KVM Xen and pfncache changes for 6.9:
- Rip out the half-baked support for using gfn_to_pfn caches to manage pages
that are "mapped" into guests via physical addresses.
- Add support for using gfn_to_pfn caches with only a host virtual address,
i.e. to bypass the "gfn" stage of the cache. The primary use case is
overlay pages, where the guest may change the gfn used to reference the
overlay page, but the backing hva+pfn remains the same.
- Add an ioctl() to allow mapping Xen's shared_info page using an hva instead
of a gpa, so that userspace doesn't need to reconfigure and invalidate the
cache/mapping if the guest changes the gpa (but userspace keeps the resolved
hva the same).
- When possible, use a single host TSC value when computing the deadline for
Xen timers in order to improve the accuracy of the timer emulation.
- Inject pending upcall events when the vCPU software-enables its APIC to fix
a bug where an upcall can be lost (and to follow Xen's behavior).
- Fall back to the slow path instead of warning if "fast" IRQ delivery of Xen
events fails, e.g. if the guest has aliased xAPIC IDs.
- Extend gfn_to_pfn_cache's mutex to cover (de)activation (in addition to
refresh), and drop a now-redundant acquisition of xen_lock (that was
protecting the shared_info cache) to fix a deadlock due to recursively
acquiring xen_lock.
KVM x86 misc changes for 6.9:
- Explicitly initialize a variety of on-stack variables in the emulator that
triggered KMSAN false positives (though in fairness in KMSAN, it's comically
difficult to see that the uninitialized memory is never truly consumed).
- Fix the deubgregs ABI for 32-bit KVM, and clean up code related to reading
DR6 and DR7.
- Rework the "force immediate exit" code so that vendor code ultimately
decides how and when to force the exit. This allows VMX to further optimize
handling preemption timer exits, and allows SVM to avoid sending a duplicate
IPI (SVM also has a need to force an exit).
- Fix a long-standing bug where kvm_has_noapic_vcpu could be left elevated if
vCPU creation ultimately failed, and add WARN to guard against similar bugs.
- Provide a dedicated arch hook for checking if a different vCPU was in-kernel
(for directed yield), and simplify the logic for checking if the currently
loaded vCPU is in-kernel.
- Misc cleanups and fixes.
KVM common MMU changes for 6.9:
- Harden KVM against underflowing the active mmu_notifier invalidation
count, so that "bad" invalidations (usually due to bugs elsehwere in the
kernel) are detected earlier and are less likely to hang the kernel.
- Fix a benign bug in __kvm_mmu_topup_memory_cache() where the object size
and number of objects parameters to kvmalloc_array() were swapped.
KVM async page fault changes for 6.9:
- Always flush the async page fault workqueue when a work item is being
removed, especially during vCPU destruction, to ensure that there are no
workers running in KVM code when all references to KVM-the-module are gone,
i.e. to prevent a use-after-free if kvm.ko is unloaded.
- Grab a reference to the VM's mm_struct in the async #PF worker itself instead
of gifting the worker a reference, e.g. so that there's no need to remember
to *conditionally* clean up after the worker.
KVM/arm64 updates for 6.9
- Infrastructure for building KVM's trap configuration based on the
architectural features (or lack thereof) advertised in the VM's ID
registers
- Support for mapping vfio-pci BARs as Normal-NC (vaguely similar to
x86's WC) at stage-2, improving the performance of interacting with
assigned devices that can tolerate it
- Conversion of KVM's representation of LPIs to an xarray, utilized to
address serialization some of the serialization on the LPI injection
path
- Support for _architectural_ VHE-only systems, advertised through the
absence of FEAT_E2H0 in the CPU's ID register
- Miscellaneous cleanups, fixes, and spelling corrections to KVM and
selftests
KVM GUEST_MEMFD fixes for 6.8:
- Make KVM_MEM_GUEST_MEMFD mutually exclusive with KVM_MEM_READONLY to
avoid creating ABI that KVM can't sanely support.
- Update documentation for KVM_SW_PROTECTED_VM to make it abundantly
clear that such VMs are purely a development and testing vehicle, and
come with zero guarantees.
- Limit KVM_SW_PROTECTED_VM guests to the TDP MMU, as the long term plan
is to support confidential VMs with deterministic private memory (SNP
and TDX) only in the TDP MMU.
- Fix a bug in a GUEST_MEMFD negative test that resulted in false passes
when verifying that KVM_MEM_GUEST_MEMFD memslots can't be dirty logged.
The locking on the gfn_to_pfn_cache is... interesting. And awful.
There is a rwlock in ->lock which readers take to ensure protection
against concurrent changes. But __kvm_gpc_refresh() makes assumptions
that certain fields will not change even while it drops the write lock
and performs MM operations to revalidate the target PFN and kernel
mapping.
Commit 93984f19e7 ("KVM: Fully serialize gfn=>pfn cache refresh via
mutex") partly addressed that — not by fixing it, but by adding a new
mutex, ->refresh_lock. This prevented concurrent __kvm_gpc_refresh()
calls on a given gfn_to_pfn_cache, but is still only a partial solution.
There is still a theoretical race where __kvm_gpc_refresh() runs in
parallel with kvm_gpc_deactivate(). While __kvm_gpc_refresh() has
dropped the write lock, kvm_gpc_deactivate() clears the ->active flag
and unmaps ->khva. Then __kvm_gpc_refresh() determines that the previous
->pfn and ->khva are still valid, and reinstalls those values into the
structure. This leaves the gfn_to_pfn_cache with the ->valid bit set,
but ->active clear. And a ->khva which looks like a reasonable kernel
address but is actually unmapped.
All it takes is a subsequent reactivation to cause that ->khva to be
dereferenced. This would theoretically cause an oops which would look
something like this:
[1724749.564994] BUG: unable to handle page fault for address: ffffaa3540ace0e0
[1724749.565039] RIP: 0010:__kvm_xen_has_interrupt+0x8b/0xb0
I say "theoretically" because theoretically, that oops that was seen in
production cannot happen. The code which uses the gfn_to_pfn_cache is
supposed to have its *own* locking, to further paper over the fact that
the gfn_to_pfn_cache's own papering-over (->refresh_lock) of its own
rwlock abuse is not sufficient.
For the Xen vcpu_info that external lock is the vcpu->mutex, and for the
shared info it's kvm->arch.xen.xen_lock. Those locks ought to protect
the gfn_to_pfn_cache against concurrent deactivation vs. refresh in all
but the cases where the vcpu or kvm object is being *destroyed*, in
which case the subsequent reactivation should never happen.
Theoretically.
Nevertheless, this locking abuse is awful and should be fixed, even if
no clear explanation can be found for how the oops happened. So expand
the use of the ->refresh_lock mutex to ensure serialization of
activate/deactivate vs. refresh and make the pfncache locking entirely
self-sufficient.
This means that a future commit can simplify the locking in the callers,
such as the Xen emulation code which has an outstanding problem with
recursive locking of kvm->arch.xen.xen_lock, which will no longer be
necessary.
The rwlock abuse described above is still not best practice, although
it's harmless now that the ->refresh_lock is held for the entire duration
while the offending code drops the write lock, does some other stuff,
then takes the write lock again and assumes nothing changed. That can
also be fixed^W cleaned up in a subsequent commit, but this commit is
a simpler basis for the Xen deadlock fix mentioned above.
Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>
Reviewed-by: Paul Durrant <paul@xen.org>
Link: https://lore.kernel.org/r/20240227115648.3104-5-dwmw2@infradead.org
[sean: use guard(mutex) to fix a missed unlock]
Signed-off-by: Sean Christopherson <seanjc@google.com>
The general expectation with debugfs is that any initialization failure
is nonfatal. Nevertheless, kvm_arch_create_vm_debugfs() allows
implementations to return an error and kvm_create_vm_debugfs() allows
that to fail VM creation.
Change to a void return to discourage architectures from making debugfs
failures fatal for the VM. Seems like everyone already had the right
idea, as all implementations already return 0 unconditionally.
Acked-by: Marc Zyngier <maz@kernel.org>
Acked-by: Paolo Bonzini <pbonzini@redhat.com>
Link: https://lore.kernel.org/r/20240216155941.2029458-1-oliver.upton@linux.dev
Signed-off-by: Oliver Upton <oliver.upton@linux.dev>
Disallow creating read-only memslots that support GUEST_MEMFD, as
GUEST_MEMFD is fundamentally incompatible with KVM's semantics for
read-only memslots. Read-only memslots allow the userspace VMM to emulate
option ROMs by filling the backing memory with readable, executable code
and data, while triggering emulated MMIO on writes. GUEST_MEMFD doesn't
currently support writes from userspace and KVM doesn't support emulated
MMIO on private accesses, i.e. the guest can only ever read zeros, and
writes will always be treated as errors.
Cc: Fuad Tabba <tabba@google.com>
Cc: Michael Roth <michael.roth@amd.com>
Cc: Isaku Yamahata <isaku.yamahata@gmail.com>
Cc: Yu Zhang <yu.c.zhang@linux.intel.com>
Cc: Chao Peng <chao.p.peng@linux.intel.com>
Fixes: a7800aa80e ("KVM: Add KVM_CREATE_GUEST_MEMFD ioctl() for guest-specific backing memory")
Link: https://lore.kernel.org/r/20240222190612.2942589-2-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
gcc-14 notices that the arguments to kvmalloc_array() are mixed up:
arch/x86/kvm/../../../virt/kvm/kvm_main.c: In function '__kvm_mmu_topup_memory_cache':
arch/x86/kvm/../../../virt/kvm/kvm_main.c:424:53: error: 'kvmalloc_array' sizes specified with 'sizeof' in the earlier argument and not in the later argument [-Werror=calloc-transposed-args]
424 | mc->objects = kvmalloc_array(sizeof(void *), capacity, gfp);
| ^~~~
arch/x86/kvm/../../../virt/kvm/kvm_main.c:424:53: note: earlier argument should specify number of elements, later size of each element
The code still works correctly, but the incorrect order prevents the compiler
from properly tracking the object sizes.
Fixes: 837f66c712 ("KVM: Allow for different capacities in kvm_mmu_memory_cache structs")
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Reviewed-by: Marc Zyngier <maz@kernel.org>
Link: https://lore.kernel.org/r/20240212112419.1186065-1-arnd@kernel.org
Signed-off-by: Sean Christopherson <seanjc@google.com>
Plumb in a dedicated hook for querying whether or not a vCPU was preempted
in-kernel. Unlike literally every other architecture, x86's VMX can check
if a vCPU is in kernel context if and only if the vCPU is loaded on the
current pCPU.
x86's kvm_arch_vcpu_in_kernel() works around the limitation by querying
kvm_get_running_vcpu() and redirecting to vcpu->arch.preempted_in_kernel
as needed. But that's unnecessary, confusing, and fragile, e.g. x86 has
had at least one bug where KVM incorrectly used a stale
preempted_in_kernel.
No functional change intended.
Reviewed-by: Yuan Yao <yuan.yao@intel.com>
Link: https://lore.kernel.org/r/20240110003938.490206-2-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
When processing mmu_notifier invalidations for gpc caches, pre-check for
overlap with the invalidation event while holding gpc->lock for read, and
only take gpc->lock for write if the cache needs to be invalidated. Doing
a pre-check without taking gpc->lock for write avoids unnecessarily
contending the lock for unrelated invalidations, which is very beneficial
for caches that are heavily used (but rarely subjected to mmu_notifier
invalidations).
Signed-off-by: Paul Durrant <pdurrant@amazon.com>
Reviewed-by: David Woodhouse <dwmw@amazon.co.uk>
Link: https://lore.kernel.org/r/20240215152916.1158-20-paul@xen.org
Signed-off-by: Sean Christopherson <seanjc@google.com>
Some pfncache pages may actually be overlays on guest memory that have a
fixed HVA within the VMM. It's pointless to invalidate such cached
mappings if the overlay is moved so allow a cache to be activated directly
with the HVA to cater for such cases. A subsequent patch will make use
of this facility.
Signed-off-by: Paul Durrant <pdurrant@amazon.com>
Reviewed-by: David Woodhouse <dwmw@amazon.co.uk>
Link: https://lore.kernel.org/r/20240215152916.1158-10-paul@xen.org
Signed-off-by: Sean Christopherson <seanjc@google.com>
Currently the pfncache page offset is sometimes determined using the gpa
and sometimes the khva, whilst the uhva is always page-aligned. After a
subsequent patch is applied the gpa will not always be valid so adjust
the code to include the page offset in the uhva and use it consistently
as the source of truth.
Also, where a page-aligned address is required, use PAGE_ALIGN_DOWN()
for clarity.
No functional change intended.
Signed-off-by: Paul Durrant <pdurrant@amazon.com>
Reviewed-by: David Woodhouse <dwmw@amazon.co.uk>
Link: https://lore.kernel.org/r/20240215152916.1158-8-paul@xen.org
Signed-off-by: Sean Christopherson <seanjc@google.com>
As noted in [1] the KVM_GUEST_USES_PFN usage flag is never set by any
callers of kvm_gpc_init(), and for good reason: the implementation is
incomplete/broken. And it's not clear that there will ever be a user of
KVM_GUEST_USES_PFN, as coordinating vCPUs with mmu_notifier events is
non-trivial.
Remove KVM_GUEST_USES_PFN and all related code, e.g. dropping
KVM_GUEST_USES_PFN also makes the 'vcpu' argument redundant, to avoid
having to reason about broken code as __kvm_gpc_refresh() evolves.
Moreover, all existing callers specify KVM_HOST_USES_PFN so the usage
check in hva_to_pfn_retry() and hence the 'usage' argument to
kvm_gpc_init() are also redundant.
[1] https://lore.kernel.org/all/ZQiR8IpqOZrOpzHC@google.com
Signed-off-by: Paul Durrant <pdurrant@amazon.com>
Reviewed-by: David Woodhouse <dwmw@amazon.co.uk>
Link: https://lore.kernel.org/r/20240215152916.1158-6-paul@xen.org
[sean: explicitly call out that guest usage is incomplete]
Signed-off-by: Sean Christopherson <seanjc@google.com>
There is a pfncache unmap helper but mapping is open-coded. Arguably this
is fine because mapping is done in only one place, hva_to_pfn_retry(), but
adding the helper does make that function more readable.
No functional change intended.
Signed-off-by: Paul Durrant <pdurrant@amazon.com>
Reviewed-by: David Woodhouse <dwmw@amazon.co.uk>
Link: https://lore.kernel.org/r/20240215152916.1158-2-paul@xen.org
Signed-off-by: Sean Christopherson <seanjc@google.com>
Cleanups to Kconfig definitions for KVM
* replace HAVE_KVM with an architecture-dependent symbol, when CONFIG_KVM
may or may not be available depending on CPU capabilities (MIPS)
* replace HAVE_KVM with IS_ENABLED(CONFIG_KVM) for host-side code that is
not part of the KVM module, so that it is completely compiled out
* factor common "select" statements in common code instead of requiring
each architecture to specify it
CONFIG_IRQ_BYPASS_MANAGER is a dependency of the common code included by
CONFIG_HAVE_KVM_IRQ_BYPASS. There is no advantage in adding the corresponding
"select" directive to each architecture.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
KVM uses __KVM_HAVE_* symbols in the architecture-dependent uapi/asm/kvm.h to mask
unused definitions in include/uapi/linux/kvm.h. __KVM_HAVE_READONLY_MEM however
was nothing but a misguided attempt to define KVM_CAP_READONLY_MEM only on
architectures where KVM_CHECK_EXTENSION(KVM_CAP_READONLY_MEM) could possibly
return nonzero. This however does not make sense, and it prevented userspace
from supporting this architecture-independent feature without recompilation.
Therefore, these days __KVM_HAVE_READONLY_MEM does not mask anything and
is only used in virt/kvm/kvm_main.c. Userspace does not need to test it
and there should be no need for it to exist. Remove it and replace it
with a Kconfig symbol within Linux source code.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Nullify the async #PF worker's local "apf" pointer immediately after the
point where the structure can be freed by the vCPU. The existing comment
is helpful, but easy to overlook as there is no associated code.
Update the comment to clarify that it can be freed by as soon as the lock
is dropped, as "after this point" isn't strictly accurate, nor does it
help understand what prevents the structure from being freed earlier.
Reviewed-by: Xu Yilun <yilun.xu@intel.com>
Reviewed-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Link: https://lore.kernel.org/r/20240110011533.503302-5-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Get a reference to the target VM's address space in async_pf_execute()
instead of gifting a reference from kvm_setup_async_pf(). Keeping the
address space alive just to service an async #PF is counter-productive,
i.e. if the process is exiting and all vCPUs are dead, then NOT doing
get_user_pages_remote() and freeing the address space asap is desirable.
Handling the mm reference entirely within async_pf_execute() also
simplifies the async #PF flows as a whole, e.g. it's not immediately
obvious when the worker task vs. the vCPU task is responsible for putting
the gifted mm reference.
Reviewed-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Reviewed-by: Xu Yilun <yilun.xu@intel.com>
Link: https://lore.kernel.org/r/20240110011533.503302-4-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
