struct dentry is one of the most critical structures in the kernel. So it's
sad to see it going neglected.
With CONFIG_PROFILING turned on (which is probably the common case at least
for distros and kernel developers), sizeof(struct dcache) == 208 here
(64-bit). This gives 19 objects per slab.
I packed d_mounted into a hole, and took another 4 bytes off the inline
name length to take the padding out from the end of the structure. This
shinks it to 200 bytes. I could have gone the other way and increased the
length to 40, but I'm aiming for a magic number, read on...
I then got rid of the d_cookie pointer. This shrinks it to 192 bytes. Rant:
why was this ever a good idea? The cookie system should increase its hash
size or use a tree or something if lookups are a problem. Also the "fast
dcookie lookups" in oprofile should be moved into the dcookie code -- how
can oprofile possibly care about the dcookie_mutex? It gets dropped after
get_dcookie() returns so it can't be providing any sort of protection.
At 192 bytes, 21 objects fit into a 4K page, saving about 3MB on my system
with ~140 000 entries allocated. 192 is also a multiple of 64, so we get
nice cacheline alignment on 64 and 32 byte line systems -- any given dentry
will now require 3 cachelines to touch all fields wheras previously it
would require 4.
I know the inline name size was chosen quite carefully, however with the
reduction in cacheline footprint, it should actually be just about as fast
to do a name lookup for a 36 character name as it was before the patch (and
faster for other sizes). The memory footprint savings for names which are
<= 32 or > 36 bytes long should more than make up for the memory cost for
33-36 byte names.
Performance is a feature...
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
The invlpg and sync walkers lack knowledge of large host sptes,
descending to non-existant pagetable level.
Stop at directory level in such case.
Fixes SMP Windows XP with hugepages.
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
While most accesses to the i8259 are with the kvm mutex taken, the call
to kvm_pic_read_irq() is not. We can't easily take the kvm mutex there
since the function is called with interrupts disabled.
Fix by adding a spinlock to the virtual interrupt controller. Since we
can't send an IPI under the spinlock (we also take the same spinlock in
an irq disabled context), we defer the IPI until the spinlock is released.
Similarly, we defer irq ack notifications until after spinlock release to
avoid lock recursion.
Signed-off-by: Avi Kivity <avi@redhat.com>
The pte.g bit is meaningless if global pages are disabled; deferring
mmu page synchronization on these ptes will lead to the guest using stale
shadow ptes.
Fixes Vista x86 smp bootloader failure.
Signed-off-by: Avi Kivity <avi@redhat.com>
Fix kvm_arch_vcpu_ioctl_[gs]et_regs() to do something meaningful on
ia64. Old versions could never have worked since they required
pointers to be set in the ioctl payload which were never being set by
the ioctl handler for get_regs.
In addition reserve extra space for future extensions.
The change of layout of struct kvm_regs doesn't require adding a new
CAP since get/set regs never worked on ia64 until now.
This version doesn't support copying the KVM kernel stack in/out of
the kernel. This should be implemented in a seperate ioctl call if
ever needed.
Signed-off-by: Jes Sorensen <jes@sgi.com>
Acked-by : Xiantao Zhang <xiantao.zhang@intel.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
There is no point in doing the ready_for_nmi_injection/
request_nmi_window dance with user space. First, we don't do this for
in-kernel irqchip anyway, while the code path is the same as for user
space irqchip mode. And second, there is nothing to loose if a pending
NMI is overwritten by another one (in contrast to IRQs where we have to
save the number). Actually, there is even the risk of raising spurious
NMIs this way because the reason for the held-back NMI might already be
handled while processing the first one.
Therefore this patch creates a simplified user space NMI injection
interface, exporting it under KVM_CAP_USER_NMI and dropping the old
KVM_CAP_NMI capability. And this time we also take care to provide the
interface only on archs supporting NMIs via KVM (right now only x86).
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
As with the kernel irqchip, don't allow an NMI to stomp over an already
injected IRQ; instead wait for the IRQ injection to be completed.
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
walk_shadow assumes the caller verified validity of the pdptr pointer in
question, which is not the case for the invlpg handler.
Fixes oops during Solaris 10 install.
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
kvm_get_tsc_khz() currently returns the previously-calculated preset_lpj
value, but it is in loops-per-jiffy, not kHz. The current code works
correctly only when HZ=1000.
Signed-off-by: Eduardo Habkost <ehabkost@redhat.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
If the guest executes invlpg, peek into the pagetable and attempt to
prepopulate the shadow entry.
Also stop dirty fault updates from interfering with the fork detector.
2% improvement on RHEL3/AIM7.
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
Skip syncing global pages on cr3 switch (but not on cr4/cr0). This is
important for Linux 32-bit guests with PAE, where the kmap page is
marked as global.
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
Collapse remote TLB flushes on root sync.
kernbench is 2.7% faster on 4-way guest. Improvements have been seen
with other loads such as AIM7.
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
Instead of invoking the handler directly collect pages into
an array so the caller can work with it.
Simplifies TLB flush collapsing.
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
The VMMCALL instruction doesn't get recognised and isn't processed
by the emulator.
This is seen on an Intel host that tries to execute the VMMCALL
instruction after a guest live migrates from an AMD host.
Signed-off-by: Amit Shah <amit.shah@redhat.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
Add emulation of shld and shrd instructions
Signed-off-by: Guillaume Thouvenin <guillaume.thouvenin@ext.bull.net>
Signed-off-by: Avi Kivity <avi@redhat.com>
Add the assembler code for instruction with three operands and one
operand is stored in ECX register
Signed-off-by: Guillaume Thouvenin <guillaume.thouvenin@ext.bull.net>
Signed-off-by: Avi Kivity <avi@redhat.com>
Add SrcOne operand type when we need to decode an implied '1' like with
regular shift instruction
Signed-off-by: Guillaume Thouvenin <guillaume.thouvenin@ext.bull.net>
Signed-off-by: Avi Kivity <avi@redhat.com>
Instruction like shld has three operands, so we need to add a Src2
decode set. We start with Src2None, Src2CL, and Src2ImmByte, Src2One to
support shld/shrd and we will expand it later.
Signed-off-by: Guillaume Thouvenin <guillaume.thouvenin@ext.bull.net>
Signed-off-by: Avi Kivity <avi@redhat.com>
Extend the opcode descriptor to 32 bits. This is needed by the
introduction of a new Src2 operand type.
Signed-off-by: Guillaume Thouvenin <guillaume.thouvenin@ext.bull.net>
Signed-off-by: Avi Kivity <avi@redhat.com>
The only significant changes were to kvmppc_exit_timing_write() and
kvmppc_exit_timing_show(), both of which were dramatically simplified.
Signed-off-by: Hollis Blanchard <hollisb@us.ibm.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
Existing KVM statistics are either just counters (kvm_stat) reported for
KVM generally or trace based aproaches like kvm_trace.
For KVM on powerpc we had the need to track the timings of the different exit
types. While this could be achieved parsing data created with a kvm_trace
extension this adds too much overhead (at least on embedded PowerPC) slowing
down the workloads we wanted to measure.
Therefore this patch adds a in-kernel exit timing statistic to the powerpc kvm
code. These statistic is available per vm&vcpu under the kvm debugfs directory.
As this statistic is low, but still some overhead it can be enabled via a
.config entry and should be off by default.
Since this patch touched all powerpc kvm_stat code anyway this code is now
merged and simplified together with the exit timing statistic code (still
working with exit timing disabled in .config).
Signed-off-by: Christian Ehrhardt <ehrhardt@linux.vnet.ibm.com>
Signed-off-by: Hollis Blanchard <hollisb@us.ibm.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
Store shadow TLB entries in memory, but only use it on host context switch
(instead of every guest entry). This improves performance for most workloads on
440 by reducing the guest TLB miss rate.
Signed-off-by: Hollis Blanchard <hollisb@us.ibm.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
Formerly, we used to maintain a per-vcpu shadow TLB and on every entry to the
guest would load this array into the hardware TLB. This consumed 1280 bytes of
memory (64 entries of 16 bytes plus a struct page pointer each), and also
required some assembly to loop over the array on every entry.
Instead of saving a copy in memory, we can just store shadow mappings directly
into the hardware TLB, accepting that the host kernel will clobber these as
part of the normal 440 TLB round robin. When we do that we need less than half
the memory, and we have decreased the exit handling time for all guest exits,
at the cost of increased number of TLB misses because the host overwrites some
guest entries.
These savings will be increased on processors with larger TLBs or which
implement intelligent flush instructions like tlbivax (which will avoid the
need to walk arrays in software).
In addition to that and to the code simplification, we have a greater chance of
leaving other host userspace mappings in the TLB, instead of forcing all
subsequent tasks to re-fault all their mappings.
Signed-off-by: Hollis Blanchard <hollisb@us.ibm.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
KVM currently ignores the host's round robin TLB eviction selection, instead
maintaining its own TLB state and its own round robin index. However, by
participating in the normal 44x TLB selection, we can drop the alternate TLB
processing in KVM. This results in a significant performance improvement,
since that processing currently must be done on *every* guest exit.
Accordingly, KVM needs to be able to access and increment tlb_44x_index.
(KVM on 440 cannot be a module, so there is no need to export this symbol.)
Signed-off-by: Hollis Blanchard <hollisb@us.ibm.com>
Acked-by: Josh Boyer <jwboyer@linux.vnet.ibm.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
KVM on 440 has always been able to handle large guest mappings with 4K host
pages -- we must, since the guest kernel uses 256MB mappings.
This patch makes KVM work when the host has large pages too (tested with 64K).
Signed-off-by: Hollis Blanchard <hollisb@us.ibm.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
Impact: make global function static
arch/x86/kvm/vmx.c:134:3: warning: symbol 'vmx_capability' was not declared. Should it be static?
Signed-off-by: Hannes Eder <hannes@hanneseder.net>
Signed-off-by: Avi Kivity <avi@redhat.com>
The s390 backend of kvm never calls kvm_vcpu_uninit. This causes
a memory leak of vcpu->run pages.
Lets call kvm_vcpu_uninit in kvm_arch_vcpu_destroy to free
the vcpu->run.
Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
Acked-by: Carsten Otte <cotte@de.ibm.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
Currently it is impossible to unload the kvm module on s390.
This patch fixes kvm_arch_destroy_vm to release all cpus.
This make it possible to unload the module.
In addition we stop messing with the module refcount in arch code.
Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
Acked-by: Carsten Otte <cotte@de.ibm.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
The write protect verification in set_spte is unnecessary for page sync.
Its guaranteed that, if the unsync spte was writable, the target page
does not have a write protected shadow (if it had, the spte would have
been write protected under mmu_lock by rmap_write_protect before).
Same reasoning applies to mark_page_dirty: the gfn has been marked as
dirty via the pagefault path.
The cost of hash table and memslot lookups are quite significant if the
workload is pagetable write intensive resulting in increased mmu_lock
contention.
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
Currently, we only set the KVM paravirt signature in case
of CONFIG_KVM_GUEST. However, it is possible to have it turned
off, while CONFIG_KVM_CLOCK is turned on. This is also a paravirt
case, and should be shown accordingly.
Signed-off-by: Glauber Costa <glommer@redhat.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
If we're injecting an interrupt, and another one is pending, request
an interrupt window notification so we don't have excess latency on the
second interrupt.
This shouldn't happen in practice since an EOI will be issued, giving a second
chance to request an interrupt window, but...
Signed-off-by: Avi Kivity <avi@redhat.com>
Since vmm runs in an isolated address space and it is just a copy
of host's kvm-intel module, so once vmm crashes, we just crash all guests
running on it instead of crashing whole kernel.
Signed-off-by: Xiantao Zhang <xiantao.zhang@intel.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
Use printk infrastructure to print out some debug info once VM crashes.
Signed-off-by: Xiantao Zhang <xiantao.zhang@intel.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
kvm-intel module is relocated to an isolated address space
with kernel, so it can't call host kernel's printk for debug
purpose. In the module, we implement the printk to output debug
info of vmm.
Signed-off-by: Xiantao Zhang <xiantao.zhang@intel.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
On emergency_restart, we may need to use an NMI to disable virtualization
on all CPUs. We do that using nmi_shootdown_cpus() if VMX is enabled.
Note: With this patch, we will run the NMI stuff only when the CPU where
emergency_restart() was called has VMX enabled. This should work on most
cases because KVM enables VMX on all CPUs, but we may miss the small
window where KVM is doing that. Also, I don't know if all code using
VMX out there always enable VMX on all CPUs like KVM does. We have two
other alternatives for that:
a) Have an API that all code that enables VMX on any CPU should use
to tell the kernel core that it is going to enable VMX on the CPUs.
b) Always call nmi_shootdown_cpus() if the CPU supports VMX. This is
a bit intrusive and more risky, as it would run nmi_shootdown_cpus()
on emergency_reboot() even on systems where virtualization is never
enabled.
Finding a proper point to hook the nmi_shootdown_cpus() call isn't
trivial, as the non-emergency machine_restart() (that doesn't need the
NMI tricks) uses machine_emergency_restart() directly.
The solution to make this work without adding a new function or argument
to machine_ops was setting a 'reboot_emergency' flag that tells if
native_machine_emergency_restart() needs to do the virt cleanup or not.
Signed-off-by: Eduardo Habkost <ehabkost@redhat.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
We need to disable virtualization extensions on all CPUs before booting
the kdump kernel, otherwise the kdump kernel booting will fail, and
rebooting after the kdump kernel did its task may also fail.
We do it using cpu_emergency_vmxoff() and cpu_emergency_svm_disable(),
that should always work, because those functions check if the CPUs
support SVM or VMX before doing their tasks.
Signed-off-by: Eduardo Habkost <ehabkost@redhat.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
This function can be used by the reboot or kdump code to forcibly
disable SVM on the CPU.
Signed-off-by: Eduardo Habkost <ehabkost@redhat.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
Use a trick to keep the printk()s on has_svm() working as before. gcc
will take care of not generating code for the 'msg' stuff when the
function is called with a NULL msg argument.
Signed-off-by: Eduardo Habkost <ehabkost@redhat.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
Al Viro