SMEP is disabled if CPU is in non-paging mode in hardware.
However KVM always uses paging mode to emulate guest non-paging
mode with TDP. To emulate this behavior, SMEP needs to be manually
disabled when guest switches to non-paging mode.
We met an issue that, SMP Linux guest with recent kernel (enable
SMEP support, for example, 3.5.3) would crash with triple fault if
setting unrestricted_guest=0. This is because KVM uses an identity
mapping page table to emulate the non-paging mode, where the page
table is set with USER flag. If SMEP is still enabled in this case,
guest will meet unhandlable page fault and then crash.
Reviewed-by: Gleb Natapov <gleb@redhat.com>
Reviewed-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Dongxiao Xu <dongxiao.xu@intel.com>
Signed-off-by: Xiantao Zhang <xiantao.zhang@intel.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
This reverts commit bd4c86eaa6.
There is not user for kvm_mmu_isolate_page() any more.
Signed-off-by: Gleb Natapov <gleb@redhat.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Gust page walker puts only present ptes into ptes[] array. No need to
check it again.
Signed-off-by: Gleb Natapov <gleb@redhat.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Move base_role.nxe initialisation to where all other roles are initialized.
Signed-off-by: Gleb Natapov <gleb@redhat.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
spte_is_locklessly_modifiable() checks that both SPTE_HOST_WRITEABLE and
SPTE_MMU_WRITEABLE are present on spte. Make it more explicit.
Signed-off-by: Gleb Natapov <gleb@redhat.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
As Xiao pointed out, there are a few problems with it:
- kvm_arch_commit_memory_region() write protects the memory slot only
for GET_DIRTY_LOG when modifying the flags.
- FNAME(sync_page) uses the old spte value to set a new one without
checking KVM_MEM_READONLY flag.
Since we flush all shadow pages when creating a new slot, the simplest
fix is to disallow such problematic flag changes: this is safe because
no one is doing such things.
Reviewed-by: Gleb Natapov <gleb@redhat.com>
Signed-off-by: Takuya Yoshikawa <yoshikawa_takuya_b1@lab.ntt.co.jp>
Cc: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com>
Cc: Alex Williamson <alex.williamson@redhat.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
KVM_SET_USER_MEMORY_REGION forces __kvm_set_memory_region() to identify
what kind of change is being requested by checking the arguments. The
current code does this checking at various points in code and each
condition being used there is not easy to understand at first glance.
This patch consolidates these checks and introduces an enum to name the
possible changes to clean up the code.
Although this does not introduce any functional changes, there is one
change which optimizes the code a bit: if we have nothing to change, the
new code returns 0 immediately.
Note that the return value for this case cannot be changed since QEMU
relies on it: we noticed this when we changed it to -EINVAL and got a
section mismatch error at the final stage of live migration.
Signed-off-by: Takuya Yoshikawa <yoshikawa_takuya_b1@lab.ntt.co.jp>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Instructions with long displacement have a signed displacement.
Currently the sign bit is interpreted as 2^20: Lets fix it by doing the
sign extension from 20bit to 32bit and then use it as a signed variable
in the addition (see kvm_s390_get_base_disp_rsy).
Furthermore, there are lots of "int" in that code. This is problematic,
because shifting on a signed integer is undefined/implementation defined
if the bit value happens to be negative.
Fortunately the promotion rules will make the right hand side unsigned
anyway, so there is no real problem right now.
Let's convert them anyway to unsigned where appropriate to avoid
problems if the code is changed or copy/pasted later on.
Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
Reviewed-by: Cornelia Huck <cornelia.huck@de.ibm.com>
Signed-off-by: Gleb Natapov <gleb@redhat.com>
virtio_ccw_setup_vq() failed to unwind correctly on errors. In
particular, it failed to delete the virtqueue on errors, leading to
list corruption when virtio_ccw_del_vqs() iterated over a virtqueue
that had not been added to the vcdev's list.
Fix this with redoing the error unwinding in virtio_ccw_setup_vq(),
using a single path for all errors.
Signed-off-by: Cornelia Huck <cornelia.huck@de.ibm.com>
Reviewed-by: Christian Borntraeger <borntraeger@de.ibm.com>
Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
Signed-off-by: Gleb Natapov <gleb@redhat.com>
On store status we need to copy the current state of registers
into a save area. Currently we might save stale versions:
The sie state descriptor doesnt have fields for guest ACRS,FPRS,
those registers are simply stored in the host registers. The host
program must copy these away if needed. We do that in vcpu_put/load.
If we now do a store status in KVM code between vcpu_put/load, the
saved values are not up-to-date. Lets collect the ACRS/FPRS before
saving them.
This also fixes some strange problems with hotplug and virtio-ccw,
since the low level machine check handler (on hotplug a machine check
will happen) will revalidate all registers with the content of the
save area.
Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
CC: stable@vger.kernel.org
Signed-off-by: Gleb Natapov <gleb@redhat.com>
yield_to returns -ESRCH, When source and target of yield_to
run queue length is one. When we see three successive failures of
yield_to we assume we are in potential undercommit case and abort
from PLE handler.
The assumption is backed by low probability of wrong decision
for even worst case scenarios such as average runqueue length
between 1 and 2.
More detail on rationale behind using three tries:
if p is the probability of finding rq length one on a particular cpu,
and if we do n tries, then probability of exiting ple handler is:
p^(n+1) [ because we would have come across one source with rq length
1 and n target cpu rqs with length 1 ]
so
num tries: probability of aborting ple handler (1.5x overcommit)
1 1/4
2 1/8
3 1/16
We can increase this probability with more tries, but the problem is
the overhead.
Also, If we have tried three times that means we would have iterated
over 3 good eligible vcpus along with many non-eligible candidates. In
worst case if we iterate all the vcpus, we reduce 1x performance and
overcommit performance get hit.
note that we do not update last boosted vcpu in failure cases.
Thank Avi for raising question on aborting after first fail from yield_to.
Reviewed-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com>
Signed-off-by: Raghavendra K T <raghavendra.kt@linux.vnet.ibm.com>
Tested-by: Chegu Vinod <chegu_vinod@hp.com>
Signed-off-by: Gleb Natapov <gleb@redhat.com>
In case of undercomitted scenarios, especially in large guests
yield_to overhead is significantly high. when run queue length of
source and target is one, take an opportunity to bail out and return
-ESRCH. This return condition can be further exploited to quickly come
out of PLE handler.
(History: Raghavendra initially worked on break out of kvm ple handler upon
seeing source runqueue length = 1, but it had to export rq length).
Peter came up with the elegant idea of return -ESRCH in scheduler core.
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Raghavendra, Checking the rq length of target vcpu condition added.(thanks Avi)
Reviewed-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com>
Signed-off-by: Raghavendra K T <raghavendra.kt@linux.vnet.ibm.com>
Acked-by: Andrew Jones <drjones@redhat.com>
Tested-by: Chegu Vinod <chegu_vinod@hp.com>
Signed-off-by: Gleb Natapov <gleb@redhat.com>
Virtual interrupt delivery avoids KVM to inject vAPIC interrupts
manually, which is fully taken care of by the hardware. This needs
some special awareness into existing interrupr injection path:
- for pending interrupt, instead of direct injection, we may need
update architecture specific indicators before resuming to guest.
- A pending interrupt, which is masked by ISR, should be also
considered in above update action, since hardware will decide
when to inject it at right time. Current has_interrupt and
get_interrupt only returns a valid vector from injection p.o.v.
Reviewed-by: Marcelo Tosatti <mtosatti@redhat.com>
Signed-off-by: Kevin Tian <kevin.tian@intel.com>
Signed-off-by: Yang Zhang <yang.z.zhang@Intel.com>
Signed-off-by: Gleb Natapov <gleb@redhat.com>
basically to benefit from apicv, we need to enable virtualized x2apic mode.
Currently, we only enable it when guest is really using x2apic.
Also, clear MSR bitmap for corresponding x2apic MSRs when guest enabled x2apic:
0x800 - 0x8ff: no read intercept for apicv register virtualization,
except APIC ID and TMCCT which need software's assistance to
get right value.
Reviewed-by: Marcelo Tosatti <mtosatti@redhat.com>
Signed-off-by: Kevin Tian <kevin.tian@intel.com>
Signed-off-by: Yang Zhang <yang.z.zhang@Intel.com>
Signed-off-by: Gleb Natapov <gleb@redhat.com>
We've been ignoring read-only mappings and programming everything
into the iommu as read-write. Fix this to only include the write
access flag when read-only is not set.
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Signed-off-by: Gleb Natapov <gleb@redhat.com>
Memory slot flags can be altered without changing other parameters of
the slot. The read-only attribute is the only one the IOMMU cares
about, so generate an un-map, re-map when this occurs. This also
avoid unnecessarily re-mapping the slot when no IOMMU visible changes
are made.
Reviewed-by: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com>
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Signed-off-by: Gleb Natapov <gleb@redhat.com>
'pushq' doesn't exist on i386. Replace with 'push', which should work
since the operand is a register.
Signed-off-by: Avi Kivity <avi.kivity@gmail.com>
Signed-off-by: Gleb Natapov <gleb@redhat.com>
If emulate_invalid_guest_state=false vmx->emulation_required is never
actually used, but it ends up to be always set to true since
handle_invalid_guest_state(), the only place it is reset back to
false, is never called. This, besides been not very clean, makes vmexit
and vmentry path to check emulate_invalid_guest_state needlessly.
The patch fixes that by keeping emulation_required coherent with
emulate_invalid_guest_state setting.
Signed-off-by: Gleb Natapov <gleb@redhat.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
If VMX reports segment as unusable, zero descriptor passed by the emulator
before returning. Such descriptor will be considered not present by the
emulator.
Signed-off-by: Gleb Natapov <gleb@redhat.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Usability is returned in unusable field, so not need to clobber entire
AR. Callers have to know how to deal with unusable segments already
since if emulate_invalid_guest_state=true AR is not zeroed.
Signed-off-by: Gleb Natapov <gleb@redhat.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
vmx->rmode.vm86_active is never true is unrestricted guest is enabled.
Make it more explicit that neither enter_pmode() nor enter_rmode() is
called in this case.
Signed-off-by: Gleb Natapov <gleb@redhat.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
There is no reason for it. If state is suitable for vmentry it
will be detected during guest entry and no emulation will happen.
Signed-off-by: Gleb Natapov <gleb@redhat.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Since vmx_get_cpl() always returns 0 when VCPU is in real mode it is no
longer needed. Also reset CPL cache to zero during transaction to
protected mode since transaction may happen while CS.selectors & 3 != 0,
but in reality CPL is 0.
Signed-off-by: Gleb Natapov <gleb@redhat.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Make fastop opcodes usable in other emulations.
Reviewed-by: Gleb Natapov <gleb@redhat.com>
Signed-off-by: Avi Kivity <avi.kivity@gmail.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
This is a bit of a special case since we don't have the usual
byte/word/long/quad switch; instead we switch on the condition code embedded
in the instruction.
Reviewed-by: Gleb Natapov <gleb@redhat.com>
Signed-off-by: Avi Kivity <avi.kivity@gmail.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
The current reexecute_instruction can not well detect the failed instruction
emulation. It allows guest to retry all the instructions except it accesses
on error pfn
For example, some cases are nested-write-protect - if the page we want to
write is used as PDE but it chains to itself. Under this case, we should
stop the emulation and report the case to userspace
Reviewed-by: Gleb Natapov <gleb@redhat.com>
Signed-off-by: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Currently, reexecute_instruction refused to retry all instructions if
tdp is enabled. If nested npt is used, the emulation may be caused by
shadow page, it can be fixed by dropping the shadow page. And the only
condition that tdp can not retry the instruction is the access fault
on error pfn
Reviewed-by: Gleb Natapov <gleb@redhat.com>
Signed-off-by: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Little cleanup for reexecute_instruction, also use gpa_to_gfn in
retry_instruction
Reviewed-by: Gleb Natapov <gleb@redhat.com>
Signed-off-by: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
One such variable, slot, is enough for holding a pointer temporarily.
We also remove another local variable named slot, which is limited in
a block, since it is confusing to have the same name in this function.
Reviewed-by: Marcelo Tosatti <mtosatti@redhat.com>
Signed-off-by: Takuya Yoshikawa <yoshikawa_takuya_b1@lab.ntt.co.jp>
Signed-off-by: Gleb Natapov <gleb@redhat.com>
Don't need the check for deleting an existing slot or just modifiying
the flags.
Reviewed-by: Marcelo Tosatti <mtosatti@redhat.com>
Signed-off-by: Takuya Yoshikawa <yoshikawa_takuya_b1@lab.ntt.co.jp>
Signed-off-by: Gleb Natapov <gleb@redhat.com>
This makes the separation between the sanity checks and the rest of the
code a bit clearer.
Reviewed-by: Marcelo Tosatti <mtosatti@redhat.com>
Signed-off-by: Takuya Yoshikawa <yoshikawa_takuya_b1@lab.ntt.co.jp>
Signed-off-by: Gleb Natapov <gleb@redhat.com>
the variable inti should be freed in the branch CPUSTAT_STOPPED.
Signed-off-by: Cong Ding <dinggnu@gmail.com>
Signed-off-by: Cornelia Huck <cornelia.huck@de.ibm.com>
Signed-off-by: Gleb Natapov <gleb@redhat.com>
If the userspace starts dirty logging for a large slot, say 64GB of
memory, kvm_mmu_slot_remove_write_access() needs to hold mmu_lock for
a long time such as tens of milliseconds. This patch controls the lock
hold time by asking the scheduler if we need to reschedule for others.
One penalty for this is that we need to flush TLBs before releasing
mmu_lock. But since holding mmu_lock for a long time does affect not
only the guest, vCPU threads in other words, but also the host as a
whole, we should pay for that.
In practice, the cost will not be so high because we can protect a fair
amount of memory before being rescheduled: on my test environment,
cond_resched_lock() was called only once for protecting 12GB of memory
even without THP. We can also revisit Avi's "unlocked TLB flush" work
later for completely suppressing extra TLB flushes if needed.
Reviewed-by: Marcelo Tosatti <mtosatti@redhat.com>
Signed-off-by: Takuya Yoshikawa <yoshikawa_takuya_b1@lab.ntt.co.jp>
Signed-off-by: Gleb Natapov <gleb@redhat.com>
Better to place mmu_lock handling and TLB flushing code together since
this is a self-contained function.
Reviewed-by: Marcelo Tosatti <mtosatti@redhat.com>
Signed-off-by: Takuya Yoshikawa <yoshikawa_takuya_b1@lab.ntt.co.jp>
Signed-off-by: Gleb Natapov <gleb@redhat.com>
No reason to make callers take mmu_lock since we do not need to protect
kvm_mmu_change_mmu_pages() and kvm_mmu_slot_remove_write_access()
together by mmu_lock in kvm_arch_commit_memory_region(): the former
calls kvm_mmu_commit_zap_page() and flushes TLBs by itself.
Note: we do not need to protect kvm->arch.n_requested_mmu_pages by
mmu_lock as can be seen from the fact that it is read locklessly.
Reviewed-by: Marcelo Tosatti <mtosatti@redhat.com>
Signed-off-by: Takuya Yoshikawa <yoshikawa_takuya_b1@lab.ntt.co.jp>
Signed-off-by: Gleb Natapov <gleb@redhat.com>
This makes it possible to release mmu_lock and reschedule conditionally
in a later patch. Although this may increase the time needed to protect
the whole slot when we start dirty logging, the kernel should not allow
the userspace to trigger something that will hold a spinlock for such a
long time as tens of milliseconds: actually there is no limit since it
is roughly proportional to the number of guest pages.
Another point to note is that this patch removes the only user of
slot_bitmap which will cause some problems when we increase the number
of slots further.
Reviewed-by: Marcelo Tosatti <mtosatti@redhat.com>
Signed-off-by: Takuya Yoshikawa <yoshikawa_takuya_b1@lab.ntt.co.jp>
Signed-off-by: Gleb Natapov <gleb@redhat.com>
