KVM: s390: Fixes for kvm/next (3.19) and stable
1. We should flush TLBs for load control instruction emulation (stable)
2. A workaround for a compiler bug that renders ACCESS_ONCE broken (stable)
3. Fix program check handling for load control
4. Documentation Fix
Documentation uses incorrect attribute names for some vm device
attributes: fix this.
Signed-off-by: Dominik Dingel <dingel@linux.vnet.ibm.com>
Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
According to the architecture all instructions are suppressing if memory
access is prohibited due to DAT protection, unless stated otherwise for
an instruction.
The lctl[g]/stctl[g] implementations handled this incorrectly since
control register handling was done piecemeal, which means they had
terminating instead of suppressing semantics.
This patch fixes this.
Signed-off-by: Heiko Carstens <heiko.carstens@de.ibm.com>
Reviewed-by: Thomas Huth <thuth@linux.vnet.ibm.com>
Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com>
Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
some control register changes will flush some aspects of the CPU, e.g.
POP explicitely mentions that for CR9-CR11 "TLBs may be cleared".
Instead of trying to be clever and only flush on specific CRs, let
play safe and flush on all lctl(g) as future machines might define
new bits in CRs. Load control intercept should not happen that often.
Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
Acked-by: Cornelia Huck <cornelia.huck@de.ibm.com>
Reviewed-by: David Hildenbrand <dahi@linux.vnet.ibm.com>
Cc: stable@vger.kernel.org
ipte_unlock_siif uses cmpxchg to replace the in-memory data of the ipte
lock together with ACCESS_ONCE for the intial read.
union ipte_control {
unsigned long val;
struct {
unsigned long k : 1;
unsigned long kh : 31;
unsigned long kg : 32;
};
};
[...]
static void ipte_unlock_siif(struct kvm_vcpu *vcpu)
{
union ipte_control old, new, *ic;
ic = &vcpu->kvm->arch.sca->ipte_control;
do {
new = old = ACCESS_ONCE(*ic);
new.kh--;
if (!new.kh)
new.k = 0;
} while (cmpxchg(&ic->val, old.val, new.val) != old.val);
if (!new.kh)
wake_up(&vcpu->kvm->arch.ipte_wq);
}
The new value, is loaded twice from memory with gcc 4.7.2 of
fedora 18, despite the ACCESS_ONCE:
--->
l %r4,0(%r3) <--- load first 32 bit of lock (k and kh) in r4
alfi %r4,2147483647 <--- add -1 to r4
llgtr %r4,%r4 <--- zero out the sign bit of r4
lg %r1,0(%r3) <--- load all 64 bit of lock into new
lgr %r2,%r1 <--- load the same into old
risbg %r1,%r4,1,31,32 <--- shift and insert r4 into the bits 1-31 of
new
llihf %r4,2147483647
ngrk %r4,%r1,%r4
jne aa0 <ipte_unlock+0xf8>
nihh %r1,32767
lgr %r4,%r2
csg %r4,%r1,0(%r3)
cgr %r2,%r4
jne a70 <ipte_unlock+0xc8>
If the memory value changes between the first load (l) and the second
load (lg) we are broken. If that happens VCPU threads will hang
(unkillable) in handle_ipte_interlock.
Andreas Krebbel analyzed this and tracked it down to a compiler bug in
that version:
"while it is not that obvious the C99 standard basically forbids
duplicating the memory access also in that case. For an argumentation of
a similiar case please see:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=22278#c43
For the implementation-defined cases regarding volatile there are some
GCC-specific clarifications which can be found here:
https://gcc.gnu.org/onlinedocs/gcc/Volatiles.html#Volatiles
I've tracked down the problem with a reduced testcase. The problem was
that during a tree level optimization (SRA - scalar replacement of
aggregates) the volatile marker is lost. And an RTL level optimizer (CSE
- common subexpression elimination) then propagated the memory read into
its second use introducing another access to the memory location. So
indeed Christian's suspicion that the union access has something to do
with it is correct (since it triggered the SRA optimization).
This issue has been reported and fixed in the GCC 4.8 development cycle:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=58145"
This patch replaces the ACCESS_ONCE scheme with a barrier() based scheme
that should work for all supported compilers.
Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
Cc: stable@vger.kernel.org # v3.16+
We can use get_cpu() and put_cpu() to replace
preempt_disable()/cpu = smp_processor_id() and
preempt_enable() for slightly better code.
Signed-off-by: Tiejun Chen <tiejun.chen@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
We mirror a subset of these registers in separate variables.
Using them directly should be faster.
Signed-off-by: Radim Krčmář <rkrcmar@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
APIC-write VM exits are "trap-like": they save CS:RIP values for the
instruction after the write, and more importantly, the handler will
already see the new value in the virtual-APIC page. This means that
apic_reg_write cannot use kvm_apic_get_reg to omit timer cancelation
when mode changes.
timer_mode_mask shouldn't be changing as it depends on cpuid.
Signed-off-by: Radim Krčmář <rkrcmar@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
APIC-write VM exits are "trap-like": they save CS:RIP values for the
instruction after the write, and more importantly, the handler will
already see the new value in the virtual-APIC page.
This caused a bug if you used KVM_SET_IRQCHIP to set the SW-enabled bit
in the SPIV register. The chain of events is as follows:
* When the irqchip is added to the destination VM, the apic_sw_disabled
static key is incremented (1)
* When the KVM_SET_IRQCHIP ioctl is invoked, it is decremented (0)
* When the guest disables the bit in the SPIV register, e.g. as part of
shutdown, apic_set_spiv does not notice the change and the static key is
_not_ incremented.
* When the guest is destroyed, the static key is decremented (-1),
resulting in this trace:
WARNING: at kernel/jump_label.c:81 __static_key_slow_dec+0xa6/0xb0()
jump label: negative count!
[<ffffffff816bf898>] dump_stack+0x19/0x1b
[<ffffffff8107c6f1>] warn_slowpath_common+0x61/0x80
[<ffffffff8107c76c>] warn_slowpath_fmt+0x5c/0x80
[<ffffffff811931e6>] __static_key_slow_dec+0xa6/0xb0
[<ffffffff81193226>] static_key_slow_dec_deferred+0x16/0x20
[<ffffffffa0637698>] kvm_free_lapic+0x88/0xa0 [kvm]
[<ffffffffa061c63e>] kvm_arch_vcpu_uninit+0x2e/0xe0 [kvm]
[<ffffffffa05ff301>] kvm_vcpu_uninit+0x21/0x40 [kvm]
[<ffffffffa067cec7>] vmx_free_vcpu+0x47/0x70 [kvm_intel]
[<ffffffffa061bc50>] kvm_arch_vcpu_free+0x50/0x60 [kvm]
[<ffffffffa061ca22>] kvm_arch_destroy_vm+0x102/0x260 [kvm]
[<ffffffff810b68fd>] ? synchronize_srcu+0x1d/0x20
[<ffffffffa06030d1>] kvm_put_kvm+0xe1/0x1c0 [kvm]
[<ffffffffa06036f8>] kvm_vcpu_release+0x18/0x20 [kvm]
[<ffffffff81215c62>] __fput+0x102/0x310
[<ffffffff81215f4e>] ____fput+0xe/0x10
[<ffffffff810ab664>] task_work_run+0xb4/0xe0
[<ffffffff81083944>] do_exit+0x304/0xc60
[<ffffffff816c8dfc>] ? _raw_spin_unlock_irq+0x2c/0x50
[<ffffffff810fd22d>] ? trace_hardirqs_on_caller+0xfd/0x1c0
[<ffffffff8108432c>] do_group_exit+0x4c/0xc0
[<ffffffff810843b4>] SyS_exit_group+0x14/0x20
[<ffffffff816d33a9>] system_call_fastpath+0x16/0x1b
Signed-off-by: Radim Krčmář <rkrcmar@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
commit 72dc67a696 ("KVM: remove the usage of the mmap_sem for the protection of the memory slots.")
changed the lock which will be taken. This should be reflected in the function
commentary.
Signed-off-by: Dominik Dingel <dingel@linux.vnet.ibm.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
No kernel ever reported KVM_CAP_DEVICE_MSIX, KVM_CAP_DEVICE_MSI,
KVM_CAP_DEVICE_ASSIGNMENT, KVM_CAP_DEVICE_DEASSIGNMENT.
This makes the documentation wrong, and no application ever
written to use these capabilities has a chance to work correctly.
The only way to detect support is to try, and test errno for ENOTTY.
That's unfortunate, but we can't fix the past.
Document the actual semantics, and drop the definitions from
the exported header to make it easier for application
developers to note and fix the bug.
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The check in kvm_set_lapic_tscdeadline_msr() was trying to prevent a
situation where we lose a pending deadline timer in a MSR write.
Losing it is fine, because it effectively occurs before the timer fired,
so we should be able to cancel or postpone it.
Another problem comes from interaction with QEMU, or other userspace
that can set deadline MSR without a good reason, when timer is already
pending: one guest's deadline request results in more than one
interrupt because one is injected immediately on MSR write from
userspace and one through hrtimer later.
The solution is to remove the injection when replacing a pending timer
and to improve the usual QEMU path, we inject without a hrtimer when the
deadline has already passed.
Signed-off-by: Radim Krčmář <rkrcmar@redhat.com>
Reported-by: Nadav Amit <namit@cs.technion.ac.il>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Make the code reusable.
If the timer was already pending, we shouldn't be waiting in a queue,
so wake_up can be skipped, simplifying the path.
There is no 'reinject' case => the comment is removed.
Current race behaves correctly.
Signed-off-by: Radim Krčmář <rkrcmar@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
When commit 6adba52742 (KVM: Let host know whether the guest can
handle async PF in non-userspace context.) is introduced, actually
bit 2 still is reserved and should be zero. Instead, bit 1 is 1 to
indicate if asynchronous page faults can be injected when vcpu is
in cpl == 0, and also please see this,
in the file kvm_para.h, #define KVM_ASYNC_PF_SEND_ALWAYS (1 << 1).
Signed-off-by: Tiejun Chen <tiejun.chen@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
If DR4/5 is accessed when it is unavailable (since CR4.DE is set), then #UD
should be generated even if CPL>0. This is according to Intel SDM Table 6-2:
"Priority Among Simultaneous Exceptions and Interrupts".
Note, that this may happen on the first DR access, even if the host does not
sets debug breakpoints. Obviously, it occurs when the host debugs the guest.
This patch moves the DR4/5 checks from __kvm_set_dr/_kvm_get_dr to handle_dr.
The emulator already checks DR4/5 availability in check_dr_read. Nested
virutalization related calls to kvm_set_dr/kvm_get_dr would not like to inject
exceptions to the guest.
As for SVM, the patch follows the previous logic as much as possible. Anyhow,
it appears the DR interception code might be buggy - even if the DR access
may cause an exception, the instruction is skipped.
Signed-off-by: Nadav Amit <namit@cs.technion.ac.il>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
When read access is performed using a readable code segment, the "conforming"
and "non-conforming" checks should not be done. As a result, read using
non-conforming readable code segment fails.
This is according to Intel SDM 5.6.1 ("Accessing Data in Code Segments").
The fix is not to perform the "non-conforming" checks if the access is not a
fetch; the relevant checks are already done when loading the segment.
Signed-off-by: Nadav Amit <namit@cs.technion.ac.il>
Reviewed-by: Radim Krčmář <rkrcmar@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
DR7.LE should be cleared during task-switch. This feature is poorly documented.
For reference, see:
http://pdos.csail.mit.edu/6.828/2005/readings/i386/s12_02.htm
SDM [17.2.4]:
This feature is not supported in the P6 family processors, later IA-32
processors, and Intel 64 processors.
AMD [2:13.1.1.4]:
This bit is ignored by implementations of the AMD64 architecture.
Intel's formulation could mean that it isn't even zeroed, but current
hardware indeed does not behave like that.
Signed-off-by: Nadav Amit <namit@cs.technion.ac.il>
Reviewed-by: Radim Krčmář <rkrcmar@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
In long-mode, when the address size is 4 bytes, the linear address is not
truncated as the emulator mistakenly does. Instead, the offset within the
segment (the ea field) should be truncated according to the address size.
As Intel SDM says: "In 64-bit mode, the effective address components are added
and the effective address is truncated ... before adding the full 64-bit
segment base."
Signed-off-by: Nadav Amit <namit@cs.technion.ac.il>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Intel SDM 17.2.4 (Debug Control Register (DR7)) says: "The processor clears the
GD flag upon entering to the debug exception handler." This sentence may be
misunderstood as if it happens only on #DB due to debug-register protection,
but it happens regardless to the cause of the #DB.
Fix the behavior to match both real hardware and Bochs.
Signed-off-by: Nadav Amit <namit@cs.technion.ac.il>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
KVM does not deliver x2APIC broadcast messages with physical mode. Intel SDM
(10.12.9 ICR Operation in x2APIC Mode) states: "A destination ID value of
FFFF_FFFFH is used for broadcast of interrupts in both logical destination and
physical destination modes."
In addition, the local-apic enables cluster mode broadcast. As Intel SDM
10.6.2.2 says: "Broadcast to all local APICs is achieved by setting all
destination bits to one." This patch enables cluster mode broadcast.
The fix tries to combine broadcast in different modes through a unified code.
One rare case occurs when the source of IPI has its APIC disabled. In such
case, the source can still issue IPIs, but since the source is not obliged to
have the same LAPIC mode as the enabled ones, we cannot rely on it.
Since it is a rare case, it is unoptimized and done on the slow-path.
Signed-off-by: Nadav Amit <namit@cs.technion.ac.il>
Reviewed-by: Radim Krčmář <rkrcmar@redhat.com>
Reviewed-by: Wanpeng Li <wanpeng.li@linux.intel.com>
[As per Radim's review, use unsigned int for X2APIC_BROADCAST, return bool from
kvm_apic_broadcast. - Paolo]
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
CR4.TSD is guest-owned; don't trap writes to it in VMX guests. This
avoids a VM exit on context switches into or out of a PR_TSC_SIGSEGV
task.
I think that this fixes an unintentional side-effect of:
4c38609ac5 KVM: VMX: Make guest cr4 mask more conservative
Signed-off-by: Andy Lutomirski <luto@amacapital.net>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
If the operand size is not 64-bit, then the sysexit instruction should assign
ECX to RSP and EDX to RIP. The current code assigns the full 64-bits.
Fix it by masking.
Signed-off-by: Nadav Amit <namit@cs.technion.ac.il>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
In 64-bit, stack operations default to 64-bits, but can be overriden (to
16-bit) using opsize override prefix. In contrast, near-branches are always
64-bit. This patch distinguish between the different behaviors.
Signed-off-by: Nadav Amit <namit@cs.technion.ac.il>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Breaking grp45 to the relevant functions to speed up the emulation and simplify
the code. In addition, it is necassary the next patch will distinguish between
far and near branches according to the flags.
Signed-off-by: Nadav Amit <namit@cs.technion.ac.il>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
