Rename kvm_tdp_mmu_invalidate_all_roots() to
kvm_tdp_mmu_invalidate_roots(), and make it enum kvm_tdp_mmu_root_types
as an argument.
kvm_tdp_mmu_invalidate_roots() is called with different root types. For
kvm_mmu_zap_all_fast() it only operates on shared roots. But when tearing
down a VM it needs to invalidate all roots. Have the callers only
invalidate the required roots instead of all roots.
Within kvm_tdp_mmu_invalidate_roots(), respect the root type
passed by checking the root type in root iterator.
Suggested-by: Chao Gao <chao.gao@intel.com>
Signed-off-by: Isaku Yamahata <isaku.yamahata@intel.com>
Signed-off-by: Rick Edgecombe <rick.p.edgecombe@intel.com>
Message-ID: <20240718211230.1492011-17-rick.p.edgecombe@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Add the ability for the TDP MMU to maintain a mirror of a separate
mapping.
Like other Coco technologies, TDX has the concept of private and shared
memory. For TDX the private and shared mappings are managed on separate
EPT roots. The private half is managed indirectly through calls into a
protected runtime environment called the TDX module, where the shared half
is managed within KVM in normal page tables.
In order to handle both shared and private memory, KVM needs to learn to
handle faults and other operations on the correct root for the operation.
KVM could learn the concept of private roots, and operate on them by
calling out to operations that call into the TDX module. But there are two
problems with that:
1. Calls into the TDX module are relatively slow compared to the simple
accesses required to read a PTE managed directly by KVM.
2. Other Coco technologies deal with private memory completely differently
and it will make the code confusing when being read from their
perspective. Special operations added for TDX that set private or zap
private memory will have nothing to do with these other private memory
technologies. (SEV, etc).
To handle these, instead teach the TDP MMU about a new concept "mirror
roots". Such roots maintain page tables that are not actually mapped,
and are just used to traverse quickly to determine if the mid level page
tables need to be installed. When the memory be mirrored needs to actually
be changed, calls can be made to via x86_ops.
private KVM page fault |
| |
V |
private GPA | CPU protected EPTP
| | |
V | V
mirror PT root | external PT root
| | |
V | V
mirror PT --hook to propagate-->external PT
| | |
\--------------------+------\ |
| | |
| V V
| private guest page
|
|
non-encrypted memory | encrypted memory
|
Leave calling out to actually update the private page tables that are being
mirrored for later changes. Just implement the handling of MMU operations
on to mirrored roots.
In order to direct operations to correct root, add root types
KVM_DIRECT_ROOTS and KVM_MIRROR_ROOTS. Tie the usage of mirrored/direct
roots to private/shared with conditionals. It could also be implemented by
making the kvm_tdp_mmu_root_types and kvm_gfn_range_filter enum bits line
up such that conversion could be a direct assignment with a case. Don't do
this because the mapping of private to mirrored is confusing enough. So it
is worth not hiding the logic in type casting.
Cleanup the mirror root in kvm_mmu_destroy() instead of the normal place
in kvm_mmu_free_roots(), because the private root that is being cannot be
rebuilt like a normal root. It needs to persist for the lifetime of the VM.
The TDX module will also need to be provided with page tables to use for
the actual mapping being mirrored by the mirrored page tables. Allocate
these in the mapping path using the recently added
kvm_mmu_alloc_external_spt().
Don't support 2M page for now. This is avoided by forcing 4k pages in the
fault. Add a KVM_BUG_ON() to verify.
Signed-off-by: Isaku Yamahata <isaku.yamahata@intel.com>
Co-developed-by: Kai Huang <kai.huang@intel.com>
Signed-off-by: Kai Huang <kai.huang@intel.com>
Co-developed-by: Yan Zhao <yan.y.zhao@intel.com>
Signed-off-by: Yan Zhao <yan.y.zhao@intel.com>
Co-developed-by: Rick Edgecombe <rick.p.edgecombe@intel.com>
Signed-off-by: Rick Edgecombe <rick.p.edgecombe@intel.com>
Message-ID: <20240718211230.1492011-13-rick.p.edgecombe@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Define an enum kvm_tdp_mmu_root_types to specify the KVM MMU root type [1]
so that the iterator on the root page table can consistently filter the
root page table type instead of only_valid.
TDX KVM will operate on KVM page tables with specified types. Shared page
table, private page table, or both. Introduce an enum instead of bool
only_valid so that we can easily enhance page table types applicable to
shared, private, or both in addition to valid or not. Replace
only_valid=false with KVM_ANY_ROOTS and only_valid=true with
KVM_ANY_VALID_ROOTS. Use KVM_ANY_ROOTS and KVM_ANY_VALID_ROOTS to wrap
KVM_VALID_ROOTS to avoid further code churn when direct vs mirror root
concepts are introduced in future patches.
Link: https://lore.kernel.org/kvm/ZivazWQw1oCU8VBC@google.com/ [1]
Suggested-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Isaku Yamahata <isaku.yamahata@intel.com>
Signed-off-by: Rick Edgecombe <rick.p.edgecombe@intel.com>
Message-ID: <20240718211230.1492011-11-rick.p.edgecombe@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The kvm_tdp_mmu_alloc_root() function currently always returns 0. This
allows for the caller, mmu_alloc_direct_roots(), to call
kvm_tdp_mmu_alloc_root() and also return 0 in one line:
return kvm_tdp_mmu_alloc_root(vcpu);
So it is useful even though the return value of kvm_tdp_mmu_alloc_root()
is always the same. However, in future changes, kvm_tdp_mmu_alloc_root()
will be called twice in mmu_alloc_direct_roots(). This will force the
first call to either awkwardly handle the return value that will always
be zero or ignore it. So change kvm_tdp_mmu_alloc_root() to return void.
Do it in a separate change so the future change will be cleaner.
Signed-off-by: Rick Edgecombe <rick.p.edgecombe@intel.com>
Reviewed-by: Paolo Bonzini <pbonzini@redhat.com>
Message-ID: <20240718211230.1492011-7-rick.p.edgecombe@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Prepare for a future TDX patch which asserts that atomic zapping
(i.e. zapping with mmu_lock taken for read) don't operate on mirror roots.
When tearing down a VM, all roots have to be zapped (including mirror
roots once they're in place) so do that with the mmu_lock taken for write.
kvm_mmu_uninit_tdp_mmu() is invoked either before or after executing any
atomic operations on SPTEs by vCPU threads. Therefore, it will not impact
vCPU threads performance if kvm_tdp_mmu_zap_invalidated_roots() acquires
mmu_lock for write to zap invalid roots.
Co-developed-by: Yan Zhao <yan.y.zhao@intel.com>
Signed-off-by: Yan Zhao <yan.y.zhao@intel.com>
Signed-off-by: Rick Edgecombe <rick.p.edgecombe@intel.com>
Message-ID: <20240718211230.1492011-2-rick.p.edgecombe@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Recover TDP MMU huge page mappings in-place instead of zapping them when
dirty logging is disabled, and rename functions that recover huge page
mappings when dirty logging is disabled to move away from the "zap
collapsible spte" terminology.
Before KVM flushes TLBs, guest accesses may be translated through either
the (stale) small SPTE or the (new) huge SPTE. This is already possible
when KVM is doing eager page splitting (where TLB flushes are also
batched), and when vCPUs are faulting in huge mappings (where TLBs are
flushed after the new huge SPTE is installed).
Recovering huge pages reduces the number of page faults when dirty
logging is disabled:
$ perf stat -e kvm:kvm_page_fault -- ./dirty_log_perf_test -s anonymous_hugetlb_2mb -v 64 -e -b 4g
Before: 393,599 kvm:kvm_page_fault
After: 262,575 kvm:kvm_page_fault
vCPU throughput and the latency of disabling dirty-logging are about
equal compared to zapping, but avoiding faults can be beneficial to
remove vCPU jitter in extreme scenarios.
Signed-off-by: David Matlack <dmatlack@google.com>
Link: https://lore.kernel.org/r/20240823235648.3236880-5-dmatlack@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Drop the return value from kvm_tdp_mmu_clear_dirty_slot() as its sole
caller ignores the result (KVM flushes after clearing dirty logs based on
the logs themselves, not based on SPTEs).
Cc: David Matlack <dmatlack@google.com>
Link: https://lore.kernel.org/r/20241011021051.1557902-7-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Pass fault->gfn into kvm_tdp_mmu_fast_pf_get_last_sptep(), instead of
passing fault->addr and then converting it to a GFN.
Future changes will make fault->addr and fault->gfn differ when running
TDX guests. The GFN will be conceptually the same as it is for normal VMs,
but fault->addr may contain a TDX specific bit that differentiates between
"shared" and "private" memory. This bit will be used to direct faults to
be handled on different roots, either the normal "direct" root or a new
type of root that handles private memory. The TDP iterators will process
the traditional GFN concept and apply the required TDX specifics depending
on the root type. For this reason, it needs to operate on regular GFN and
not the addr, which may contain these special TDX specific bits.
Today kvm_tdp_mmu_fast_pf_get_last_sptep() takes fault->addr and then
immediately converts it to a GFN with a bit shift. However, this would
unfortunately retain the TDX specific bits in what is supposed to be a
traditional GFN. Excluding TDX's needs, it is also is unnecessary to pass
fault->addr and convert it to a GFN when the GFN is already on hand.
So instead just pass the GFN into kvm_tdp_mmu_fast_pf_get_last_sptep() and
use it directly.
Signed-off-by: Rick Edgecombe <rick.p.edgecombe@intel.com>
Message-ID: <20240619223614.290657-9-rick.p.edgecombe@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The .change_pte() MMU notifier callback was intended as an
optimization. The original point of it was that KSM could tell KVM to flip
its secondary PTE to a new location without having to first zap it. At
the time there was also an .invalidate_page() callback; both of them were
*not* bracketed by calls to mmu_notifier_invalidate_range_{start,end}(),
and .invalidate_page() also doubled as a fallback implementation of
.change_pte().
Later on, however, both callbacks were changed to occur within an
invalidate_range_start/end() block.
In the case of .change_pte(), commit 6bdb913f0a ("mm: wrap calls to
set_pte_at_notify with invalidate_range_start and invalidate_range_end",
2012-10-09) did so to remove the fallback from .invalidate_page() to
.change_pte() and allow sleepable .invalidate_page() hooks.
This however made KVM's usage of the .change_pte() callback completely
moot, because KVM unmaps the sPTEs during .invalidate_range_start()
and therefore .change_pte() has no hope of finding a sPTE to change.
Drop the generic KVM code that dispatches to kvm_set_spte_gfn(), as
well as all the architecture specific implementations.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Acked-by: Anup Patel <anup@brainfault.org>
Acked-by: Michael Ellerman <mpe@ellerman.id.au> (powerpc)
Reviewed-by: Bibo Mao <maobibo@loongson.cn>
Message-ID: <20240405115815.3226315-2-pbonzini@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
When allocating a new TDP MMU root, check for a usable root while holding
mmu_lock for read and only acquire mmu_lock for write if a new root needs
to be created. There is no need to serialize other MMU operations if a
vCPU is simply grabbing a reference to an existing root, holding mmu_lock
for write is "necessary" (spoiler alert, it's not strictly necessary) only
to ensure KVM doesn't end up with duplicate roots.
Allowing vCPUs to get "new" roots in parallel is beneficial to VM boot and
to setups that frequently delete memslots, i.e. which force all vCPUs to
reload all roots.
Link: https://lore.kernel.org/r/20240111020048.844847-7-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Neither tdp_mmu_next_root nor kvm_tdp_mmu_put_root need to know
if the lock is taken for read or write. Either way, protection
is achieved via RCU and tdp_mmu_pages_lock. Remove the argument
and just assert that the lock is taken.
Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Link: https://lore.kernel.org/r/20231125083400.1399197-2-pbonzini@redhat.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Stop zapping invalidate TDP MMU roots via work queue now that KVM
preserves TDP MMU roots until they are explicitly invalidated. Zapping
roots asynchronously was effectively a workaround to avoid stalling a vCPU
for an extended during if a vCPU unloaded a root, which at the time
happened whenever the guest toggled CR0.WP (a frequent operation for some
guest kernels).
While a clever hack, zapping roots via an unbound worker had subtle,
unintended consequences on host scheduling, especially when zapping
multiple roots, e.g. as part of a memslot. Because the work of zapping a
root is no longer bound to the task that initiated the zap, things like
the CPU affinity and priority of the original task get lost. Losing the
affinity and priority can be especially problematic if unbound workqueues
aren't affined to a small number of CPUs, as zapping multiple roots can
cause KVM to heavily utilize the majority of CPUs in the system, *beyond*
the CPUs KVM is already using to run vCPUs.
When deleting a memslot via KVM_SET_USER_MEMORY_REGION, the async root
zap can result in KVM occupying all logical CPUs for ~8ms, and result in
high priority tasks not being scheduled in in a timely manner. In v5.15,
which doesn't preserve unloaded roots, the issues were even more noticeable
as KVM would zap roots more frequently and could occupy all CPUs for 50ms+.
Consuming all CPUs for an extended duration can lead to significant jitter
throughout the system, e.g. on ChromeOS with virtio-gpu, deleting memslots
is a semi-frequent operation as memslots are deleted and recreated with
different host virtual addresses to react to host GPU drivers allocating
and freeing GPU blobs. On ChromeOS, the jitter manifests as audio blips
during games due to the audio server's tasks not getting scheduled in
promptly, despite the tasks having a high realtime priority.
Deleting memslots isn't exactly a fast path and should be avoided when
possible, and ChromeOS is working towards utilizing MAP_FIXED to avoid the
memslot shenanigans, but KVM is squarely in the wrong. Not to mention
that removing the async zapping eliminates a non-trivial amount of
complexity.
Note, one of the subtle behaviors hidden behind the async zapping is that
KVM would zap invalidated roots only once (ignoring partial zaps from
things like mmu_notifier events). Preserve this behavior by adding a flag
to identify roots that are scheduled to be zapped versus roots that have
already been zapped but not yet freed.
Add a comment calling out why kvm_tdp_mmu_invalidate_all_roots() can
encounter invalid roots, as it's not at all obvious why zapping
invalidated roots shouldn't simply zap all invalid roots.
Reported-by: Pattara Teerapong <pteerapong@google.com>
Cc: David Stevens <stevensd@google.com>
Cc: Yiwei Zhang<zzyiwei@google.com>
Cc: Paul Hsia <paulhsia@google.com>
Cc: stable@vger.kernel.org
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20230916003916.2545000-4-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
All callers except the MMU notifier want to process all address spaces.
Remove the address space ID argument of for_each_tdp_mmu_root_yield_safe()
and switch the MMU notifier to use __for_each_tdp_mmu_root_yield_safe().
Extracted out of a patch by Sean Christopherson <seanjc@google.com>
Cc: stable@vger.kernel.org
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The mmu_notifier path is a bit of a special snowflake, e.g. it zaps only a
single address space (because it's per-slot), and can't always yield.
Because of this, it calls kvm_tdp_mmu_zap_leafs() in ways that no one
else does.
Iterate manually over the leafs in response to an mmu_notifier
invalidation, instead of invoking kvm_tdp_mmu_zap_leafs(). Drop the
@can_yield param from kvm_tdp_mmu_zap_leafs() as its sole remaining
caller unconditionally passes "true".
Cc: stable@vger.kernel.org
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20230916003916.2545000-2-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Simplify and optimize the logic for detecting if the current/active MMU
is a TDP MMU. If the TDP MMU is globally enabled, then the active MMU is
a TDP MMU if it is direct. When TDP is enabled, so called nonpaging MMUs
are never used as the only form of shadow paging KVM uses is for nested
TDP, and the active MMU can't be direct in that case.
Rename the helper and take the vCPU instead of an arbitrary MMU, as
nonpaging MMUs can show up in the walk_mmu if L1 is using nested TDP and
L2 has paging disabled. Taking the vCPU has the added bonus of cleaning
up the callers, all of which check the current MMU but wrap code that
consumes the vCPU.
No functional change intended.
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20221012181702.3663607-9-seanjc@google.com>
[Use tdp_mmu_enabled variable. - Paolo]
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Move kvm_mmu_{init,uninit}_tdp_mmu() behind tdp_mmu_enabled. This makes
these functions consistent with the rest of the calls into the TDP MMU
from mmu.c, and which is now possible since tdp_mmu_enabled is only
modified when the x86 vendor module is loaded. i.e. It will never change
during the lifetime of a VM.
This change also enabled removing the stub definitions for 32-bit KVM,
as the compiler will just optimize the calls out like it does for all
the other TDP MMU functions.
No functional change intended.
Signed-off-by: David Matlack <dmatlack@google.com>
Reviewed-by: Isaku Yamahata <isaku.yamahata@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20220921173546.2674386-3-dmatlack@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Add a helper to convert a SPTE to its shadow page to deduplicate a
variety of flows and hopefully avoid future bugs, e.g. if KVM attempts to
get the shadow page for a SPTE without dropping high bits.
Opportunistically add a comment in mmu_free_root_page() documenting why
it treats the root HPA as a SPTE.
No functional change intended.
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20221019165618.927057-7-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Re-introduce zapping only leaf SPTEs in kvm_zap_gfn_range() and
kvm_tdp_mmu_unmap_gfn_range(), this time without losing a pending TLB
flush when processing multiple roots (including nested TDP shadow roots).
Dropping the TLB flush resulted in random crashes when running Hyper-V
Server 2019 in a guest with KSM enabled in the host (or any source of
mmu_notifier invalidations, KSM is just the easiest to force).
This effectively revert commits 873dd12217
and fcb93eb6d0, and thus restores commit
cf3e26427c, plus this delta on top:
bool kvm_tdp_mmu_zap_leafs(struct kvm *kvm, int as_id, gfn_t start, gfn_t end,
struct kvm_mmu_page *root;
for_each_tdp_mmu_root_yield_safe(kvm, root, as_id)
- flush = tdp_mmu_zap_leafs(kvm, root, start, end, can_yield, false);
+ flush = tdp_mmu_zap_leafs(kvm, root, start, end, can_yield, flush);
return flush;
}
Cc: Ben Gardon <bgardon@google.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Tested-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Message-Id: <20220325230348.2587437-1-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
If kvm->arch.tdp_mmu_zap_wq cannot be created, the failure has
to be propagated up to kvm_mmu_init_vm and kvm_arch_init_vm.
kvm_arch_init_vm also has to undo all the initialization, so
group all the MMU initialization code at the beginning and
handle cleaning up of kvm_page_track_init.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
This reverts commit cf3e26427c.
Multi-vCPU Hyper-V guests started crashing randomly on boot with the
latest kvm/queue and the problem can be bisected the problem to this
particular patch. Basically, I'm not able to boot e.g. 16-vCPU guest
successfully anymore. Both Intel and AMD seem to be affected. Reverting
the commit saves the day.
Reported-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Zap only leaf SPTEs in the TDP MMU's zap_gfn_range(), and rename various
functions accordingly. When removing mappings for functional correctness
(except for the stupid VFIO GPU passthrough memslots bug), zapping the
leaf SPTEs is sufficient as the paging structures themselves do not point
at guest memory and do not directly impact the final translation (in the
TDP MMU).
Note, this aligns the TDP MMU with the legacy/full MMU, which zaps only
the rmaps, a.k.a. leaf SPTEs, in kvm_zap_gfn_range() and
kvm_unmap_gfn_range().
Signed-off-by: Sean Christopherson <seanjc@google.com>
Reviewed-by: Ben Gardon <bgardon@google.com>
Message-Id: <20220226001546.360188-18-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
When recovering a potential hugepage that was shattered for the iTLB
multihit workaround, precisely zap only the target page instead of
iterating over the TDP MMU to find the SP that was passed in. This will
allow future simplification of zap_gfn_range() by having it zap only
leaf SPTEs.
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20220226001546.360188-14-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The root_hpa and root_pgd fields form essentially a struct kvm_mmu_root_info.
Use the struct to have more consistency between mmu->root and
mmu->prev_roots.
The patch is entirely search and replace except for cached_root_available,
which does not need a temporary struct kvm_mmu_root_info anymore.
Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
When using KVM_DIRTY_LOG_INITIALLY_SET, huge pages are not
write-protected when dirty logging is enabled on the memslot. Instead
they are write-protected once userspace invokes KVM_CLEAR_DIRTY_LOG for
the first time and only for the specific sub-region being cleared.
Enhance KVM_CLEAR_DIRTY_LOG to also try to split huge pages prior to
write-protecting to avoid causing write-protection faults on vCPU
threads. This also allows userspace to smear the cost of huge page
splitting across multiple ioctls, rather than splitting the entire
memslot as is the case when initially-all-set is not used.
Signed-off-by: David Matlack <dmatlack@google.com>
Message-Id: <20220119230739.2234394-17-dmatlack@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
When dirty logging is enabled without initially-all-set, try to split
all huge pages in the memslot down to 4KB pages so that vCPUs do not
have to take expensive write-protection faults to split huge pages.
Eager page splitting is best-effort only. This commit only adds the
support for the TDP MMU, and even there splitting may fail due to out
of memory conditions. Failures to split a huge page is fine from a
correctness standpoint because KVM will always follow up splitting by
write-protecting any remaining huge pages.
Eager page splitting moves the cost of splitting huge pages off of the
vCPU threads and onto the thread enabling dirty logging on the memslot.
This is useful because:
1. Splitting on the vCPU thread interrupts vCPUs execution and is
disruptive to customers whereas splitting on VM ioctl threads can
run in parallel with vCPU execution.
2. Splitting all huge pages at once is more efficient because it does
not require performing VM-exit handling or walking the page table for
every 4KiB page in the memslot, and greatly reduces the amount of
contention on the mmu_lock.
For example, when running dirty_log_perf_test with 96 virtual CPUs, 1GiB
per vCPU, and 1GiB HugeTLB memory, the time it takes vCPUs to write to
all of their memory after dirty logging is enabled decreased by 95% from
2.94s to 0.14s.
Eager Page Splitting is over 100x more efficient than the current
implementation of splitting on fault under the read lock. For example,
taking the same workload as above, Eager Page Splitting reduced the CPU
required to split all huge pages from ~270 CPU-seconds ((2.94s - 0.14s)
* 96 vCPU threads) to only 1.55 CPU-seconds.
Eager page splitting does increase the amount of time it takes to enable
dirty logging since it has split all huge pages. For example, the time
it took to enable dirty logging in the 96GiB region of the
aforementioned test increased from 0.001s to 1.55s.
Reviewed-by: Peter Xu <peterx@redhat.com>
Signed-off-by: David Matlack <dmatlack@google.com>
Message-Id: <20220119230739.2234394-16-dmatlack@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The "struct kvm *kvm" parameter of kvm_tdp_mmu_get_root() is not used,
so remove it. No functional change intended.
Signed-off-by: Jinrong Liang <cloudliang@tencent.com>
Message-Id: <20220125095909.38122-5-cloudliang@tencent.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Move the check for an invalid root out of kvm_tdp_mmu_get_root() and into
the one place it actually matters, tdp_mmu_next_root(), as the other user
already has an implicit validity check. A future bug fix will need to
get references to invalid roots to honor mmu_notifier requests; there's
no point in forcing what will be a common path to open code getting a
reference to a root.
No functional change intended.
Cc: stable@vger.kernel.org
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20211215011557.399940-3-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Drop the "flush" param and return values to/from the TDP MMU's helper for
zapping collapsible SPTEs. Because the helper runs with mmu_lock held
for read, not write, it uses tdp_mmu_zap_spte_atomic(), and the atomic
zap handles the necessary remote TLB flush.
Similarly, because mmu_lock is dropped and re-acquired between zapping
legacy MMUs and zapping TDP MMUs, kvm_mmu_zap_collapsible_sptes() must
handle remote TLB flushes from the legacy MMU before calling into the TDP
MMU.
Fixes: e2209710cc ("KVM: x86/mmu: Skip rmap operations if rmaps not allocated")
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20211120045046.3940942-4-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Unify the flags for rmaps and page tracking data, using a
single flag in struct kvm_arch and a single loop to go
over all the address spaces and memslots. This avoids
code duplication between alloc_all_memslots_rmaps and
kvm_page_track_enable_mmu_write_tracking.
Signed-off-by: David Stevens <stevensd@chromium.org>
[This patch is the delta between David's v2 and v3, with conflicts
fixed and my own commit message. - Paolo]
Co-developed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Pass struct kvm_page_fault to kvm_tdp_mmu_map() instead of
extracting the arguments from the struct.
Suggested-by: Isaku Yamahata <isaku.yamahata@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
This together with the next patch will fix a future race between
kvm_zap_gfn_range and the page fault handler, which will happen
when AVIC memslot is going to be only partially disabled.
The performance impact is minimal since kvm_zap_gfn_range is only
called by users, update_mtrr() and kvm_post_set_cr0().
Both only use it if the guest has non-coherent DMA, in order to
honor the guest's UC memtype.
MTRR and CD setup only happens at boot, and generally in an area
where the page tables should be small (for CD) or should not
include the affected GFNs at all (for MTRRs).
This is based on a patch suggested by Sean Christopherson:
https://lkml.org/lkml/2021/7/22/1025
Signed-off-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Maxim Levitsky <mlevitsk@redhat.com>
Message-Id: <20210810205251.424103-2-mlevitsk@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
As alluded to in commit f36f3f2846 ("KVM: add "new" argument to
kvm_arch_commit_memory_region"), a bunch of other places where struct
kvm_memory_slot is used, needs to be refactored to preserve the
"const"ness of struct kvm_memory_slot across-the-board.
Signed-off-by: Hamza Mahfooz <someguy@effective-light.com>
Message-Id: <20210713023338.57108-1-someguy@effective-light.com>
[Do not touch body of slot_rmap_walk_init. - Paolo]
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Make fast_page_fault interoperate with the TDP MMU by leveraging
walk_shadow_page_lockless_{begin,end} to acquire the RCU read lock and
introducing a new helper function kvm_tdp_mmu_fast_pf_get_last_sptep to
grab the lowest level sptep.
Suggested-by: Ben Gardon <bgardon@google.com>
Signed-off-by: David Matlack <dmatlack@google.com>
Message-Id: <20210713220957.3493520-5-dmatlack@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Acquire the RCU read lock in walk_shadow_page_lockless_begin and release
it in walk_shadow_page_lockless_end when the TDP MMU is enabled. This
should not introduce any functional changes but is used in the following
commit to make fast_page_fault interoperate with the TDP MMU.
Signed-off-by: David Matlack <dmatlack@google.com>
Message-Id: <20210713220957.3493520-4-dmatlack@google.com>
[Use if...else instead of if(){return;}]
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Treat a NULL shadow page in the "is a TDP MMU" check as valid, non-TDP
root. KVM uses a "direct" PAE paging MMU when TDP is disabled and the
guest is running with paging disabled. In that case, root_hpa points at
the pae_root page (of which only 32 bytes are used), not a standard
shadow page, and the WARN fires (a lot).
Fixes: 0b873fd7fb ("KVM: x86/mmu: Remove redundant is_tdp_mmu_enabled check")
Cc: David Matlack <dmatlack@google.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210622072454.3449146-1-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
This change simplifies the call sites slightly and also abstracts away
the implementation detail of looking at root_hpa as the mechanism for
determining if the mmu is the TDP MMU.
Suggested-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: David Matlack <dmatlack@google.com>
Message-Id: <20210617231948.2591431-4-dmatlack@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
This check is redundant because the root shadow page will only be a TDP
MMU page if is_tdp_mmu_enabled() returns true, and is_tdp_mmu_enabled()
never changes for the lifetime of a VM.
It's possible that this check was added for performance reasons but it
is unlikely that it is useful in practice since to_shadow_page() is
cheap. That being said, this patch also caches the return value of
is_tdp_mmu_root() in direct_page_fault() since there's no reason to
duplicate the call so many times, so performance is not a concern.
Suggested-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: David Matlack <dmatlack@google.com>
Message-Id: <20210617231948.2591431-3-dmatlack@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
If is_tdp_mmu_root is not inlined, the elimination of TDP MMU calls as dead
code might not work out. To avoid this, explicitly declare the stubbed
is_tdp_mmu_root on 32-bit hosts.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
TDP MMU iterator's level is identical to page table's actual level. For
instance, for the last level page table (whose entry points to one 4K
page), iter->level is 1 (PG_LEVEL_4K), and in case of 5 level paging,
the iter->level is mmu->shadow_root_level, which is 5. However, struct
kvm_mmu_page's level currently is not set correctly when it is allocated
in kvm_tdp_mmu_map(). When iterator hits non-present SPTE and needs to
allocate a new child page table, currently iter->level, which is the
level of the page table where the non-present SPTE belongs to, is used.
This results in struct kvm_mmu_page's level always having its parent's
level (excpet root table's level, which is initialized explicitly using
mmu->shadow_root_level).
This is kinda wrong, and not consistent with existing non TDP MMU code.
Fortuantely sp->role.level is only used in handle_removed_tdp_mmu_page()
and kvm_tdp_mmu_zap_sp(), and they are already aware of this and behave
correctly. However to make it consistent with legacy MMU code (and fix
the issue that both root page table and its child page table have
shadow_root_level), use iter->level - 1 in kvm_tdp_mmu_map(), and change
handle_removed_tdp_mmu_page() and kvm_tdp_mmu_zap_sp() accordingly.
Reviewed-by: Ben Gardon <bgardon@google.com>
Signed-off-by: Kai Huang <kai.huang@intel.com>
Message-Id: <bcb6569b6e96cb78aaa7b50640e6e6b53291a74e.1623717884.git.kai.huang@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
If the TDP MMU is in use, wait to allocate the rmaps until the shadow
MMU is actually used. (i.e. a nested VM is launched.) This saves memory
equal to 0.2% of guest memory in cases where the TDP MMU is used and
there are no nested guests involved.
Signed-off-by: Ben Gardon <bgardon@google.com>
Message-Id: <20210518173414.450044-8-bgardon@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Prepare for write protecting large page lazily during dirty log tracking,
for which we will only need to write protect gfns at large page
granularity.
No functional or performance change expected.
Signed-off-by: Keqian Zhu <zhukeqian1@huawei.com>
Message-Id: <20210429034115.35560-2-zhukeqian1@huawei.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
To avoid saddling a vCPU thread with the work of tearing down an entire
paging structure, take a reference on each root before they become
obsolete, so that the thread initiating the fast invalidation can tear
down the paging structure and (most likely) release the last reference.
As a bonus, this teardown can happen under the MMU lock in read mode so
as not to block the progress of vCPU threads.
Signed-off-by: Ben Gardon <bgardon@google.com>
Message-Id: <20210401233736.638171-14-bgardon@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Provide a real mechanism for fast invalidation by marking roots as
invalid so that their reference count will quickly fall to zero
and they will be torn down.
One negative side affect of this approach is that a vCPU thread will
likely drop the last reference to a root and be saddled with the work of
tearing down an entire paging structure. This issue will be resolved in
a later commit.
Signed-off-by: Ben Gardon <bgardon@google.com>
Message-Id: <20210401233736.638171-13-bgardon@google.com>
[Move the loop to tdp_mmu.c, otherwise compilation fails on 32-bit. - Paolo]
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
To reduce lock contention and interference with page fault handlers,
allow the TDP MMU function to zap a GFN range to operate under the MMU
read lock.
Signed-off-by: Ben Gardon <bgardon@google.com>
Message-Id: <20210401233736.638171-10-bgardon@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
To reduce dependence on the MMU write lock, don't rely on the assumption
that the atomic operation in kvm_tdp_mmu_get_root will always succeed.
By not relying on that assumption, threads do not need to hold the MMU
lock in write mode in order to take a reference on a TDP MMU root.
In the root iterator, this change means that some roots might have to be
skipped if they are found to have a zero refcount. This will still never
happen as of this patch, but a future patch will need that flexibility to
make the root iterator safe under the MMU read lock.
Signed-off-by: Ben Gardon <bgardon@google.com>
Message-Id: <20210401233736.638171-8-bgardon@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
In order to parallelize more operations for the TDP MMU, make the
refcount on TDP MMU roots atomic, so that a future patch can allow
multiple threads to take a reference on the root concurrently, while
holding the MMU lock in read mode.
Signed-off-by: Ben Gardon <bgardon@google.com>
Message-Id: <20210401233736.638171-7-bgardon@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
kvm_tdp_mmu_put_root and kvm_tdp_mmu_free_root are always called
together, so merge the functions to simplify TDP MMU root refcounting /
freeing.
Signed-off-by: Ben Gardon <bgardon@google.com>
Message-Id: <20210401233736.638171-5-bgardon@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The TDP MMU is almost the only user of kvm_mmu_get_root and
kvm_mmu_put_root. There is only one use of put_root in mmu.c for the
legacy / shadow MMU. Open code that one use and move the get / put
functions to the TDP MMU so they can be extended in future commits.
No functional change intended.
Signed-off-by: Ben Gardon <bgardon@google.com>
Message-Id: <20210401233736.638171-3-bgardon@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
kvm_tdp_mmu_zap_collapsible_sptes unnecessarily removes the const
qualifier from its memlsot argument, leading to a compiler warning. Add
the const annotation and pass it to subsequent functions.
Signed-off-by: Ben Gardon <bgardon@google.com>
Message-Id: <20210401233736.638171-2-bgardon@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Move the hva->gfn lookup for MMU notifiers into common code. Every arch
does a similar lookup, and some arch code is all but identical across
multiple architectures.
In addition to consolidating code, this will allow introducing
optimizations that will benefit all architectures without incurring
multiple walks of the memslots, e.g. by taking mmu_lock if and only if a
relevant range exists in the memslots.
The use of __always_inline to avoid indirect call retpolines, as done by
x86, may also benefit other architectures.
Consolidating the lookups also fixes a wart in x86, where the legacy MMU
and TDP MMU each do their own memslot walks.
Lastly, future enhancements to the memslot implementation, e.g. to add an
interval tree to track host address, will need to touch far less arch
specific code.
MIPS, PPC, and arm64 will be converted one at a time in future patches.
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210402005658.3024832-3-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
