During kernel build in dom0, half of the resulting slow reflections
are alt DTLB misses. Since we know Linux identity maps region 7,
Xen can handle the miss itself. This reduces the system time component
of the kernel build by nearly 15%.
Signed-off-by: Anthony Xu <anthony.xu@intel.com>
1) This patch moved some processing from vcpu_initialise() and
added a new function vcpu_late_initialise().
It executes the following initializations for VCPU of
dom0/domU.
- Allocate the VHPT
- Allocate the privregs area and assign these pages into
guest pseudo physical address space.
- Set the tlbflush_timestamp.
It is executed in the following sequence.
dom0:
start_kernel()
->domain_create()
->alloc_vcpu(VCPU0)
->alloc_vcpu_struct(VCPU0)
->vcpu_initialise(VCPU0)
->vcpu_late_initialise(VCPU0)
->construct_dom0
->alloc_vcpu(othe VCPUs)
->alloc_vcpu_struct(other VCPUs)
->vcpu_initialise(other VCPUs)
ia64_hypercall(FW_HYPERCALL_IPI)
->fw_hypercall_ipi(XEN_SAL_BOOT_RENDEZ_VEC)
->arch_set_info_guest(other VCPUs)
->vcpu_late_initialise(other VCPUs)
domU:
do_domctl(XEN_DOMCTL_createdomain)
->domain_create()
do_domctl(XEN_DOMCTL_max_vcpus)
->alloc_vcpu(all VCPUs)
->alloc_vcpu_struct(all VCPUs)
->vcpu_initialise(all VCPUs)
do_domctl(XEN_DOMCTL_setvcpucontext)
->set_info_guest(VCPU0)
->arch_set_info_guest(VCPU0)
->vcpu_late_initialise(VCPU0)
ia64_hypercall(FW_HYPERCALL_IPI)
->fw_hypercall_ipi(XEN_SAL_BOOT_RENDEZ_VEC)
->arch_set_info_guest(other VCPUs)
->vcpu_late_initialise(other VCPUs)
2) This patch modified the domain_set_shared_info_va().
Currently, initialization of arch.privregs->interrupt_mask_addr
of all VCPUs is executed in domain_set_shared_info_va().
However, allocation of privregs area is late by modified of 1).
Therefore, this patch modified initialization of
arch.privregs->interrupt_mask_addr to the following sequence.
dom0 and domU:
ia64_hypercall(FW_HYPERCALL_SET_SHARED_INFO_VA)
->domain_set_shared_info_va()
Initialize interrupt_mask_addr of VCPU0
ia64_hypercall(FW_HYPERCALL_IPI)
->fw_hypercall_ipi(XEN_SAL_BOOT_RENDEZ_VEC)
->arch_set_info_guest(other VCPUs)
->vcpu_late_initialise(other VCPUs)
Initialize interrupt_mask_addr of other VCPUs
Signed-off-by: Masaki Kanno <kanno.masaki@jp.fujitsu.com>
Allow extra application objects to be linked against a partial link of
minios with non-global names localised. For now we assume that global
names ill be assigned a xenos_ prefix in due course. We may decide to
choose a different prefix.
Those who wish to finish the link from an external Makefile can link
their own objects against mini-os.o (which must be first in link
order!) and specify the appropriate architecture linker script.
Signed-off-by: Keir Fraser <keir@xensource.com>
Using the new hypercall path, scratch registers are not saved/restored.
So after returning from hypercall, r20 may contain garbage data.
This triggered General Exception.
Signed-off-by: Anthony Xu <anthony.xu@intel.com>
Some architectures, notably ia64, can automatically place
the crash kernel into an appropriate place if the start
address for crashkernel is specified as zero or omitted.
E.g. crashkernel=256M or crashkernel=256M@0
While xen does not actually have support for ia64 kexec,
I am working on it. And in any case this change should be harmless.
Just a small bit of infastructure that will make things easier in the
future.
This should also be possible on x86, now relocatable kernels
are reloacatble on x86. So once xen moves up to using linux ~2.6.19
it would make sense to look into implementing this.
The patch does 3 things.
* Firstly, parse_crashkernel() is modified to allows the size and start elements of kexec_crash_area
to be set to any value, including zero. Previously if either size or
start were specified as zero, they would both end up as zero.
* Secondly, when kexec_get() is called, if either the size or start
elements of kexec_crash_area are zero, then zero is passed
back to the hypercall caller for both values.
This gives the same behaviour as having parse_crashkernel() set
size and start to zero if either of them are zero, but it allows
architecture sepcific code called between the invocation of
parse_crashkernel() and kexec_get() to modify start (and size if
there was a reason). In particular, this allows the
architecture specific code to find a good start point if 0 is
specified.
* Lastly, it ads an additional check to the x86 setup code.
As this code currently does not know how to deal with a
0 start address, it doesn't reserve memory if start is 0.
This is neccessary as previously if start was specified as zero,
parse_crashkernel() would set size to zero, but that is
no longer the case, so a stronger check is needed.
I would really appreciate it if this patch was merged,
as I don't believe it harms anything, and it does allow
a nice path for moving forwards.
Signed-off-by: Simon Horman <horms@verge.net.au>
Use non-trivial dma_{un,}map_page only when CONFIG_HIGHMEM, allowing
to not compile swiotlb_{un,}map_page in that case, which in neither
case need exporting.
Signed-off-by: Jan Beulich <jbeulich@novell.com>
Use address width needed by device rather than dma_bits in
dma_alloc_coherent().
Probe supported address width in swiotlb initialization.
Signed-off-by: Jan Beulich <jbeulich@novell.com>
The ipv6-no-autoconf patch didn't disable IPv6 addrconf completely.
This means that the Xen bridge device still interfered with normal
IPv6 operation by engaging the IPv6 network with a bogus MAC address.
For details please refer to
https://bugzilla.redhat.com/bugzilla/show_bug.cgi?id=200360
The following patch completely disables IPv6 on the Xen bridge device
by temporarily setting the MTU to a value less than the minimum
allowed for IPv6.
Upstream will provide a cleaner way to disable IPv6 addrconf in
future, possibly in the form of a proc sysctl. Of course if the Xen
loopback device is removed it would render this change unnecessary.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
The functions die_if_kernel and vmx_die_if_kernel can certainly
return. This disqualifies them from using the noreturn keyword
which is reserved for functions that never return.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
- include LOCK prefix presence in calculation of which CR/DR is being
targeted by moves to/from these registers
- fail any other opcodes when beingused with lock prefix
Signed-off-by: Jan Beulich <jbeulich@novell.com>
While, as discussed, the performance impact of this option is
certainly higher than on native Linux, the option should not be
entirely disallowed if people want to sacrifice performance for less
lowmem pressure.
Signed-off-by: Jan Beulich <jbeulich@novell.com>
PUDs,PMDs,PTEs are all marked as ForeignPage so that they can be
grabbed from tlb_remove_page() at the appropriate time and freed in a
special way.
Signed-off-by: Jan Beulich <jbeulich@novell.com>
