During guest entry/exit, we switch over to/from the guest MMU context
and we cannot take exceptions in the hypervisor code.

Since ftrace may be enabled and since it can result in us taking a trap,
disable ftrace by setting paca->ftrace_enabled to zero. There are two
paths through which we enter/exit a guest:
1. If we are the vcore runner, then we enter the guest via
__kvmppc_vcore_entry() and we disable ftrace around this. This is always
the case for Power9, and for the primary thread on Power8.
2. If we are a secondary thread in Power8, then we would be in nap due
to SMT being disabled. We are woken up by an IPI to enter the guest. In
this scenario, we enter the guest through kvm_start_guest(). We disable
ftrace at this point. In this scenario, ftrace would only get re-enabled
on the secondary thread when SMT is re-enabled (via start_secondary()).
Signed-off-by: NNaveen N. Rao <naveen.n.rao@linux.vnet.ibm.com>
Signed-off-by: NMichael Ellerman <mpe@ellerman.id.au>

SLOF checks for 'sc 1' (hypercall) support by issuing a hcall with
H_SET_DABR. Since the recent commit e8ebedbf ("KVM: PPC: Book3S
HV: Return error from h_set_dabr() on POWER9") changed H_SET_DABR to
return H_UNSUPPORTED on Power9, we see guest boot failures, the
symptom is the boot seems to just stop in SLOF, eg:

  SLOF ***************************************************************
  QEMU Starting
   Build Date = Sep 24 2017 12:23:07
   FW Version = buildd@ release 20170724
  <no further output>

SLOF can cope if H_SET_DABR returns H_HARDWARE. So wwitch the return
value to H_HARDWARE instead of H_UNSUPPORTED so that we don't break
the guest boot.

That does mean we return a different error to PowerVM in this case,
but that's probably not a big concern.

Fixes: e8ebedbf ("KVM: PPC: Book3S HV: Return error from h_set_dabr() on POWER9")
Signed-off-by: NAneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Signed-off-by: NMichael Ellerman <mpe@ellerman.id.au>

The "lppaca" is a structure registered with the hypervisor. This is
unnecessary when running on non-virtualised platforms. One field from
the lppaca (pmcregs_in_use) is also used by the host, so move the host
part out into the paca (lppaca field is still updated in
guest mode).
Signed-off-by: NNicholas Piggin <npiggin@gmail.com>
[mpe: Fix non-pseries build with some #ifdefs]
Signed-off-by: NMichael Ellerman <mpe@ellerman.id.au>

POWER9 with the DAWR disabled causes problems for partition
migration. Either we have to fail the migration (since we lose the
DAWR) or we silently drop the DAWR and allow the migration to pass.

This patch does the latter and allows the migration to pass (at the
cost of silently losing the DAWR). This is not ideal but hopefully the
best overall solution. This approach has been acked by Paulus.

With this patch kvmppc_set_one_reg() will store the DAWR in the vcpu
but won't actually set it on POWER9 hardware.
Signed-off-by: NMichael Neuling <mikey@neuling.org>
Signed-off-by: NMichael Ellerman <mpe@ellerman.id.au>

POWER7 compat mode guests can use h_set_dabr on POWER9. POWER9 should
use the DAWR but since it's disabled there we can't.

This returns H_UNSUPPORTED on a h_set_dabr() on POWER9 where the DAWR
is disabled.

Current Linux guests ignore this error, so they will silently not get
the DAWR (sigh). The same error code is being used by POWERVM in this
case.
Signed-off-by: NMichael Neuling <mikey@neuling.org>
Signed-off-by: NMichael Ellerman <mpe@ellerman.id.au>

This works around a hardware bug in "Nimbus" POWER9 DD2.2 processors,
where the contents of the TEXASR can get corrupted while a thread is
in fake suspend state.  The workaround is for the instruction emulation
code to use the value saved at the most recent guest exit in real
suspend mode.  We achieve this by simply not saving the TEXASR into
the vcpu struct on an exit in fake suspend state.  We also have to
take care to set the orig_texasr field only on guest exit in real
suspend state.

This also means that on guest entry in fake suspend state, TEXASR
will be restored to the value it had on the last exit in real suspend
state, effectively counteracting any hardware-caused corruption.  This
works because TEXASR may not be written in suspend state.

With this, the guest might see the wrong values in TEXASR if it reads
it while in suspend state, but will see the correct value in
non-transactional state (e.g. after a treclaim), and treclaim will
work correctly.

With this workaround, the code will actually run slightly faster, and
will operate correctly on systems without the TEXASR bug (since TEXASR
may not be written in suspend state, and is only changed by failure
recording, which will have already been done before we get into fake
suspend state).  Therefore these changes are not made subject to a CPU
feature bit.
Signed-off-by: NPaul Mackerras <paulus@ozlabs.org>
Signed-off-by: NMichael Ellerman <mpe@ellerman.id.au>

This works around a hardware bug in "Nimbus" POWER9 DD2.2 processors,
where a treclaim performed in fake suspend mode can cause subsequent
reads from the XER register to return inconsistent values for the SO
(summary overflow) bit.  The inconsistent SO bit state can potentially
be observed on any thread in the core.  We have to do the treclaim
because that is the only way to get the thread out of suspend state
(fake or real) and into non-transactional state.

The workaround for the bug is to force the core into SMT4 mode before
doing the treclaim.  This patch adds the code to do that, conditional
on the CPU_FTR_P9_TM_XER_SO_BUG feature bit.
Signed-off-by: NSuraj Jitindar Singh <sjitindarsingh@gmail.com>
Signed-off-by: NPaul Mackerras <paulus@ozlabs.org>
Signed-off-by: NMichael Ellerman <mpe@ellerman.id.au>

POWER9 has hardware bugs relating to transactional memory and thread
reconfiguration (changes to hardware SMT mode).  Specifically, the core
does not have enough storage to store a complete checkpoint of all the
architected state for all four threads.  The DD2.2 version of POWER9
includes hardware modifications designed to allow hypervisor software
to implement workarounds for these problems.  This patch implements
those workarounds in KVM code so that KVM guests see a full, working
transactional memory implementation.

The problems center around the use of TM suspended state, where the
CPU has a checkpointed state but execution is not transactional.  The
workaround is to implement a "fake suspend" state, which looks to the
guest like suspended state but the CPU does not store a checkpoint.
In this state, any instruction that would cause a transition to
transactional state (rfid, rfebb, mtmsrd, tresume) or would use the
checkpointed state (treclaim) causes a "soft patch" interrupt (vector
0x1500) to the hypervisor so that it can be emulated.  The trechkpt
instruction also causes a soft patch interrupt.

On POWER9 DD2.2, we avoid returning to the guest in any state which
would require a checkpoint to be present.  The trechkpt in the guest
entry path which would normally create that checkpoint is replaced by
either a transition to fake suspend state, if the guest is in suspend
state, or a rollback to the pre-transactional state if the guest is in
transactional state.  Fake suspend state is indicated by a flag in the
PACA plus a new bit in the PSSCR.  The new PSSCR bit is write-only and
reads back as 0.

On exit from the guest, if the guest is in fake suspend state, we still
do the treclaim instruction as we would in real suspend state, in order
to get into non-transactional state, but we do not save the resulting
register state since there was no checkpoint.

Emulation of the instructions that cause a softpatch interrupt is
handled in two paths.  If the guest is in real suspend mode, we call
kvmhv_p9_tm_emulation_early() to handle the cases where the guest is
transitioning to transactional state.  This is called before we do the
treclaim in the guest exit path; because we haven't done treclaim, we
can get back to the guest with the transaction still active.  If the
instruction is a case that kvmhv_p9_tm_emulation_early() doesn't
handle, or if the guest is in fake suspend state, then we proceed to
do the complete guest exit path and subsequently call
kvmhv_p9_tm_emulation() in host context with the MMU on.  This handles
all the cases including the cases that generate program interrupts
(illegal instruction or TM Bad Thing) and facility unavailable
interrupts.

The emulation is reasonably straightforward and is mostly concerned
with checking for exception conditions and updating the state of
registers such as MSR and CR0.  The treclaim emulation takes care to
ensure that the TEXASR register gets updated as if it were the guest
treclaim instruction that had done failure recording, not the treclaim
done in hypervisor state in the guest exit path.

With this, the KVM_CAP_PPC_HTM capability returns true (1) even if
transactional memory is not available to host userspace.
Signed-off-by: NPaul Mackerras <paulus@ozlabs.org>
Signed-off-by: NMichael Ellerman <mpe@ellerman.id.au>

Since commit 6964e6a4 ("KVM: PPC: Book3S HV: Do SLB load/unload
with guest LPCR value loaded", 2018-01-11), we have been seeing
occasional machine check interrupts on POWER8 systems when running
KVM guests, due to SLB multihit errors.

This turns out to be due to the guest exit code reloading the host
SLB entries from the SLB shadow buffer when the SLB was not previously
cleared in the guest entry path.  This can happen because the path
which skips from the guest entry code to the guest exit code without
entering the guest now does the skip before the SLB is cleared and
loaded with guest values, but the host values are loaded after the
point in the guest exit path that we skip to.

To fix this, we move the code that reloads the host SLB values up
so that it occurs just before the point in the guest exit code (the
label guest_bypass:) where we skip to from the guest entry path.
Reported-by: NAlexey Kardashevskiy <aik@ozlabs.ru>
Fixes: 6964e6a4 ("KVM: PPC: Book3S HV: Do SLB load/unload with guest LPCR value loaded")
Tested-by: NAlexey Kardashevskiy <aik@ozlabs.ru>
Signed-off-by: NPaul Mackerras <paulus@ozlabs.org>

This fixes a bug where the trap number that is returned by
__kvmppc_vcore_entry gets corrupted.  The effect of the corruption
is that IPIs get ignored on POWER9 systems when the IPI is sent via
a doorbell interrupt to a CPU which is executing in a KVM guest.
The effect of the IPI being ignored is often that another CPU locks
up inside smp_call_function_many() (and if that CPU is holding a
spinlock, other CPUs then lock up inside raw_spin_lock()).

The trap number is currently held in register r12 for most of the
assembly-language part of the guest exit path.  In that path, we
call kvmppc_subcore_exit_guest(), which is a C function, without
restoring r12 afterwards.  Depending on the kernel config and the
compiler, it may modify r12 or it may not, so some config/compiler
combinations see the bug and others don't.

To fix this, we arrange for the trap number to be stored on the
stack from the 'guest_bypass:' label until the end of the function,
then the trap number is loaded and returned in r12 as before.

Cc: stable@vger.kernel.org # v4.8+
Fixes: fd7bacbc ("KVM: PPC: Book3S HV: Fix TB corruption in guest exit path on HMI interrupt")
Signed-off-by: NPaul Mackerras <paulus@ozlabs.org>

We ended up with code that did a conditional branch inside a feature
section to code outside of the feature section. Depending on how the
object file gets organized, that might mean we exceed the 14bit
relocation limit for conditional branches:

  arch/powerpc/kvm/built-in.o:arch/powerpc/kvm/book3s_hv_rmhandlers.S:416:(__ftr_alt_97+0x8): relocation truncated to fit: R_PPC64_REL14 against `.text'+1ca4

So instead of doing a conditional branch outside of the feature section,
let's just jump at the end of the same, making the branch very short.
Signed-off-by: NAlexander Graf <agraf@suse.de>
Signed-off-by: NPaul Mackerras <paulus@ozlabs.org>

Rename the paca->soft_enabled to paca->irq_soft_mask as it is no
longer used as a flag for interrupt state, but a mask.
Signed-off-by: NMadhavan Srinivasan <maddy@linux.vnet.ibm.com>
Signed-off-by: NMichael Ellerman <mpe@ellerman.id.au>

This works on top of the single escalation support. When in single
escalation, with this change, we will keep the escalation interrupt
disabled unless the VCPU is in H_CEDE (idle). In any other case, we
know the VCPU will be rescheduled and thus there is no need to take
escalation interrupts in the host whenever a guest interrupt fires.
Signed-off-by: NBenjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: NPaul Mackerras <paulus@ozlabs.org>

Signed-off-by: NBenjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: NPaul Mackerras <paulus@ozlabs.org>

Signed-off-by: NBenjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: NPaul Mackerras <paulus@ozlabs.org>

The prodded flag is only cleared at the beginning of H_CEDE,
so every time we have an escalation, we will cause the *next*
H_CEDE to return immediately.

Instead use a dedicated "irq_pending" flag to indicate that
a guest interrupt is pending for the VCPU. We don't reuse the
existing exception bitmap so as to avoid expensive atomic ops.
Signed-off-by: NBenjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: NPaul Mackerras <paulus@ozlabs.org>

Hypervisor maintenance interrupts (HMIs) are generated by various
causes, signalled by bits in the hypervisor maintenance exception
register (HMER).  In most cases calling OPAL to handle the interrupt
is the correct thing to do, but the "debug trigger" HMIs signalled by
PPC bit 17 (bit 46) of HMER are used to invoke software workarounds
for hardware bugs, and OPAL does not have any code to handle this
cause.  The debug trigger HMI is used in POWER9 DD2.0 and DD2.1 chips
to work around a hardware bug in executing vector load instructions to
cache inhibited memory.  In POWER9 DD2.2 chips, it is generated when
conditions are detected relating to threads being in TM (transactional
memory) suspended mode when the core SMT configuration needs to be
reconfigured.

The kernel currently has code to detect the vector CI load condition,
but only when the HMI occurs in the host, not when it occurs in a
guest.  If a HMI occurs in the guest, it is always passed to OPAL, and
then we always re-sync the timebase, because the HMI cause might have
been a timebase error, for which OPAL would re-sync the timebase, thus
removing the timebase offset which KVM applied for the guest.  Since
we don't know what OPAL did, we don't know whether to subtract the
timebase offset from the timebase, so instead we re-sync the timebase.

This adds code to determine explicitly what the cause of a debug
trigger HMI will be.  This is based on a new device-tree property
under the CPU nodes called ibm,hmi-special-triggers, if it is
present, or otherwise based on the PVR (processor version register).
The handling of debug trigger HMIs is pulled out into a separate
function which can be called from the KVM guest exit code.  If this
function handles and clears the HMI, and no other HMI causes remain,
then we skip calling OPAL and we proceed to subtract the guest
timebase offset from the timebase.

The overall handling for HMIs that occur in the host (i.e. not in a
KVM guest) is largely unchanged, except that we now don't set the flag
for the vector CI load workaround on DD2.2 processors.

This also removes a BUG_ON in the KVM code.  BUG_ON is generally not
useful in KVM guest entry/exit code since it is difficult to handle
the resulting trap gracefully.
Signed-off-by: NPaul Mackerras <paulus@ozlabs.org>
Signed-off-by: NMichael Ellerman <mpe@ellerman.id.au>

This moves the code that loads and unloads the guest SLB values so that
it is done while the guest LPCR value is loaded in the LPCR register.
The reason for doing this is that on POWER9, the behaviour of the
slbmte instruction depends on the LPCR[UPRT] bit.  If UPRT is 1, as
it is for a radix host (or guest), the SLB index is truncated to
2 bits.  This means that for a HPT guest on a radix host, the SLB
was not being loaded correctly, causing the guest to crash.

The SLB is now loaded much later in the guest entry path, after the
LPCR is loaded, which for a secondary thread is after it sees that
the primary thread has switched the MMU to the guest.  The loop that
waits for the primary thread has a branch out to the exit code that
is taken if it sees that other threads have commenced exiting the
guest.  Since we have now not loaded the SLB at this point, we make
this path branch to a new label 'guest_bypass' and we move the SLB
unload code to before this label.
Signed-off-by: NPaul Mackerras <paulus@ozlabs.org>

This fixes a bug where it is possible to enter a guest on a POWER9
system without having the XIVE (interrupt controller) context loaded.
This can happen because we unload the XIVE context from the CPU
before doing the real-mode handling for machine checks.  After the
real-mode handler runs, it is possible that we re-enter the guest
via a fast path which does not load the XIVE context.

To fix this, we move the unloading of the XIVE context to come after
the real-mode machine check handler is called.

Fixes: 5af50993 ("KVM: PPC: Book3S HV: Native usage of the XIVE interrupt controller")
Cc: stable@vger.kernel.org # v4.11+
Signed-off-by: NPaul Mackerras <paulus@ozlabs.org>

On Book3S in HV mode, we don't use the vcpu->arch.dec field at all.
Instead, all logic is built around vcpu->arch.dec_expires.

So let's remove the one remaining piece of code that was setting it.
Signed-off-by: NAlexander Graf <agraf@suse.de>
Signed-off-by: NPaul Mackerras <paulus@ozlabs.org>

This commit does simple conversions of rfi/rfid to the new macros that
include the expected destination context. By simple we mean cases
where there is a single well known destination context, and it's
simply a matter of substituting the instruction for the appropriate
macro.
Signed-off-by: NNicholas Piggin <npiggin@gmail.com>
Signed-off-by: NMichael Ellerman <mpe@ellerman.id.au>

This patch removes the restriction that a radix host can only run
radix guests, allowing us to run HPT (hashed page table) guests as
well.  This is useful because it provides a way to run old guest
kernels that know about POWER8 but not POWER9.

Unfortunately, POWER9 currently has a restriction that all threads
in a given code must either all be in HPT mode, or all in radix mode.
This means that when entering a HPT guest, we have to obtain control
of all 4 threads in the core and get them to switch their LPIDR and
LPCR registers, even if they are not going to run a guest.  On guest
exit we also have to get all threads to switch LPIDR and LPCR back
to host values.

To make this feasible, we require that KVM not be in the "independent
threads" mode, and that the CPU cores be in single-threaded mode from
the host kernel's perspective (only thread 0 online; threads 1, 2 and
3 offline).  That allows us to use the same code as on POWER8 for
obtaining control of the secondary threads.

To manage the LPCR/LPIDR changes required, we extend the kvm_split_info
struct to contain the information needed by the secondary threads.
All threads perform a barrier synchronization (where all threads wait
for every other thread to reach the synchronization point) on guest
entry, both before and after loading LPCR and LPIDR.  On guest exit,
they all once again perform a barrier synchronization both before
and after loading host values into LPCR and LPIDR.

Finally, it is also currently necessary to flush the entire TLB every
time we enter a HPT guest on a radix host.  We do this on thread 0
with a loop of tlbiel instructions.
Signed-off-by: NPaul Mackerras <paulus@ozlabs.org>

This patch allows for a mode on POWER9 hosts where we control all the
threads of a core, much as we do on POWER8.  The mode is controlled by
a module parameter on the kvm_hv module, called "indep_threads_mode".
The normal mode on POWER9 is the "independent threads" mode, with
indep_threads_mode=Y, where the host is in SMT4 mode (or in fact any
desired SMT mode) and each thread independently enters and exits from
KVM guests without reference to what other threads in the core are
doing.

If indep_threads_mode is set to N at the point when a VM is started,
KVM will expect every core that the guest runs on to be in single
threaded mode (that is, threads 1, 2 and 3 offline), and will set the
flag that prevents secondary threads from coming online.  We can still
use all four threads; the code that implements dynamic micro-threading
on POWER8 will become active in over-commit situations and will allow
up to three other VCPUs to be run on the secondary threads of the core
whenever a VCPU is run.

The reason for wanting this mode is that this will allow us to run HPT
guests on a radix host on a POWER9 machine that does not support
"mixed mode", that is, having some threads in a core be in HPT mode
while other threads are in radix mode.  It will also make it possible
to implement a "strict threads" mode in future, if desired.
Signed-off-by: NPaul Mackerras <paulus@ozlabs.org>

This reverts commit 94a04bc2.

In order to run HPT guests on a radix POWER9 host, we will have to run
the host in single-threaded mode, because POWER9 processors do not
currently support running some threads of a core in HPT mode while
others are in radix mode ("mixed mode").

That means that we will need the same mechanisms that are used on
POWER8 to make the secondary threads available to KVM, which were
disabled on POWER9 by commit 94a04bc2.
Signed-off-by: NPaul Mackerras <paulus@ozlabs.org>
Signed-off-by: NMichael Ellerman <mpe@ellerman.id.au>

On POWER9 systems, we push the VCPU context onto the XIVE (eXternal
Interrupt Virtualization Engine) hardware when entering a guest,
and pull the context off the XIVE when exiting the guest.  The push
is done with cache-inhibited stores, and the pull with cache-inhibited
loads.

Testing has revealed that it is possible (though very rare) for
the stores to get reordered with the loads so that we end up with the
guest VCPU context still loaded on the XIVE after we have exited the
guest.  When that happens, it is possible for the same VCPU context
to then get loaded on another CPU, which causes the machine to
checkstop.

To fix this, we add I/O barrier instructions (eieio) before and
after the push and pull operations.  As partial compensation for the
potential slowdown caused by the extra barriers, we remove the eieio
instructions between the two stores in the push operation, and between
the two loads in the pull operation.  (The architecture requires
loads to cache-inhibited, guarded storage to be kept in order, and
requires stores to cache-inhibited, guarded storage likewise to be
kept in order, but allows such loads and stores to be reordered with
respect to each other.)
Reported-by: NCarol L Soto <clsoto@us.ibm.com>
Signed-off-by: NPaul Mackerras <paulus@ozlabs.org>

At present, if an interrupt (i.e. an exception or trap) occurs in the
code where KVM is switching the MMU to or from guest context, we jump
to kvmppc_bad_host_intr, where we simply spin with interrupts disabled.
In this situation, it is hard to debug what happened because we get no
indication as to which interrupt occurred or where.  Typically we get
a cascade of stall and soft lockup warnings from other CPUs.

In order to get more information for debugging, this adds code to
create a stack frame on the emergency stack and save register values
to it.  We start half-way down the emergency stack in order to give
ourselves some chance of being able to do a stack trace on secondary
threads that are already on the emergency stack.

On POWER7 or POWER8, we then just spin, as before, because we don't
know what state the MMU context is in or what other threads are doing,
and we can't switch back to host context without coordinating with
other threads.  On POWER9 we can do better; there we load up the host
MMU context and jump to C code, which prints an oops message to the
console and panics.
Signed-off-by: NPaul Mackerras <paulus@ozlabs.org>

- Add another case where msgsync is required.
- Required barrier sequence for global doorbells is msgsync ; lwsync

When msgsnd is used for IPIs to other cores, msgsync must be executed by
the target to order stores performed on the source before its msgsnd
(provided the source executes the appropriate sync).

Fixes: 1704a81c ("KVM: PPC: Book3S HV: Use msgsnd for IPIs to other cores on POWER9")
Cc: stable@vger.kernel.org # v4.10+
Signed-off-by: NNicholas Piggin <npiggin@gmail.com>
Signed-off-by: NPaul Mackerras <paulus@ozlabs.org>

On POWER9 DD2.1 and below, sometimes on a Hypervisor Data Storage
Interrupt (HDSI) the HDSISR is not be updated at all.

To work around this we put a canary value into the HDSISR before
returning to a guest and then check for this canary when we take a
HDSI. If we find the canary on a HDSI, we know the hardware didn't
update the HDSISR. In this case we return to the guest to retake the
HDSI which should correctly update the HDSISR the second time HDSI
entry.

After talking to Paulus we've applied this workaround to all POWER9
CPUs. The workaround of returning to the guest shouldn't ever be
triggered on well behaving CPU. The extra instructions should have
negligible performance impact.
Signed-off-by: NMichael Neuling <mikey@neuling.org>
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>

Aneesh Kumar reported seeing host crashes when running recent kernels
on POWER8.  The symptom was an oops like this:

Unable to handle kernel paging request for data at address 0xf00000000786c620
Faulting instruction address: 0xc00000000030e1e4
Oops: Kernel access of bad area, sig: 11 [#1]
LE SMP NR_CPUS=2048 NUMA PowerNV
Modules linked in: powernv_op_panel
CPU: 24 PID: 6663 Comm: qemu-system-ppc Tainted: G        W 4.13.0-rc7-43932-gfc36c59 #2
task: c000000fdeadfe80 task.stack: c000000fdeb68000
NIP:  c00000000030e1e4 LR: c00000000030de6c CTR: c000000000103620
REGS: c000000fdeb6b450 TRAP: 0300   Tainted: G        W        (4.13.0-rc7-43932-gfc36c59)
MSR:  9000000000009033 <SF,HV,EE,ME,IR,DR,RI,LE>  CR: 24044428  XER: 20000000
CFAR: c00000000030e134 DAR: f00000000786c620 DSISR: 40000000 SOFTE: 0
GPR00: 0000000000000000 c000000fdeb6b6d0 c0000000010bd000 000000000000e1b0
GPR04: c00000000115e168 c000001fffa6e4b0 c00000000115d000 c000001e1b180386
GPR08: f000000000000000 c000000f9a8913e0 f00000000786c600 00007fff587d0000
GPR12: c000000fdeb68000 c00000000fb0f000 0000000000000001 00007fff587cffff
GPR16: 0000000000000000 c000000000000000 00000000003fffff c000000fdebfe1f8
GPR20: 0000000000000004 c000000fdeb6b8a8 0000000000000001 0008000000000040
GPR24: 07000000000000c0 00007fff587cffff c000000fdec20bf8 00007fff587d0000
GPR28: c000000fdeca9ac0 00007fff587d0000 00007fff587c0000 00007fff587d0000
NIP [c00000000030e1e4] __get_user_pages_fast+0x434/0x1070
LR [c00000000030de6c] __get_user_pages_fast+0xbc/0x1070
Call Trace:
[c000000fdeb6b6d0] [c00000000139dab8] lock_classes+0x0/0x35fe50 (unreliable)
[c000000fdeb6b7e0] [c00000000030ef38] get_user_pages_fast+0xf8/0x120
[c000000fdeb6b830] [c000000000112318] kvmppc_book3s_hv_page_fault+0x308/0xf30
[c000000fdeb6b960] [c00000000010e10c] kvmppc_vcpu_run_hv+0xfdc/0x1f00
[c000000fdeb6bb20] [c0000000000e915c] kvmppc_vcpu_run+0x2c/0x40
[c000000fdeb6bb40] [c0000000000e5650] kvm_arch_vcpu_ioctl_run+0x110/0x300
[c000000fdeb6bbe0] [c0000000000d6468] kvm_vcpu_ioctl+0x528/0x900
[c000000fdeb6bd40] [c0000000003bc04c] do_vfs_ioctl+0xcc/0x950
[c000000fdeb6bde0] [c0000000003bc930] SyS_ioctl+0x60/0x100
[c000000fdeb6be30] [c00000000000b96c] system_call+0x58/0x6c
Instruction dump:
7ca81a14 2fa50000 41de0010 7cc8182a 68c60002 78c6ffe2 0b060000 3cc2000a
794a3664 390610d8 e9080000 7d485214 <e90a0020> 7d435378 790507e1 408202f0
---[ end trace fad4a342d0414aa2 ]---

It turns out that what has happened is that the SLB entry for the
vmmemap region hasn't been reloaded on exit from a guest, and it has
the wrong page size.  Then, when the host next accesses the vmemmap
region, it gets a page fault.

Commit a25bd72b ("powerpc/mm/radix: Workaround prefetch issue with
KVM", 2017-07-24) modified the guest exit code so that it now only clears
out the SLB for hash guest.  The code tests the radix flag and puts the
result in a non-volatile CR field, CR2, and later branches based on CR2.

Unfortunately, the kvmppc_save_tm function, which gets called between
those two points, modifies all the user-visible registers in the case
where the guest was in transactional or suspended state, except for a
few which it restores (namely r1, r2, r9 and r13).  Thus the hash/radix indication in CR2 gets corrupted.

This fixes the problem by re-doing the comparison just before the
result is needed.  For good measure, this also adds comments next to
the call sites of kvmppc_save_tm and kvmppc_restore_tm pointing out
that non-volatile register state will be lost.

Cc: stable@vger.kernel.org # v4.13
Fixes: a25bd72b ("powerpc/mm/radix: Workaround prefetch issue with KVM")
Tested-by: NAneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Signed-off-by: NPaul Mackerras <paulus@ozlabs.org>

Commit 2f272463 ("KVM: PPC: Book3S HV: Cope with host using large
decrementer mode", 2017-05-22) added code to treat the hypervisor
decrementer (HDEC) as a 64-bit value on POWER9 rather than 32-bit.
Unfortunately, that commit missed one place where HDEC is treated
as a 32-bit value.  This fixes it.

This bug should not have any user-visible consequences that I can
think of, beyond an occasional unnecessary exit to the host kernel.
If the hypervisor decrementer has gone negative, then the bottom
32 bits will be negative for about 4 seconds after that, so as
long as we get out of the guest within those 4 seconds we won't
conclude that the HDEC interrupt is spurious.
Reported-by: NSuraj Jitindar Singh <sjitindarsingh@gmail.com>
Fixes: 2f272463 ("KVM: PPC: Book3S HV: Cope with host using large decrementer mode")
Signed-off-by: NPaul Mackerras <paulus@ozlabs.org>

binutils >= 2.26 now warns about misuse of register expressions in
assembler operands that are actually literals. In this instance r0 is
being used where a literal 0 should be used.
Signed-off-by: NAndreas Schwab <schwab@linux-m68k.org>
[mpe: Split into separate KVM patch, tweak change log]
Signed-off-by: NMichael Ellerman <mpe@ellerman.id.au>
Signed-off-by: NPaul Mackerras <paulus@ozlabs.org>

POWER9 CPUs have independent MMU contexts per thread, so KVM does not
need to quiesce secondary threads, so the hwthread_req/hwthread_state
protocol does not have to be used. So patch it away on POWER9, and patch
away the branch from the Linux idle wakeup to kvm_start_guest that is
never used.

Add a warning and error out of kvmppc_grab_hwthread in case it is ever
called on POWER9.

This avoids a hwsync in the idle wakeup path on POWER9.
Signed-off-by: NNicholas Piggin <npiggin@gmail.com>
Acked-by: NPaul Mackerras <paulus@ozlabs.org>
[mpe: Use WARN(...) instead of WARN_ON()/pr_err(...)]
Signed-off-by: NMichael Ellerman <mpe@ellerman.id.au>

When msgsnd is used for IPIs to other cores, msgsync must be executed by
the target to order stores performed on the source before its msgsnd
(provided the source executes the appropriate sync).

Fixes: 1704a81c ("KVM: PPC: Book3S HV: Use msgsnd for IPIs to other cores on POWER9")
Signed-off-by: NNicholas Piggin <npiggin@gmail.com>
Signed-off-by: NPaul Mackerras <paulus@ozlabs.org>

There's a somewhat architectural issue with Radix MMU and KVM.

When coming out of a guest with AIL (Alternate Interrupt Location, ie,
MMU enabled), we start executing hypervisor code with the PID register
still containing whatever the guest has been using.

The problem is that the CPU can (and will) then start prefetching or
speculatively load from whatever host context has that same PID (if
any), thus bringing translations for that context into the TLB, which
Linux doesn't know about.

This can cause stale translations and subsequent crashes.

Fixing this in a way that is neither racy nor a huge performance
impact is difficult. We could just make the host invalidations always
use broadcast forms but that would hurt single threaded programs for
example.

We chose to fix it instead by partitioning the PID space between guest
and host. This is possible because today Linux only use 19 out of the
20 bits of PID space, so existing guests will work if we make the host
use the top half of the 20 bits space.

We additionally add support for a property to indicate to Linux the
size of the PID register which will be useful if we eventually have
processors with a larger PID space available.

There is still an issue with malicious guests purposefully setting the
PID register to a value in the hosts PID range. Hopefully future HW
can prevent that, but in the meantime, we handle it with a pair of
kludges:

 - On the way out of a guest, before we clear the current VCPU in the
   PACA, we check the PID and if it's outside of the permitted range
   we flush the TLB for that PID.

 - When context switching, if the mm is "new" on that CPU (the
   corresponding bit was set for the first time in the mm cpumask), we
   check if any sibling thread is in KVM (has a non-NULL VCPU pointer
   in the PACA). If that is the case, we also flush the PID for that
   CPU (core).

This second part is needed to handle the case where a process is
migrated (or starts a new pthread) on a sibling thread of the CPU
coming out of KVM, as there's a window where stale translations can
exist before we detect it and flush them out.

A future optimization could be added by keeping track of whether the
PID has ever been used and avoid doing that for completely fresh PIDs.
We could similarily mark PIDs that have been the subject of a global
invalidation as "fresh". But for now this will do.
Signed-off-by: NBenjamin Herrenschmidt <benh@kernel.crashing.org>
[mpe: Rework the asm to build with CONFIG_PPC_RADIX_MMU=n, drop
      unneeded include of kvm_book3s_asm.h]
Signed-off-by: NMichael Ellerman <mpe@ellerman.id.au>

At present, interrupts are hard-disabled fairly late in the guest
entry path, in the assembly code.  Since we check for pending signals
for the vCPU(s) task(s) earlier in the guest entry path, it is
possible for a signal to be delivered before we enter the guest but
not be noticed until after we exit the guest for some other reason.

Similarly, it is possible for the scheduler to request a reschedule
while we are in the guest entry path, and we won't notice until after
we have run the guest, potentially for a whole timeslice.

Furthermore, with a radix guest on POWER9, we can take the interrupt
with the MMU on.  In this case we end up leaving interrupts
hard-disabled after the guest exit, and they are likely to stay
hard-disabled until we exit to userspace or context-switch to
another process.  This was masking the fact that we were also not
setting the RI (recoverable interrupt) bit in the MSR, meaning
that if we had taken an interrupt, it would have crashed the host
kernel with an unrecoverable interrupt message.

To close these races, we need to check for signals and reschedule
requests after hard-disabling interrupts, and then keep interrupts
hard-disabled until we enter the guest.  If there is a signal or a
reschedule request from another CPU, it will send an IPI, which will
cause a guest exit.

This puts the interrupt disabling before we call kvmppc_start_thread()
for all the secondary threads of this core that are going to run vCPUs.
The reason for that is that once we have started the secondary threads
there is no easy way to back out without going through at least part
of the guest entry path.  However, kvmppc_start_thread() includes some
code for radix guests which needs to call smp_call_function(), which
must be called with interrupts enabled.  To solve this problem, this
patch moves that code into a separate function that is called earlier.

When the guest exit is caused by an external interrupt, a hypervisor
doorbell or a hypervisor maintenance interrupt, we now handle these
using the replay facility.  __kvmppc_vcore_entry() now returns the
trap number that caused the exit on this thread, and instead of the
assembly code jumping to the handler entry, we return to C code with
interrupts still hard-disabled and set the irq_happened flag in the
PACA, so that when we do local_irq_enable() the appropriate handler
gets called.

With all this, we now have the interrupt soft-enable flag clear while
we are in the guest.  This is useful because code in the real-mode
hypercall handlers that checks whether interrupts are enabled will
now see that they are disabled, which is correct, since interrupts
are hard-disabled in the real-mode code.
Signed-off-by: NPaul Mackerras <paulus@ozlabs.org>

Enhance KVM to cause a guest exit with KVM_EXIT_NMI
exit reason upon a machine check exception (MCE) in
the guest address space if the KVM_CAP_PPC_FWNMI
capability is enabled (instead of delivering a 0x200
interrupt to guest). This enables QEMU to build error
log and deliver machine check exception to guest via
guest registered machine check handler.

This approach simplifies the delivery of machine
check exception to guest OS compared to the earlier
approach of KVM directly invoking 0x200 guest interrupt
vector.

This design/approach is based on the feedback for the
QEMU patches to handle machine check exception. Details
of earlier approach of handling machine check exception
in QEMU and related discussions can be found at:

https://lists.nongnu.org/archive/html/qemu-devel/2014-11/msg00813.html

Note:

This patch now directly invokes machine_check_print_event_info()
from kvmppc_handle_exit_hv() to print the event to host console
at the time of guest exit before the exception is passed on to the
guest. Hence, the host-side handling which was performed earlier
via machine_check_fwnmi is removed.

The reasons for this approach is (i) it is not possible
to distinguish whether the exception occurred in the
guest or the host from the pt_regs passed on the
machine_check_exception(). Hence machine_check_exception()
calls panic, instead of passing on the exception to
the guest, if the machine check exception is not
recoverable. (ii) the approach introduced in this
patch gives opportunity to the host kernel to perform
actions in virtual mode before passing on the exception
to the guest. This approach does not require complex
tweaks to machine_check_fwnmi and friends.
Signed-off-by: NAravinda Prasad <aravinda@linux.vnet.ibm.com>
Reviewed-by: NDavid Gibson <david@gibson.dropbear.id.au>
Signed-off-by: NMahesh Salgaonkar <mahesh@linux.vnet.ibm.com>
Signed-off-by: NPaul Mackerras <paulus@ozlabs.org>

Idle code now always runs at the 0xc... effective address whether
in real or virtual mode. This means rfid can be ditched, along
with a lot of SRR manipulations.

In the wakeup path, carry SRR1 around in r12. Use mtmsrd to change
MSR states as required.

This also balances the return prediction for the idle call, by
doing blr rather than rfid to return to the idle caller.

On POWER9, 2-process context switch on different cores, with snooze
disabled, increases performance by 2%.
Signed-off-by: NNicholas Piggin <npiggin@gmail.com>
[mpe: Incorporate v2 fixes from Nick]
Signed-off-by: NMichael Ellerman <mpe@ellerman.id.au>

On POWER9, we no longer have the restriction that we had on POWER8
where all threads in a core have to be in the same partition, so
the CPU threads are now independent.  However, we still want to be
able to run guests with a virtual SMT topology, if only to allow
migration of guests from POWER8 systems to POWER9.

A guest that has a virtual SMT mode greater than 1 will expect to
be able to use the doorbell facility; it will expect the msgsndp
and msgclrp instructions to work appropriately and to be able to read
sensible values from the TIR (thread identification register) and
DPDES (directed privileged doorbell exception status) special-purpose
registers.  However, since each CPU thread is a separate sub-processor
in POWER9, these instructions and registers can only be used within
a single CPU thread.

In order for these instructions to appear to act correctly according
to the guest's virtual SMT mode, we have to trap and emulate them.
We cause them to trap by clearing the HFSCR_MSGP bit in the HFSCR
register.  The emulation is triggered by the hypervisor facility
unavailable interrupt that occurs when the guest uses them.

To cause a doorbell interrupt to occur within the guest, we set the
DPDES register to 1.  If the guest has interrupts enabled, the CPU
will generate a doorbell interrupt and clear the DPDES register in
hardware.  The DPDES hardware register for the guest is saved in the
vcpu->arch.vcore->dpdes field.  Since this gets written by the guest
exit code, other VCPUs wishing to cause a doorbell interrupt don't
write that field directly, but instead set a vcpu->arch.doorbell_request
flag.  This is consumed and set to 0 by the guest entry code, which
then sets DPDES to 1.

Emulating reads of the DPDES register is somewhat involved, because
it requires reading the doorbell pending interrupt status of all of the
VCPU threads in the virtual core, and if any of those VCPUs are
running, their doorbell status is only up-to-date in the hardware
DPDES registers of the CPUs where they are running.  In order to get
a reasonable approximation of the current doorbell status, we send
those CPUs an IPI, causing an exit from the guest which will update
the vcpu->arch.vcore->dpdes field.  We then use that value in
constructing the emulated DPDES register value.
Signed-off-by: NPaul Mackerras <paulus@ozlabs.org>

This adds code to allow us to use a different value for the HFSCR
(Hypervisor Facilities Status and Control Register) when running the
guest from that which applies in the host.  The reason for doing this
is to allow us to trap the msgsndp instruction and related operations
in future so that they can be virtualized.  We also save the value of
HFSCR when a hypervisor facility unavailable interrupt occurs, because
the high byte of HFSCR indicates which facility the guest attempted to
access.

We save and restore the host value on guest entry/exit because some
bits of it affect host userspace execution.

We only do all this on POWER9, not on POWER8, because we are not
intending to virtualize any of the facilities controlled by HFSCR on
POWER8.  In particular, the HFSCR bit that controls execution of
msgsndp and related operations does not exist on POWER8.  The HFSCR
doesn't exist at all on POWER7.
Signed-off-by: NPaul Mackerras <paulus@ozlabs.org>

This allows userspace (e.g. QEMU) to enable large decrementer mode for
the guest when running on a POWER9 host, by setting the LPCR_LD bit in
the guest LPCR value.  With this, the guest exit code saves 64 bits of
the guest DEC value on exit.  Other places that use the guest DEC
value check the LPCR_LD bit in the guest LPCR value, and if it is set,
omit the 32-bit sign extension that would otherwise be done.

This doesn't change the DEC emulation used by PR KVM because PR KVM
is not supported on POWER9 yet.

This is partly based on an earlier patch by Oliver O'Halloran.
Signed-off-by: NPaul Mackerras <paulus@ozlabs.org>