Many msrs and the like will only be used by the host if we schedule() or
return to userspace. Therefore, we avoid saving them if we handle the
exit within the kernel, and if a reschedule is not requested.
Based on a patch from Eddie Dong <eddie.dong@intel.com> with a couple of
fixes by me.
Signed-off-by: Yaozu(Eddie) Dong <eddie.dong@intel.com>
Signed-off-by: Avi Kivity <avi@qumranet.com>
This patch enables IO bitmaps control on vmx and unmask the 0x80 port to
avoid VMEXITs caused by accessing port 0x80. 0x80 is used as delays (see
include/asm/io.h), and handling VMEXITs on its access is unnecessary but
slows things down. This patch improves kernel build test at around
3%~5%.
Because every VM uses the same io bitmap, it is shared between
all VMs rather than a per-VM data structure.
Signed-off-by: Qing He <qing.he@intel.com>
Signed-off-by: Avi Kivity <avi@qumranet.com>
The lazy fpu changes did not take into account that some vmexit handlers
can sleep. Move loading the guest state into the inner loop so that it
can be reloaded if necessary, and move loading the host state into
vmx_vcpu_put() so it can be performed whenever we relinquish the vcpu.
Signed-off-by: Avi Kivity <avi@qumranet.com>
Fix following section mismatch warning in kvm-intel.o:
WARNING: o-i386/drivers/kvm/kvm-intel.o(.init.text+0xbd): Section mismatch: reference to .exit.text: (between 'hardware_setup' and 'vmx_disabled_by_bios')
The function free_kvm_area is used in the function alloc_kvm_area which
is marked __init.
The __exit area is discarded by some archs during link-time if a
module is built-in resulting in an oops.
Note: This warning is only seen by my local copy of modpost
but the change will soon hit upstream.
Signed-off-by: Sam Ravnborg <sam@ravnborg.org>
Cc: Avi Kivity <avi@qumranet.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
First thing mm.h does is including sched.h solely for can_do_mlock() inline
function which has "current" dereference inside. By dealing with can_do_mlock()
mm.h can be detached from sched.h which is good. See below, why.
This patch
a) removes unconditional inclusion of sched.h from mm.h
b) makes can_do_mlock() normal function in mm/mlock.c
c) exports can_do_mlock() to not break compilation
d) adds sched.h inclusions back to files that were getting it indirectly.
e) adds less bloated headers to some files (asm/signal.h, jiffies.h) that were
getting them indirectly
Net result is:
a) mm.h users would get less code to open, read, preprocess, parse, ... if
they don't need sched.h
b) sched.h stops being dependency for significant number of files:
on x86_64 allmodconfig touching sched.h results in recompile of 4083 files,
after patch it's only 3744 (-8.3%).
Cross-compile tested on
all arm defconfigs, all mips defconfigs, all powerpc defconfigs,
alpha alpha-up
arm
i386 i386-up i386-defconfig i386-allnoconfig
ia64 ia64-up
m68k
mips
parisc parisc-up
powerpc powerpc-up
s390 s390-up
sparc sparc-up
sparc64 sparc64-up
um-x86_64
x86_64 x86_64-up x86_64-defconfig x86_64-allnoconfig
as well as my two usual configs.
Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Only save/restore the FPU host state when the guest is actually using the
FPU.
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
Signed-off-by: Avi Kivity <avi@qumranet.com>
Set all of the host mask bits for CR0 so that we can maintain a proper
shadow of CR0. This exposes CR0.TS, paving the way for lazy fpu handling.
Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
Signed-off-by: Avi Kivity <avi@qumranet.com>
Make the exit statistics per-vcpu instead of global. This gives a 3.5%
boost when running one virtual machine per core on my two socket dual core
(4 cores total) machine.
Signed-off-by: Avi Kivity <avi@qumranet.com>
Intel hosts only support syscall/sysret in long more (and only if efer.sce
is enabled), so only reload the related MSR_K6_STAR if the guest will
actually be able to use it.
This reduces vmexit cost by about 500 cycles (6400 -> 5870) on my setup.
Signed-off-by: Avi Kivity <avi@qumranet.com>
Some msrs are only used by x86_64 instructions, and are therefore
not needed when the guest is legacy mode. By not bothering to switch
them, we reduce vmexit latency by 2400 cycles (from about 8800) when
running a 32-bt guest on a 64-bit host.
Signed-off-by: Avi Kivity <avi@qumranet.com>
THe automatically switched msrs are never changed on the host (with
the exception of MSR_KERNEL_GS_BASE) and thus there is no need to save
them on every vm entry.
This reduces vmexit latency by ~400 cycles on i386 and by ~900 cycles (10%)
on x86_64.
Signed-off-by: Avi Kivity <avi@qumranet.com>
The expression
sp - 6 < sp
where sp is a u16 is undefined in C since 'sp - 6' is promoted to int,
and signed overflow is undefined in C. gcc 4.2 actually warns about it.
Replace with a simpler test.
Signed-off-by: Eric Sesterhenn <snakebyte@gmx.de>
Signed-off-by: Avi Kivity <avi@qumranet.com>
Mapping a guest page to a host page is a common operation. Currently,
one has first to find the memory slot where the page belongs (gfn_to_memslot),
then locate the page itself (gfn_to_page()).
This is clumsy, and also won't work well with memory aliases. So simplify
gfn_to_page() not to require memory slot translation first, and instead do it
internally.
Signed-off-by: Avi Kivity <avi@qumranet.com>
As usual, we need to mangle segment registers when emulating real mode
as vm86 has specific constraints. We special case the reset segment base,
and set the "access rights" (or descriptor flags) to vm86 comaptible values.
This fixes reboot on vmx.
Signed-off-by: Avi Kivity <avi@qumranet.com>
set_cr0_no_modeswitch() was a hack to avoid corrupting segment registers.
As we now cache the protected mode values on entry to real mode, this
isn't an issue anymore, and it interferes with reboot (which usually _is_
a modeswitch).
Signed-off-by: Avi Kivity <avi@qumranet.com>
The reset state has cs.selector == 0xf000 and cs.base == 0xffff0000,
which aren't compatible with vm86 mode, which is used for real mode
virtualization.
When we create a vcpu, we set cs.base to 0xf0000, but if we get there by
way of a reset, the values are inconsistent and vmx refuses to enter
guest mode.
Workaround by detecting the state and munging it appropriately.
Signed-off-by: Avi Kivity <avi@qumranet.com>
The current string pio interface communicates using guest virtual addresses,
relying on userspace to translate addresses and to check permissions. This
interface cannot fully support guest smp, as the check needs to take into
account two pages at one in case an unaligned string transfer straddles a
page boundary.
Change the interface not to communicate guest addresses at all; instead use
a buffer page (mmaped by userspace) and do transfers there. The kernel
manages the virtual to physical translation and can perform the checks
atomically by taking the appropriate locks.
Signed-off-by: Avi Kivity <avi@qumranet.com>
This is redundant, as we also return -EINTR from the ioctl, but it
allows us to examine the exit_reason field on resume without seeing
old data.
Signed-off-by: Avi Kivity <avi@qumranet.com>
Currently, userspace is told about the nature of the last exit from the
guest using two fields, exit_type and exit_reason, where exit_type has
just two enumerations (and no need for more). So fold exit_type into
exit_reason, reducing the complexity of determining what really happened.
Signed-off-by: Avi Kivity <avi@qumranet.com>
KVM used to handle cpuid by letting userspace decide what values to
return to the guest. We now handle cpuid completely in the kernel. We
still let userspace decide which values the guest will see by having
userspace set up the value table beforehand (this is necessary to allow
management software to set the cpu features to the least common denominator,
so that live migration can work).
The motivation for the change is that kvm kernel code can be impacted by
cpuid features, for example the x86 emulator.
Signed-off-by: Avi Kivity <avi@qumranet.com>
Currently when passing the a PIO emulation request to userspace, we
rely on userspace updating %rax (on 'in' instructions) and %rsi/%rdi/%rcx
(on string instructions). This (a) requires two extra ioctls for getting
and setting the registers and (b) is unfriendly to non-x86 archs, when
they get kvm ports.
So fix by doing the register fixups in the kernel and passing to userspace
only an abstract description of the PIO to be done.
Signed-off-by: Avi Kivity <avi@qumranet.com>
Instead of twiddling the rip registers directly, use the
skip_emulated_instruction() function to do that for us.
Signed-off-by: Dor Laor <dor.laor@qumranet.com>
Signed-off-by: Avi Kivity <avi@qumranet.com>
failed VM entry on VMX might still change %fs or %gs, thus make sure
that KVM always reloads the segment selectors. This is crutial on both
x86 and x86_64: x86 has __KERNEL_PDA in %fs on which things like
'current' depends and x86_64 has 0 there and needs MSR_GS_BASE to work.
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Intel virtualization extensions do not support virtualizing real mode. So
kvm uses virtualized vm86 mode to run real mode code. Unfortunately, this
virtualized vm86 mode does not support the so called "big real" mode, where
the segment selector and base do not agree with each other according to the
real mode rules (base == selector << 4).
To work around this, kvm checks whether a selector/base pair violates the
virtualized vm86 rules, and if so, forces it into conformance. On a
transition back to protected mode, if we see that the guest did not touch
a forced segment, we restore it back to the original protected mode value.
This pile of hacks breaks down if the gdt has changed in real mode, as it
can cause a segment selector to point to a system descriptor instead of a
normal data segment. In fact, this happens with the Windows bootloader
and the qemu acpi bios, where a protected mode memcpy routine issues an
innocent 'pop %es' and traps on an attempt to load a system descriptor.
"Fix" by checking if the to-be-restored selector points at a system segment,
and if so, coercing it into a normal data segment. The long term solution,
of course, is to abandon vm86 mode and use emulation for big real mode.
Signed-off-by: Avi Kivity <avi@qumranet.com>
The vmx code currently treats the guest's sysenter support msrs as 32-bit
values, which breaks 32-bit compat mode userspace on 64-bit guests. Fix by
using the native word width of the machine.
Signed-off-by: Avi Kivity <avi@qumranet.com>
Allocate a distinct inode for every vcpu in a VM. This has the following
benefits:
- the filp cachelines are no longer bounced when f_count is incremented on
every ioctl()
- the API and internal code are distinctly clearer; for example, on the
KVM_GET_REGS ioctl, there is no need to copy the vcpu number from
userspace and then copy the registers back; the vcpu identity is derived
from the fd used to make the call
Right now the performance benefits are completely theoretical since (a) we
don't support more than one vcpu per VM and (b) virtualization hardware
inefficiencies completely everwhelm any cacheline bouncing effects. But
both of these will change, and we need to prepare the API today.
Signed-off-by: Avi Kivity <avi@qumranet.com>
This adds a special MSR based hypercall API to KVM. This is to be
used by paravirtual kernels and virtual drivers.
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Avi Kivity <avi@qumranet.com>
The whole thing is rotten, but this allows vmx to boot with the guest reboot
fix.
Signed-off-by: Markus Rechberger <markus.rechberger@amd.com>
Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
Signed-off-by: Avi Kivity <avi@qumranet.com>
Convert the PDA code to use %fs rather than %gs as the segment for
per-processor data. This is because some processors show a small but
measurable performance gain for reloading a NULL segment selector (as %fs
generally is in user-space) versus a non-NULL one (as %gs generally is).
On modern processors the difference is very small, perhaps undetectable.
Some old AMD "K6 3D+" processors are noticably slower when %fs is used
rather than %gs; I have no idea why this might be, but I think they're
sufficiently rare that it doesn't matter much.
This patch also fixes the math emulator, which had not been adjusted to
match the changed struct pt_regs.
[frederik.deweerdt@gmail.com: fixit with gdb]
[mingo@elte.hu: Fix KVM too]
Signed-off-by: Jeremy Fitzhardinge <jeremy@xensource.com>
Signed-off-by: Andi Kleen <ak@suse.de>
Cc: Ian Campbell <Ian.Campbell@XenSource.com>
Acked-by: Ingo Molnar <mingo@elte.hu>
Acked-by: Zachary Amsden <zach@vmware.com>
Cc: Eric Dumazet <dada1@cosmosbay.com>
Signed-off-by: Frederik Deweerdt <frederik.deweerdt@gmail.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
On hotplug, we execute the hardware extension enable sequence. On unplug, we
decache any vcpus that last ran on the exiting cpu, and execute the hardware
extension disable sequence.
Signed-off-by: Avi Kivity <avi@qumranet.com>
Cc: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Like the inline code it replaces, this function decaches the vmcs from the cpu
it last executed on. in addition:
- vcpu_clear() works if the last cpu is also the cpu we're running on
- it is faster on larger smps by virtue of using smp_call_function_single()
Includes fix from Ingo Molnar.
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Avi Kivity <avi@qumranet.com>
Cc: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Or 32-bit userspace will get confused.
Signed-off-by: Avi Kivity <avi@qumranet.com>
Cc: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Just like svm.
Signed-off-by: Avi Kivity <avi@qumranet.com>
Cc: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Forms like "0(%rsp)" generate an instruction with an unnecessary one byte
displacement under certain circumstances. replace with the equivalent
"(%rsp)".
Signed-off-by: Avi Kivity <avi@qumranet.com>
Cc: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Intel hosts, without long mode, and with nx support disabled in the bios
have an efer that is readable but not writable. This causes a lockup on
switch to guest mode (even though it should exit with reason 34 according
to the documentation).
Signed-off-by: Avi Kivity <avi@qumranet.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The kvm mmio read path looks like:
1. guest read faults
2. kvm emulates read, calls emulator_read_emulated()
3. fails as a read requires userspace help
4. exit to userspace
5. userspace emulates read, kvm sets vcpu->mmio_read_completed
6. re-enter guest, fault again
7. kvm emulates read, calls emulator_read_emulated()
8. succeeds as vcpu->mmio_read_emulated is set
9. instruction completes and guest is resumed
A problem surfaces if the userspace exit (step 5) also requests an interrupt
injection. In that case, the guest does not re-execute the original
instruction, but the interrupt handler. The next time an mmio read is
exectued (likely for a different address), step 3 will find
vcpu->mmio_read_completed set and return the value read for the original
instruction.
The problem manifested itself in a few annoying ways:
- little squares appear randomly on console when switching virtual terminals
- ne2000 fails under nfs read load
- rtl8139 complains about "pci errors" even though the device model is
incapable of issuing them.
Fix by skipping interrupt injection if an mmio read is pending.
A better fix is to avoid re-entry into the guest, and re-emulating immediately
instead. However that's a bit more complex.
Signed-off-by: Avi Kivity <avi@qumranet.com>
Cc: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Both "=r" and "=g" breaks my build on i386:
$ make
CC [M] drivers/kvm/vmx.o
{standard input}: Assembler messages:
{standard input}:3318: Error: bad register name `%sil'
make[1]: *** [drivers/kvm/vmx.o] Error 1
make: *** [_module_drivers/kvm] Error 2
The reason is that setbe requires an 8-bit register but "=r" does not
constrain the target register to be one that has an 8-bit version on
i386.
According to
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=10153
the correct constraint is "=q".
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
No need to test for rflags.if as both VT and SVM specs assure us that on exit
caused from interrupt window opening, 'if' is set.
Signed-off-by: Dor Laor <dor.laor@qumranet.com>
Signed-off-by: Avi Kivity <avi@qumranet.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
It overwrites the right cr3 set from mmu setup. Happens only with the test
harness.
Signed-off-by: Avi Kivity <avi@qumranet.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
This will allow us to see the root cause when a vmwrite error happens.
Signed-off-by: Avi Kivity <avi@qumranet.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
