This fixes various issues in how we were handling the VSX registers
that exist on POWER7 machines. First, we were running off the end
of the current->thread.fpr[] array. Ultimately this was because the
vcpu->arch.vsr[] array is sized to be able to store both the FP
registers and the extra VSX registers (i.e. 64 entries), but PR KVM
only uses it for the extra VSX registers (i.e. 32 entries).
Secondly, calling load_up_vsx() from C code is a really bad idea,
because it jumps to fast_exception_return at the end, rather than
returning with a blr instruction. This was causing it to jump off
to a random location with random register contents, since it was using
the largely uninitialized stack frame created by kvmppc_load_up_vsx.
In fact, it isn't necessary to call either __giveup_vsx or load_up_vsx,
since giveup_fpu and load_up_fpu handle the extra VSX registers as well
as the standard FP registers on machines with VSX. Also, since VSX
instructions can access the VMX registers and the FP registers as well
as the extra VSX registers, we have to load up the FP and VMX registers
before we can turn on the MSR_VSX bit for the guest. Conversely, if
we save away any of the VSX or FP registers, we have to turn off MSR_VSX
for the guest.
To handle all this, it is more convenient for a single call to
kvmppc_giveup_ext() to handle all the state saving that needs to be done,
so we make it take a set of MSR bits rather than just one, and the switch
statement becomes a series of if statements. Similarly kvmppc_handle_ext
needs to be able to load up more than one set of registers.
Signed-off-by: Paul Mackerras <paulus@samba.org>
Signed-off-by: Alexander Graf <agraf@suse.de>
This adds basic emulation of the PURR and SPURR registers. We assume
we are emulating a single-threaded core, so these advance at the same
rate as the timebase. A Linux kernel running on a POWER7 expects to
be able to access these registers and is not prepared to handle a
program interrupt on accessing them.
This also adds a very minimal emulation of the DSCR (data stream
control register). Writes are ignored and reads return zero.
Signed-off-by: Paul Mackerras <paulus@samba.org>
Signed-off-by: Alexander Graf <agraf@suse.de>
Currently, if the guest does an H_PROTECT hcall requesting that the
permissions on a HPT entry be changed to allow writing, we make the
requested change even if the page is marked read-only in the host
Linux page tables. This is a problem since it would for instance
allow a guest to modify a page that KSM has decided can be shared
between multiple guests.
To fix this, if the new permissions for the page allow writing, we need
to look up the memslot for the page, work out the host virtual address,
and look up the Linux page tables to get the PTE for the page. If that
PTE is read-only, we reduce the HPTE permissions to read-only.
Signed-off-by: Paul Mackerras <paulus@samba.org>
Signed-off-by: Alexander Graf <agraf@suse.de>
This fixes a bug in the code which allows userspace to read out the
contents of the guest's hashed page table (HPT). On the second and
subsequent passes through the HPT, when we are reporting only those
entries that have changed, we were incorrectly initializing the index
field of the header with the index of the first entry we skipped
rather than the first changed entry. This fixes it.
Signed-off-by: Paul Mackerras <paulus@samba.org>
Signed-off-by: Alexander Graf <agraf@suse.de>
With HV-style KVM, we maintain reverse-mapping lists that enable us to
find all the HPT (hashed page table) entries that reference each guest
physical page, with the heads of the lists in the memslot->arch.rmap
arrays. When we reset the HPT (i.e. when we reboot the VM), we clear
out all the HPT entries but we were not clearing out the reverse
mapping lists. The result is that as we create new HPT entries, the
lists get corrupted, which can easily lead to loops, resulting in the
host kernel hanging when it tries to traverse those lists.
This fixes the problem by zeroing out all the reverse mapping lists
when we zero out the HPT. This incidentally means that we are also
zeroing our record of the referenced and changed bits (not the bits
in the Linux PTEs, used by the Linux MM subsystem, but the bits used
by the KVM_GET_DIRTY_LOG ioctl, and those used by kvm_age_hva() and
kvm_test_age_hva()).
Signed-off-by: Paul Mackerras <paulus@samba.org>
Signed-off-by: Alexander Graf <agraf@suse.de>
A new ioctl, KVM_PPC_GET_HTAB_FD, returns a file descriptor. Reads on
this fd return the contents of the HPT (hashed page table), writes
create and/or remove entries in the HPT. There is a new capability,
KVM_CAP_PPC_HTAB_FD, to indicate the presence of the ioctl. The ioctl
takes an argument structure with the index of the first HPT entry to
read out and a set of flags. The flags indicate whether the user is
intending to read or write the HPT, and whether to return all entries
or only the "bolted" entries (those with the bolted bit, 0x10, set in
the first doubleword).
This is intended for use in implementing qemu's savevm/loadvm and for
live migration. Therefore, on reads, the first pass returns information
about all HPTEs (or all bolted HPTEs). When the first pass reaches the
end of the HPT, it returns from the read. Subsequent reads only return
information about HPTEs that have changed since they were last read.
A read that finds no changed HPTEs in the HPT following where the last
read finished will return 0 bytes.
The format of the data provides a simple run-length compression of the
invalid entries. Each block of data starts with a header that indicates
the index (position in the HPT, which is just an array), the number of
valid entries starting at that index (may be zero), and the number of
invalid entries following those valid entries. The valid entries, 16
bytes each, follow the header. The invalid entries are not explicitly
represented.
Signed-off-by: Paul Mackerras <paulus@samba.org>
[agraf: fix documentation]
Signed-off-by: Alexander Graf <agraf@suse.de>
This makes a HPTE removal function, kvmppc_do_h_remove(), available
outside book3s_hv_rm_mmu.c. This will be used by the HPT writing
code.
Signed-off-by: Paul Mackerras <paulus@samba.org>
Signed-off-by: Alexander Graf <agraf@suse.de>
This uses a bit in our record of the guest view of the HPTE to record
when the HPTE gets modified. We use a reserved bit for this, and ensure
that this bit is always cleared in HPTE values returned to the guest.
The recording of modified HPTEs is only done if other code indicates
its interest by setting kvm->arch.hpte_mod_interest to a non-zero value.
The reason for this is that when later commits add facilities for
userspace to read the HPT, the first pass of reading the HPT will be
quicker if there are no (or very few) HPTEs marked as modified,
rather than having most HPTEs marked as modified.
Signed-off-by: Paul Mackerras <paulus@samba.org>
Signed-off-by: Alexander Graf <agraf@suse.de>
This fixes a bug where adding a new guest HPT entry via the H_ENTER
hcall would lose the "changed" bit in the reverse map information
for the guest physical page being mapped. The result was that the
KVM_GET_DIRTY_LOG could return a zero bit for the page even though
the page had been modified by the guest.
This fixes it by only modifying the index and present bits in the
reverse map entry, thus preserving the reference and change bits.
We were also unnecessarily setting the reference bit, and this
fixes that too.
Signed-off-by: Paul Mackerras <paulus@samba.org>
Signed-off-by: Alexander Graf <agraf@suse.de>
This restructures the code that creates HPT (hashed page table)
entries so that it can be called in situations where we don't have a
struct vcpu pointer, only a struct kvm pointer. It also fixes a bug
where kvmppc_map_vrma() would corrupt the guest R4 value.
Most of the work of kvmppc_virtmode_h_enter is now done by a new
function, kvmppc_virtmode_do_h_enter, which itself calls another new
function, kvmppc_do_h_enter, which contains most of the old
kvmppc_h_enter. The new kvmppc_do_h_enter takes explicit arguments
for the place to return the HPTE index, the Linux page tables to use,
and whether it is being called in real mode, thus removing the need
for it to have the vcpu as an argument.
Currently kvmppc_map_vrma creates the VRMA (virtual real mode area)
HPTEs by calling kvmppc_virtmode_h_enter, which is designed primarily
to handle H_ENTER hcalls from the guest that need to pin a page of
memory. Since H_ENTER returns the index of the created HPTE in R4,
kvmppc_virtmode_h_enter updates the guest R4, corrupting the guest R4
in the case when it gets called from kvmppc_map_vrma on the first
VCPU_RUN ioctl. With this, kvmppc_map_vrma instead calls
kvmppc_virtmode_do_h_enter with the address of a dummy word as the
place to store the HPTE index, thus avoiding corrupting the guest R4.
Signed-off-by: Paul Mackerras <paulus@samba.org>
Signed-off-by: Alexander Graf <agraf@suse.de>
In order to support the generic eventfd infrastructure on PPC, we need
to call into the generic KVM in-kernel device mmio code.
Signed-off-by: Alexander Graf <agraf@suse.de>
The current eventfd code assumes that when we have eventfd, we also have
irqfd for in-kernel interrupt delivery. This is not necessarily true. On
PPC we don't have an in-kernel irqchip yet, but we can still support easily
support eventfd.
Signed-off-by: Alexander Graf <agraf@suse.de>
CPUID.7.0.EBX[1]=1 indicates IA32_TSC_ADJUST MSR 0x3b is supported
Basic design is to emulate the MSR by allowing reads and writes to a guest
vcpu specific location to store the value of the emulated MSR while adding
the value to the vmcs tsc_offset. In this way the IA32_TSC_ADJUST value will
be included in all reads to the TSC MSR whether through rdmsr or rdtsc. This
is of course as long as the "use TSC counter offsetting" VM-execution control
is enabled as well as the IA32_TSC_ADJUST control.
However, because hardware will only return the TSC + IA32_TSC_ADJUST +
vmsc tsc_offset for a guest process when it does and rdtsc (with the correct
settings) the value of our virtualized IA32_TSC_ADJUST must be stored in one
of these three locations. The argument against storing it in the actual MSR
is performance. This is likely to be seldom used while the save/restore is
required on every transition. IA32_TSC_ADJUST was created as a way to solve
some issues with writing TSC itself so that is not an option either.
The remaining option, defined above as our solution has the problem of
returning incorrect vmcs tsc_offset values (unless we intercept and fix, not
done here) as mentioned above. However, more problematic is that storing the
data in vmcs tsc_offset will have a different semantic effect on the system
than does using the actual MSR. This is illustrated in the following example:
The hypervisor set the IA32_TSC_ADJUST, then the guest sets it and a guest
process performs a rdtsc. In this case the guest process will get
TSC + IA32_TSC_ADJUST_hyperviser + vmsc tsc_offset including
IA32_TSC_ADJUST_guest. While the total system semantics changed the semantics
as seen by the guest do not and hence this will not cause a problem.
Signed-off-by: Will Auld <will.auld@intel.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
In order to track who initiated the call (host or guest) to modify an msr
value I have changed function call parameters along the call path. The
specific change is to add a struct pointer parameter that points to (index,
data, caller) information rather than having this information passed as
individual parameters.
The initial use for this capability is for updating the IA32_TSC_ADJUST msr
while setting the tsc value. It is anticipated that this capability is
useful for other tasks.
Signed-off-by: Will Auld <will.auld@intel.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Prior to memory slot sorting this loop compared all of the user memory
slots for overlap with new entries. With memory slot sorting, we're
just checking some number of entries in the array that may or may not
be user slots. Instead, walk all the slots with kvm_for_each_memslot,
which has the added benefit of terminating early when we hit the first
empty slot, and skip comparison to private slots.
Cc: stable@vger.kernel.org
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
vmcs->cpu indicates whether it exists on the target cpu, -1 means the vmcs
does not exist on any vcpu
If vcpu load vmcs with vmcs.cpu = -1, it can be directly added to cpu's percpu
list. The list can be corrupted if the cpu prefetch the vmcs's list before
reading vmcs->cpu. Meanwhile, we should remove vmcs from the list before
making vmcs->vcpu == -1 be visible
Signed-off-by: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
In loaded_vmcs_clear, loaded_vmcs->cpu is the fist parameter passed to
smp_call_function_single, if the target cpu is downing (doing cpu hot remove),
loaded_vmcs->cpu can become -1 then -1 is passed to smp_call_function_single
It can be triggered when vcpu is being destroyed, loaded_vmcs_clear is called
in the preemptionable context
Signed-off-by: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
As Frederic pointed idle_cpu() may return false even if async fault
happened in the idle task if wake up is pending. In this case the code
will try to put idle task to sleep. Fix this by using is_idle_task() to
check for idle task.
Reported-by: Frederic Weisbecker <fweisbec@gmail.com>
Signed-off-by: Gleb Natapov <gleb@redhat.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
As requested by Glauber, do not update kvmclock area on vcpu->pcpu
migration, in case the host has stable TSC.
This is to reduce cacheline bouncing.
Acked-by: Glauber Costa <glommer@parallels.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
With master clock, a pvclock clock read calculates:
ret = system_timestamp + [ (rdtsc + tsc_offset) - tsc_timestamp ]
Where 'rdtsc' is the host TSC.
system_timestamp and tsc_timestamp are unique, one tuple
per VM: the "master clock".
Given a host with synchronized TSCs, its obvious that
guest TSC must be matched for the above to guarantee monotonicity.
Allow master clock usage only if guest TSCs are synchronized.
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
KVM added a global variable to guarantee monotonicity in the guest.
One of the reasons for that is that the time between
1. ktime_get_ts(×pec);
2. rdtscll(tsc);
Is variable. That is, given a host with stable TSC, suppose that
two VCPUs read the same time via ktime_get_ts() above.
The time required to execute 2. is not the same on those two instances
executing in different VCPUS (cache misses, interrupts...).
If the TSC value that is used by the host to interpolate when
calculating the monotonic time is the same value used to calculate
the tsc_timestamp value stored in the pvclock data structure, and
a single <system_timestamp, tsc_timestamp> tuple is visible to all
vcpus simultaneously, this problem disappears. See comment on top
of pvclock_update_vm_gtod_copy for details.
Monotonicity is then guaranteed by synchronicity of the host TSCs
and guest TSCs.
Set TSC stable pvclock flag in that case, allowing the guest to read
clock from userspace.
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Register a notifier for clocksource change event. In case
the host switches to clock other than TSC, disable master
clock usage.
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
As suggested by John, export time data similarly to how its
done by vsyscall support. This allows KVM to retrieve necessary
information to implement vsyscall support in KVM guests.
Acked-by: John Stultz <johnstul@us.ibm.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Improve performance of time system calls when using Linux pvclock,
by reading time info from fixmap visible copy of pvclock data.
Originally from Jeremy Fitzhardinge.
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Hook into generic pvclock vsyscall code, with the aim to
allow userspace to have visibility into pvclock data.
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Originally from Jeremy Fitzhardinge.
Introduce generic, non hypervisor specific, pvclock initialization
routines.
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Originally from Jeremy Fitzhardinge.
So code can be reused.
Acked-by: Glauber Costa <glommer@parallels.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Originally from Jeremy Fitzhardinge.
We can copy the information directly from "struct pvclock_vcpu_time_info",
remove pvclock_shadow_time.
Reviewed-by: Glauber Costa <glommer@parallels.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Originally from Jeremy Fitzhardinge.
pvclock_get_time_values, which contains the memory barriers
will be removed by next patch.
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
We want to expose the pvclock shared memory areas, which
the hypervisor periodically updates, to userspace.
For a linear mapping from userspace, it is necessary that
entire page sized regions are used for array of pvclock
structures.
There is no such guarantee with per cpu areas, therefore move
to memblock_alloc based allocation.
Acked-by: Glauber Costa <glommer@parallels.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Otherwise its possible for an unrelated KVM_REQ_UPDATE_CLOCK (such as due to CPU
migration) to clear the bit.
Noticed by Paolo Bonzini.
Reviewed-by: Gleb Natapov <gleb@redhat.com>
Reviewed-by: Glauber Costa <glommer@parallels.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
No need to check return value before breaking switch.
Signed-off-by: Guo Chao <yan@linux.vnet.ibm.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Return value of this function will be that of ioctl().
#include <stdio.h>
#include <linux/kvm.h>
int main () {
int fd;
fd = open ("/dev/kvm", 0);
fd = ioctl (fd, KVM_CREATE_VM, 0);
ioctl (fd, KVM_SET_TSS_ADDR, 0xfffff000);
perror ("");
return 0;
}
Output is "Operation not permitted". That's not what
we want.
Return -EINVAL in this case.
Signed-off-by: Guo Chao <yan@linux.vnet.ibm.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
We should avoid kfree()ing error pointer in kvm_vcpu_ioctl() and
kvm_arch_vcpu_ioctl().
Signed-off-by: Guo Chao <yan@linux.vnet.ibm.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
I have no access to my AMD email address anymore. Update
entry in MAINTAINERS to the new address.
Cc: Avi Kivity <avi@redhat.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Signed-off-by: Joerg Roedel <joro@8bytes.org>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
The new uapi framework splits kernel internal and user space exported
bits of header files more cleanly. Adjust the ePAPR header accordingly.
Signed-off-by: Alexander Graf <agraf@suse.de>
This fixes an error in the inline asm in try_lock_hpte() where we
were erroneously using a register number as an immediate operand.
The bug only affects an error path, and in fact the code will still
work as long as the compiler chooses some register other than r0
for the "bits" variable. Nevertheless it should still be fixed.
Signed-off-by: Paul Mackerras <paulus@samba.org>
Signed-off-by: Alexander Graf <agraf@suse.de>
Commit 55b665b026 ("KVM: PPC: Book3S HV: Provide a way for userspace
to get/set per-vCPU areas") includes a check on the length of the
dispatch trace log (DTL) to make sure the buffer is at least one entry
long. This is appropriate when registering a buffer, but the
interface also allows for any existing buffer to be unregistered by
specifying a zero address. In this case the length check is not
appropriate. This makes the check conditional on the address being
non-zero.
Signed-off-by: Paul Mackerras <paulus@samba.org>
Signed-off-by: Alexander Graf <agraf@suse.de>
Currently the code that accounts stolen time tends to overestimate the
stolen time, and will sometimes report more stolen time in a DTL
(dispatch trace log) entry than has elapsed since the last DTL entry.
This can cause guests to underflow the user or system time measured
for some tasks, leading to ridiculous CPU percentages and total runtimes
being reported by top and other utilities.
In addition, the current code was designed for the previous policy where
a vcore would only run when all the vcpus in it were runnable, and so
only counted stolen time on a per-vcore basis. Now that a vcore can
run while some of the vcpus in it are doing other things in the kernel
(e.g. handling a page fault), we need to count the time when a vcpu task
is preempted while it is not running as part of a vcore as stolen also.
To do this, we bring back the BUSY_IN_HOST vcpu state and extend the
vcpu_load/put functions to count preemption time while the vcpu is
in that state. Handling the transitions between the RUNNING and
BUSY_IN_HOST states requires checking and updating two variables
(accumulated time stolen and time last preempted), so we add a new
spinlock, vcpu->arch.tbacct_lock. This protects both the per-vcpu
stolen/preempt-time variables, and the per-vcore variables while this
vcpu is running the vcore.
Finally, we now don't count time spent in userspace as stolen time.
The task could be executing in userspace on behalf of the vcpu, or
it could be preempted, or the vcpu could be genuinely stopped. Since
we have no way of dividing up the time between these cases, we don't
count any of it as stolen.
Signed-off-by: Paul Mackerras <paulus@samba.org>
Signed-off-by: Alexander Graf <agraf@suse.de>
Currently the Book3S HV code implements a policy on multi-threaded
processors (i.e. POWER7) that requires all of the active vcpus in a
virtual core to be ready to run before we run the virtual core.
However, that causes problems on reset, because reset stops all vcpus
except vcpu 0, and can also reduce throughput since all four threads
in a virtual core have to wait whenever any one of them hits a
hypervisor page fault.
This relaxes the policy, allowing the virtual core to run as soon as
any vcpu in it is runnable. With this, the KVMPPC_VCPU_STOPPED state
and the KVMPPC_VCPU_BUSY_IN_HOST state have been combined into a single
KVMPPC_VCPU_NOTREADY state, since we no longer need to distinguish
between them.
Signed-off-by: Paul Mackerras <paulus@samba.org>
Signed-off-by: Alexander Graf <agraf@suse.de>
If a thread in a virtual core becomes runnable while other threads
in the same virtual core are already running in the guest, it is
possible for the latecomer to join the others on the core without
first pulling them all out of the guest. Currently this only happens
rarely, when a vcpu is first started. This fixes some bugs and
omissions in the code in this case.
First, we need to check for VPA updates for the latecomer and make
a DTL entry for it. Secondly, if it comes along while the master
vcpu is doing a VPA update, we don't need to do anything since the
master will pick it up in kvmppc_run_core. To handle this correctly
we introduce a new vcore state, VCORE_STARTING. Thirdly, there is
a race because we currently clear the hardware thread's hwthread_req
before waiting to see it get to nap. A latecomer thread could have
its hwthread_req cleared before it gets to test it, and therefore
never increment the nap_count, leading to messages about wait_for_nap
timeouts.
Signed-off-by: Paul Mackerras <paulus@samba.org>
Signed-off-by: Alexander Graf <agraf@suse.de>
There were a few places where we were traversing the list of runnable
threads in a virtual core, i.e. vc->runnable_threads, without holding
the vcore spinlock. This extends the places where we hold the vcore
spinlock to cover everywhere that we traverse that list.
Since we possibly need to sleep inside kvmppc_book3s_hv_page_fault,
this moves the call of it from kvmppc_handle_exit out to
kvmppc_vcpu_run, where we don't hold the vcore lock.
In kvmppc_vcore_blocked, we don't actually need to check whether
all vcpus are ceded and don't have any pending exceptions, since the
caller has already done that. The caller (kvmppc_run_vcpu) wasn't
actually checking for pending exceptions, so we add that.
The change of if to while in kvmppc_run_vcpu is to make sure that we
never call kvmppc_remove_runnable() when the vcore state is RUNNING or
EXITING.
Signed-off-by: Paul Mackerras <paulus@samba.org>
Signed-off-by: Alexander Graf <agraf@suse.de>