here?
Quoted 8 lines
Quoted 5 lines
It's already an abberevation to start with, why leave it halfways?
Especially when three remaining words have been shrunk to single characters ('E', 'P' and 'C').
I have expressed my opinion above. And as I said I don't have strong objection here. I'll change to
/dev/sgx_vepcif no one opposes.Hi Jarkko,
I am reluctant to change to
/dev/sgx_vepcnow, because there are lots of 'sgx_virt_epc' in the code. For instance, 'structsgx_virt_epc', and function names insgx/virt.care allsgx_virt_epc_xxx(), which has 'sgx_virt_epc' as prefix. I feel changing to/dev/sgx_vepconly is kinda incomplete, but I really don't want to change so many 'sgx_virt_epc' to 'sgx_vepc'.
Quoted 4 lines
(Plus I still think '
virt_epc' is more obvious than 'vepc' from userspace's perspective.)Does it make sense?
We can reconsider naming later on for sure, and maybe it's better to do so. It's probably too early to define the final name.
As far as naming goes, I'm actually wondering is this usable outside of KVM by any means? If not, then probably the best name for this device would be sgx_kvm_epc. Better to be always as explicit as possible.
/Jarkko
Jarkko Sakkinen
maintainerQuoted 42 lines · Jarkko Sakkinen
Quoted 41 lines · Kai Huang
Quoted 20 lines · Kai Huang
Quoted 20 lines
Quoted 19 lines
Quoted 11 lines
Quoted 10 lines
To support virtual EPC, add a new misc device
/dev/sgx_virt_epcto SGX core/driver to allow userspace (Qemu) to allocate "raw" EPC, and use it as "virtual EPC" for guest. Obviously, unlike EPC allocated for host SGX driver,
virtual EPC allocated via/dev/sgx_virt_epcdoesn't have enclave associated,
and how virtual EPC is used by guest is compeletely controlled by guest's SGX
software.Quoted 2 lines
I think that
/dev/sgx_vepcwould be a clear enough name for the device. This text has now a bit confusing "terminology" related to this.
/dev/sgx_virt_epcmay be clearer from userspace's perspective, for instance, if people see/dev/sgx_vepc, they may have to think about what it is, while/dev/sgx_virt_epcthey may not.But I don't have strong objection here. Does anyone has anything to say here?
Quoted 8 lines
Quoted 5 lines
It's already an abberevation to start with, why leave it halfways?
Especially when three remaining words have been shrunk to single characters ('E', 'P' and 'C').
I have expressed my opinion above. And as I said I don't have strong objection here. I'll change to
/dev/sgx_vepcif no one opposes.Hi Jarkko,
I am reluctant to change to
/dev/sgx_vepcnow, because there are lots of 'sgx_virt_epc' in the code. For instance, 'structsgx_virt_epc', and function names insgx/virt.care allsgx_virt_epc_xxx(), which has 'sgx_virt_epc' as prefix. I feel changing to/dev/sgx_vepconly is kinda incomplete, but I really don't want to change so many 'sgx_virt_epc' to 'sgx_vepc'.(Plus I still think '
virt_epc' is more obvious than 'vepc' from userspace's perspective.)
Quoted 8 lines
Quoted 1 line
Does it make sense?
We can reconsider naming later on for sure, and maybe it's better to do so. It's probably too early to define the final name.
As far as naming goes, I'm actually wondering is this usable outside of KVM by any means? If not, then probably the best name for this device would be
sgx_kvm_epc. Better to be always as explicit as possible.
You can easily do such renames with git filter-branch over a patch set:
https://git-scm.com/book/en/v2/Git-Tools-Rewriting-History
Having to rename something in too many places is not an argument. Considering it too early is.
/Jarkko