- 28 Oct, 2021 1 commit
-
-
Chang S. Bae authored
AMX TILEDATA is a very large XSAVE feature. It could have caused nasty XSAVE buffer space waste in two places: * Signal stacks * Kernel task_struct->fpu buffers To avoid this waste, neither of these buffers have AMX state by default. The non-default features are called "dynamic" features. There is an arch_prctl(ARCH_REQ_XCOMP_PERM) which allows a task to declare that it wants to use AMX or other "dynamic" XSAVE features. This arch_prctl() ensures that sufficient sigaltstack space is available before it will succeed. It also expands the task_struct buffer. Functions of this test: * Test arch_prctl(ARCH_REQ_XCOMP_PERM). Ensure that it checks for proper sigaltstack sizing and that the sizing is enforced for future sigaltstack calls. * Ensure that ARCH_REQ_XCOMP_PERM is inherited across fork() * Ensure that TILEDATA use before the prctl() is fatal * Ensure that TILEDATA is cleared across fork() Note: Generally, compiler support is needed to do something with AMX. Instead, directly load AMX state from userspace with a plain XSAVE. Do not depend on the compiler. [ dhansen: bunches of cleanups ] Signed-off-by:
Chang S. Bae <chang.seok.bae@intel.com> Signed-off-by:
Dave Hansen <dave.hansen@linux.intel.com> Signed-off-by:
Borislav Petkov <bp@suse.de> Link: https://lkml.kernel.org/r/20211026122524.7BEDAA95@davehans-spike.ostc.intel.com
-
- 26 Oct, 2021 23 commits
-
-
Chang S. Bae authored
Add the AMX state components in XFEATURE_MASK_USER_SUPPORTED and the TILE_DATA component to the dynamic states and update the permission check table accordingly. This is only effective on 64 bit kernels as for 32bit kernels XFEATURE_MASK_TILE is defined as 0. TILE_DATA is caller-saved state and the only dynamic state. Add build time sanity check to ensure the assumption that every dynamic feature is caller- saved. Make AMX state depend on XFD as it is dynamic feature. Signed-off-by:
Thomas Gleixner <tglx@linutronix.de> Signed-off-by:
-
Chang S. Bae authored
To handle the dynamic sizing of buffers on first use the XFD MSR has to be armed. Store the delta between the maximum available and the default feature bits in init_fpstate where it can be retrieved for task creation. If the delta is non zero then dynamic features are enabled. This needs also to enable the static key which guards the XFD updates. This is delayed to an initcall because the FPU setup runs before jump labels are initialized. Signed-off-by:
-
Chang S. Bae authored
When dynamically enabled states are supported the maximum and default sizes for the kernel buffers and user space interfaces are not longer identical. Put the necessary calculations in place which only take the default enabled features into account. Signed-off-by:
-
Chang S. Bae authored
The XSTATE initialization uses check_xstate_against_struct() to sanity check the size of XSTATE-enabled features. AMX is a XSAVE-enabled feature, and its size is not hard-coded but discoverable at run-time via CPUID. The AMX state is composed of state components 17 and 18, which are all user state components. The first component is the XTILECFG state of a 64-byte tile-related control register. The state component 18, called XTILEDATA, contains the actual tile data, and the state size varies on implementations. The architectural maximum, as defined in the CPUID(0x1d, 1): EAX[15:0], is a byte less than 64KB. The first implementation supports 8KB. Check the XTILEDATA state size dynamically. The feature introduces the new tile register, TMM. Define one register struct only and read the number of registers from CPUID. Cross-check the overall size with CPUID again. Signed-off-by:
-
Chang S. Bae authored
The kernel checks at boot time which features are available by walking a XSAVE feature table which contains the CPUID feature bit numbers which need to be checked whether a feature is available on a CPU or not. So far the feature numbers have been linear, but AMX will create a gap which the current code cannot handle. Make the table entries explicitly indexed and adjust the loop code accordingly to prepare for that. No functional change. Signed-off-by:
Len Brown <len.brown@intel.com> Signed-off-by:
-
Chang S. Bae authored
The fpstate embedded in struct fpu is the default state for storing the FPU registers. It's sized so that the default supported features can be stored. For dynamically enabled features the register buffer is too small. The #NM handler detects first use of a feature which is disabled in the XFD MSR. After handling permission checks it recalculates the size for kernel space and user space state and invokes fpstate_realloc() which tries to reallocate fpstate and install it. Provide the allocator function which checks whether the current buffer size is sufficient and if not allocates one. If allocation is successful the new fpstate is initialized with the new features and sizes and the now enabled features is removed from the task's XFD mask. realloc_fpstate() uses vzalloc(). If use of this mechanism grows to re-allocate buffers larger than 64KB, a more sophisticated allocation scheme that includes purpose-built reclaim capability might be justified. Signed-off-by:
-
Chang S. Bae authored
If the XFD MSR has feature bits set then #NM will be raised when user space attempts to use an instruction related to one of these features. When the task has no permissions to use that feature, raise SIGILL, which is the same behavior as #UD. If the task has permissions, calculate the new buffer size for the extended feature set and allocate a larger fpstate. In the unlikely case that vzalloc() fails, SIGSEGV is raised. The allocation function will be added in the next step. Provide a stub which fails for now. [ tglx: Updated serialization ] Signed-off-by:
-
Chang S. Bae authored
The IA32_XFD_MSR allows to arm #NM traps for XSTATE components which are enabled in XCR0. The register has to be restored before the tasks XSTATE is restored. The life time rules are the same as for FPU state. XFD is updated on return to userspace only when the FPU state of the task is not up to date in the registers. It's updated before the XRSTORS so that eventually enabled dynamic features are restored as well and not brought into init state. Also in signal handling for restoring FPU state from user space the correctness of the XFD state has to be ensured. Add it to CPU initialization and resume as well. Signed-off-by:
-
Thomas Gleixner authored
Add debug functionality to ensure that the XFD MSR is up to date for XSAVE* and XRSTOR* operations. [ tglx: Improve comment. ] Signed-off-by:
-
Chang S. Bae authored
Add storage for XFD register state to struct fpstate. This will be used to store the XFD MSR state. This will be used for switching the XFD MSR when FPU content is restored. Add a per-CPU variable to cache the current MSR value so the MSR has only to be written when the values are different. Signed-off-by:
-
Chang S. Bae authored
XFD introduces two MSRs: - IA32_XFD to enable/disable a feature controlled by XFD - IA32_XFD_ERR to expose to the #NM trap handler which feature was tried to be used for the first time. Both use the same xstate-component bitmap format, used by XCR0. Signed-off-by: -
Chang S. Bae authored
Intel's eXtended Feature Disable (XFD) feature is an extension of the XSAVE architecture. XFD allows the kernel to enable a feature state in XCR0 and to receive a #NM trap when a task uses instructions accessing that state. This is going to be used to postpone the allocation of a larger XSTATE buffer for a task to the point where it is actually using a related instruction after the permission to use that facility has been granted. XFD is not used by the kernel, but only applied to userspace. This is a matter of policy as the kernel knows how a fpstate is reallocated and the XFD state. The compacted XSAVE format is adjustable for dynamic features. Make XFD depend on XSAVES. Signed-off-by:
-
Chang S. Bae authored
On exec(), extended register states saved in the buffer is cleared. With dynamic features, each task carries variables besides the register states. The struct fpu has permission information and struct fpstate contains buffer size and feature masks. They are all dynamically updated with dynamic features. Reset the current task's entire FPU data before an exec() so that the new task starts with default permission and fpstate. Rename the register state reset function because the old naming confuses as it does not reset struct fpstate. Signed-off-by:
-
Thomas Gleixner authored
The default portion of the parent's FPU state is saved in a child task. With dynamic features enabled, the non-default portion is not saved in a child's fpstate because these register states are defined to be caller-saved. The new task's fpstate is therefore the default buffer. Fork inherits the permission of the parent. Also, do not use memcpy() when TIF_NEED_FPU_LOAD is set because it is invalid when the parent has dynamic features. Signed-off-by:
-
Chang S. Bae authored
The software reserved portion of the fxsave frame in the signal frame is copied from structures which have been set up at boot time. With dynamically enabled features the content of these structures is no longer correct because the xfeatures and size can be different per task. Calculate the software reserved portion at runtime and fill in the xfeatures and size values from the tasks active fpstate. Signed-off-by:
-
Thomas Gleixner authored
Use the current->group_leader->fpu to check for pending permissions to use extended features and validate against the resulting user space size which is stored in the group leaders fpu struct as well. This prevents a task from installing a too small sized sigaltstack after permissions to use dynamically enabled features have been granted, but the task has not (yet) used a related instruction. Signed-off-by:
-
Thomas Gleixner authored
To allow building up the infrastructure required to support dynamically enabled FPU features, add: - XFEATURES_MASK_DYNAMIC This constant will hold xfeatures which can be dynamically enabled. - fpu_state_size_dynamic() A static branch for 64-bit and a simple 'return false' for 32-bit. This helper allows to add dynamic-feature-specific changes to common code which is shared between 32-bit and 64-bit without #ifdeffery. Signed-off-by:
-
Chang S. Bae authored
Dynamically enabled XSTATE features are by default disabled for all processes. A process has to request permission to use such a feature. To support this implement a architecture specific prctl() with the options: - ARCH_GET_XCOMP_SUPP Copies the supported feature bitmap into the user space provided u64 storage. The pointer is handed in via arg2 - ARCH_GET_XCOMP_PERM Copies the process wide permitted feature bitmap into the user space provided u64 storage. The pointer is handed in via arg2 - ARCH_REQ_XCOMP_PERM Request permission for a feature set. A feature set can be mapped to a facility, e.g. AMX, and can require one or more XSTATE components to be enabled. The feature argument is the number of the highest XSTATE component which is required for a facility to work. The request argument is not a user supplied bitmap because that makes filtering harder (think seccomp) and even impossible because to support 32bit tasks the argument would have to be a pointer. The permission mechanism works this way: Task asks for permission for a facility and kernel checks whether that's supported. If supported it does: 1) Check whether permission has already been granted 2) Compute the size of the required kernel and user space buffer (sigframe) size. 3) Validate that no task has a sigaltstack installed which is smaller than the resulting sigframe size 4) Add the requested feature bit(s) to the permission bitmap of current->group_leader->fpu and store the sizes in the group leaders fpu struct as well. If that is successful then the feature is still not enabled for any of the tasks. The first usage of a related instruction will result in a #NM trap. The trap handler validates the permission bit of the tasks group leader and if permitted it installs a larger kernel buffer and transfers the permission and size info to the new fpstate container which makes all the FPU functions which require per task information aware of the extended feature set. [ tglx: Adopted to new base code, added missing serialization, massaged namings, comments and changelog ] Signed-off-by: -
Thomas Gleixner authored
The upcoming prctl() which is required to request the permission for a dynamically enabled feature will also provide an option to retrieve the supported features. If the CPU does not support XSAVE, the supported features would be 0 even when the CPU supports FP and SSE. Provide separate storage for the legacy feature set to avoid that and fill in the bits in the legacy init function. Signed-off-by:
-
Thomas Gleixner authored
Dynamically enabled features can be requested by any thread of a running process at any time. The request does neither enable the feature nor allocate larger buffers. It just stores the permission to use the feature by adding the features to the permission bitmap and by calculating the required sizes for kernel and user space. The reallocation of the kernel buffer happens when the feature is used for the first time which is caught by an exception. The permission bitmap is then checked and if the feature is permitted, then it becomes fully enabled. If not, the task dies similarly to a task which uses an undefined instruction. The size information is precomputed to allow proper sigaltstack size checks once the feature is permitted, but not yet in use because otherwise this would open race windows where too small stacks could be installed causing a later fail on signal delivery. Initialize them to the default feature set and sizes. Signed-off-by:
-
Chang S. Bae authored
Split out the size calculation from the paranoia check so it can be used for recalculating buffer sizes when dynamically enabled features are supported. Signed-off-by:
-
Thomas Gleixner authored
For historical reasons MINSIGSTKSZ is a constant which became already too small with AVX512 support. Add a mechanism to enforce strict checking of the sigaltstack size against the real size of the FPU frame. The strict check can be enabled via a config option and can also be controlled via the kernel command line option 'strict_sas_size' independent of the config switch. Enabling it might break existing applications which allocate a too small sigaltstack but 'work' because they never get a signal delivered. Though it can be handy to filter out binaries which are not yet aware of AT_MINSIGSTKSZ. Also the upcoming support for dynamically enabled FPU features requires a strict sanity check to ensure that: - Enabling of a dynamic feature, which changes the sigframe size fits into an enabled sigaltstack - Installing a too small sigaltstack after a dynamic feature has been added is not possible. Implement the base check which is controlled by config and command line options. Signed-off-by: -
Thomas Gleixner authored
New x86 FPU features will be very large, requiring ~10k of stack in signal handlers. These new features require a new approach called "dynamic features". The kernel currently tries to ensure that altstacks are reasonably sized. Right now, on x86, sys_sigaltstack() requires a size of >=2k. However, that 2k is a constant. Simply raising that 2k requirement to >10k for the new features would break existing apps which have a compiled-in size of 2k. Instead of universally enforcing a larger stack, prohibit a process from using dynamic features without properly-sized altstacks. This must be enforced in two places: * A dynamic feature can not be enabled without an large-enough altstack for each process thread. * Once a dynamic feature is enabled, any request to install a too-small altstack will be rejected The dynamic feature enabling code must examine each thread in a process to ensure that the altstacks are large enough. Add a new lock (sigaltstack_lock()) to ensure that threads can not race and change their altstack after being examined. Add the infrastructure in form of a config option and provide empty stubs for architectures which do not need dynamic altstack size checks. This implementation will be fleshed out for x86 in a future patch called x86/arch_prctl: Add controls for dynamic XSTATE components [dhansen: commit message. ] Signed-off-by:
-
- 23 Oct, 2021 4 commits
-
-
Thomas Gleixner authored
No more users. Signed-off-by:
-
Thomas Gleixner authored
For the upcoming AMX support it's necessary to do a proper integration with KVM. Currently KVM allocates two FPU structs which are used for saving the user state of the vCPU thread and restoring the guest state when entering vcpu_run() and doing the reverse operation before leaving vcpu_run(). With the new fpstate mechanism this can be reduced to one extra buffer by swapping the fpstate pointer in current::thread::fpu. This makes the upcoming support for AMX and XFD simpler because then fpstate information (features, sizes, xfd) are always consistent and it does not require any nasty workarounds. Convert the KVM FPU code over to this new scheme. Signed-off-by:
-
Thomas Gleixner authored
For the upcoming AMX support it's necessary to do a proper integration with KVM. Currently KVM allocates two FPU structs which are used for saving the user state of the vCPU thread and restoring the guest state when entering vcpu_run() and doing the reverse operation before leaving vcpu_run(). With the new fpstate mechanism this can be reduced to one extra buffer by swapping the fpstate pointer in current::thread::fpu. This makes the upcoming support for AMX and XFD simpler because then fpstate information (features, sizes, xfd) are always consistent and it does not require any nasty workarounds. Provide: - An allocator which initializes the state properly - A replacement for the existing FPU swap mechanim Aside of the reduced memory footprint, this also makes state switching more efficient when TIF_FPU_NEED_LOAD is set. It does not require a memcpy as the state is already correct in the to be swapped out fpstate. The existing interfaces will be removed once KVM is converted over. Signed-off-by:
-
Thomas Gleixner authored
For the upcoming AMX support it's necessary to do a proper integration with KVM. To avoid more nasty hackery in KVM which violate encapsulation extend struct fpu and fpstate so the fpstate switching can be consolidated and simplified. Currently KVM allocates two FPU structs which are used for saving the user state of the vCPU thread and restoring the guest state when entering vcpu_run() and doing the reverse operation before leaving vcpu_run(). With the new fpstate mechanism this can be reduced to one extra buffer by swapping the fpstate pointer in current::thread::fpu. This makes the upcoming support for AMX and XFD simpler because then fpstate information (features, sizes, xfd) are always consistent and it does not require any nasty workarounds. Add fpu::__task_fpstate to save the regular fpstate pointer while the task is inside vcpu_run(). Add some state fields to fpstate to indicate the nature of the state. Signed-off-by:
-
- 22 Oct, 2021 3 commits
-
-
Thomas Gleixner authored
Now that everything is mopped up, move all the helpers and prototypes into the core header. They are not required by the outside. Signed-off-by:
-
Thomas Gleixner authored
xfeatures_mask_fpstate() is no longer valid when dynamically enabled features come into play. Rework restore_regs_from_fpstate() so it takes a constant mask which will then be applied against the maximum feature set so that the restore operation brings all features which are not in the xsave buffer xfeature bitmap into init state. This ensures that if the previous task used a dynamically enabled feature that the task which restores has all unused components properly initialized. Cleanup the last user of xfeatures_mask_fpstate() as well and remove it. Signed-off-by:
-
Thomas Gleixner authored
Use the new fpu_user_cfg to retrieve the information instead of xfeatures_mask_uabi() which will be no longer correct when dynamically enabled features become available. Using fpu_user_cfg is appropriate when setting XCOMP_BV in the init_fpstate since it has space allocated for "max_features". But, normal fpstates might only have space for default xfeatures. Since XRSTOR* derives the format of the XSAVE buffer from XCOMP_BV, this can lead to XRSTOR reading out of bounds. So when copying actively used fpstate, simply read the XCOMP_BV features bits directly out of the fpstate instead. This correction courtesy of Dave Hansen <dave.hansen@linux.intel.com> Signed-off-by:
-
- 21 Oct, 2021 9 commits
-
-
Thomas Gleixner authored
Move the feature mask storage to the kernel and user config structs. Default and maximum feature set are the same for now. Signed-off-by:
-
Thomas Gleixner authored
Use the new kernel and user space config storage to store and retrieve the XSTATE buffer sizes. The default and the maximum size are the same for now, but will change when support for dynamically enabled features is added. Signed-off-by:
-
Thomas Gleixner authored
The size calculations are partially unreadable gunk. Clean them up. No functional change. Signed-off-by:
-
Thomas Gleixner authored
Clean the function up before making changes. No functional change. Signed-off-by:
-
Thomas Gleixner authored
Provide a struct to store information about the maximum supported and the default feature set and buffer sizes for both user and kernel space. This allows quick retrieval of this information for the upcoming support for dynamically enabled features. [ bp: Add vertical spacing between the struct members. ] Signed-off-by:
-
Thomas Gleixner authored
For dynamically enabled features it's required to get the features which are enabled for that context when restoring from sigframe. The same applies for all signal frame size calculations. Signed-off-by:
-
Thomas Gleixner authored
Prepare for dynamically enabled states per task. The function needs to retrieve the features and sizes which are valid in a fpstate context. Retrieve them from fpstate. Move the function declarations to the core header as they are not required anywhere else. Signed-off-by:
-
Thomas Gleixner authored
With dynamically enabled features the copy function must know the features and the size which is valid for the task. Retrieve them from fpstate. Signed-off-by:
-
Thomas Gleixner authored
Straight forward conversion. No functional change. Signed-off-by:
-
