The main purpose of this patch is to only ever insert the LBR MSRs into the
guest load/save list once, as a future patch wants to change the behaviour of
vmx_add_guest_msr().
The repeated processing of lbr_info and the guests MSR load/save list is
redundant, and a guest using LBR itself will have to re-enable
MSR_DEBUGCTL.LBR in its #DB handler, meaning that Xen will repeat this
redundant processing every time the guest gets a debug exception.
Rename lbr_fixup_enabled to lbr_flags to be a little more generic, and use one
bit to indicate that the MSRs have been inserted into the load/save list.
Shorten the existing FIXUP* identifiers to reduce code volume.
Furthermore, handing the guest #MC on an error isn't a legitimate action. Two
of the three failure cases are definitely hypervisor bugs, and the third is a
boundary case which shouldn't occur in practice. The guest also won't execute
correctly, so handle errors by cleanly crashing the guest.
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Acked-by: Kevin Tian <kevin.tian@intel.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
(cherry picked from commit be73a842e6)
At the moment, all modifications of the MSR lists are in current context.
However, future changes may need to put MSR_EFER into the lists from domctl
hypercall context.
Plumb a struct vcpu parameter down through the infrastructure, and use
vmx_vmcs_{enter,exit}() for safe access to the VMCS in vmx_add_msr(). Use
assertions to ensure that access is either in current context, or while the
vcpu is paused.
Note these expectations beside the fields in arch_vmx_struct, and reorder the
fields to avoid unnecessary padding.
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Acked-by: Kevin Tian <kevin.tian@intel.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
(cherry picked from commit 80599f0b77)
Instead of having multiple algorithms searching the MSR lists, implement a
single one. It has the semantics required by vmx_add_msr(), to identify the
position in which an MSR should live, if it isn't already present.
There will be a marginal improvement for vmx_find_msr() by avoiding the
function pointer calls to vmx_msr_entry_key_cmp(), and a major improvement for
vmx_add_msr() by using a binary search instead of a linear search.
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Reviewed-by: Roger Pau Monné <roger.pau@citrix.com>
Acked-by: Kevin Tian <kevin.tian@intel.com>
(cherry picked from commit 4d94828cf1)
* Use an arch_vmx_struct local variable to reduce later code volume.
* Use start/total instead of msr_area/msr_count. This is in preparation for
more finegrained handling with later changes.
* Use ent/end pointers (again for preparation), and to make the vmx_add_msr()
logic easier to follow.
* Make the memory allocation block of vmx_add_msr() unlikely, and calculate
virt_to_maddr() just once.
No practical change to functionality.
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Acked-by: Kevin Tian <kevin.tian@intel.com>
(cherry picked from commit 94fda356fc)
Collect together related infrastructure in vmcs.h, rather than having it
spread out. Turn vmx_{read,write}_guest_msr() into static inlines, as they
are simple enough.
Replace 'int type' with 'enum vmx_msr_list_type', and use switch statements
internally. Later changes are going to introduce a new type.
Rename the type identifiers for consistency with the other VMX_MSR_*
constants.
No functional change.
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Reviewed-by: Roger Pau Monné <roger.pau@citrix.com>
Acked-by: Kevin Tian <kevin.tian@intel.com>
(cherry picked from commit f54b63e861)
paging_update_paging_modes() and vmx_vlapic_msr_changed() both operate on the
VMCS being constructed. Avoid dropping and re-acquiring the reference
multiple times.
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
Acked-by: Kevin Tian <kevin.tian@intel.com>
(cherry picked from commit f30e3cf340)
Currently, whenever the guest writes a nonzero value to MSR_DEBUGCTL, Xen
updates a host MSR load list entry with the current hardware value of
MSR_DEBUGCTL.
On VMExit, hardware automatically resets MSR_DEBUGCTL to 0. Later, when the
guest writes to MSR_DEBUGCTL, the current value in hardware (0) is fed back
into guest load list. As a practical result, `ler` debugging gets lost on any
PCPU which has ever scheduled an HVM vcpu, and the common case when `ler`
debugging isn't active, guest actions result in an unnecessary load list entry
repeating the MSR_DEBUGCTL reset.
Restoration of Xen's debugging setting needs to happen from the very first
vmexit. Due to the automatic reset, Xen need take no action in the general
case, and only needs to load a value when debugging is active.
This could be fixed by using a host MSR load list entry set up during
construct_vmcs(). However, a more efficient option is to use an alternative
block in the VMExit path, keyed on whether hypervisor debugging has been
enabled.
In order to set this up, drop the per cpu ler_msr variable (as there is no
point having it per cpu when it will be the same everywhere), and use a single
read_mostly variable instead. Split calc_ler_msr() out of percpu_traps_init()
for clarity.
Finally, clean up do_debug(). Reinstate LBR early to help catch cascade
errors, which allows for the removal of the out label.
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
Reviewed-by: Kevin Tian <kevin.tian@intel.com>
(cherry picked from commit 730dc8d2c9)
As it currently stands, 'xpti=dom0' is indistinguishable from the default
value, which means it will be overridden by ARCH_CAPABILITIES_RDCL_NO on fixed
hardware.
Switch opt_xpti to use -1 as a default like all our other related options, and
clobber it as soon as we have a string to parse.
In addition, 'xpti' alone should be interpreted in its positive boolean form,
rather than resulting in a parse error.
(XEN) parameter "xpti" has invalid value "", rc=-22!
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Reviewed-by: Juergen Gross <jgross@suse.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
(cherry picked from commit 2a3b34ec47)
The calls to xpti_init_default() in parse_xpti() are buggy. The CPUID data
hasn't been fetched that early, and boot_cpu_has(X86_FEATURE_ARCH_CAPS) will
always evaluate false.
As a result, the default case won't disable XPTI on Intel hardware which
advertises ARCH_CAPABILITIES_RDCL_NO.
Simplify parse_xpti() to solely the setting of opt_xpti according to the
passed string, and have init_speculation_mitigations() call
xpti_init_default() if appropiate. Drop the force parameter, and pass caps
instead, to avoid redundant re-reading of MSR_ARCH_CAPS.
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Reviewed-by: Juergen Gross <jgross@suse.com>
Reviewed-by: Wei Liu <wei.liu2@citrix.com>
Acked-by: Jan Beulich <jbeulich@suse.com>
master commit: be5e2ff6f5
master date: 2018-07-24 11:25:54 +0100
It turns out that nothing ever prevented HVM guests from trying to set unknown
EFER bits. Generally, this results in a vmentry failure.
For Intel hardware, all implemented bits are covered by the checks.
For AMD hardware, the only EFER bit which isn't covered by the checks is TCE
(which AFAICT is specific to AMD Fam15/16 hardware). We never advertise TCE
in CPUID, but it isn't a security problem to have TCE unexpected enabled in
guest context.
Disallow the setting of bits outside of the EFER_KNOWN_MASK, which prevents
any vmentry failures for guests, yielding #GP instead.
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Reviewed-by: Roger Pau Monné <roger.pau@citrix.com>
Reviewed-by: Wei Liu <wei.liu2@citrix.com>
Acked-by: Jan Beulich <jbeulich@suse.com>
master commit: ef0269c621
master date: 2018-07-24 11:25:53 +0100
If xcr0_max exceeds xfeature_mask, then something is broken with the CPUID
policy derivation or auditing logic. If hardware rejects new_bv, then
something is broken with Xen's xstate logic.
In both cases, crash the domain with an obvious error message, to help
highlight the issues.
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
master commit: d6371ccb93
master date: 2018-07-19 19:57:26 +0100
It turns out that Xen has never enforced that a domain remain within the
xstate features advertised in CPUID.
The check of new_bv against xfeature_mask ensures that a domain stays within
the set of features that Xen has enabled in hardware (and therefore isn't a
security problem), but this does means that attempts to level a guest for
migration safety might not be effective if the guest ignores CPUID.
Check the CPUID policy in validate_xstate() (for incoming migration) and in
handle_xsetbv() (for guest XSETBV instructions). This subsumes the PKRU check
for PV guests in handle_xsetbv() (and also demonstrates that I should have
spotted this problem while reviewing c/s fbf9971241).
For migration, this is correct despite the current (mis)ordering of data
because d->arch.cpuid is the applicable max policy.
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
master commit: 361b835fa0
master date: 2018-07-19 19:57:26 +0100
c/s 4f36452b63 introduced a write to %dr6 in the #DB intercept case, but the
guests debug registers may be lazy at this point, at which point the guests
later attempt to read %dr6 will discard this value and use the older stale
value.
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Reviewed-by: Roger Pau Monné <roger.pau@citrix.com>
Acked-by: Kevin Tian <kevin.tian@intel.com>
master commit: 3cdac28056
master date: 2018-07-19 14:06:48 +0100
Shared resources (L1 cache and TLB in particular) present a risk of
information leak via side channels. Provide a means to avoid use of
hyperthreads in such cases.
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Reviewed-by: Roger Pau Monné <roger.pau@citrix.com>
Reviewed-by: Andrew Cooper <andrew.cooper3@citrix.com>
master commit: d8f974f1a6
master date: 2018-07-19 13:43:33 +0100
Reportedly Intel CPUs which can't broadcast #MC to all targeted
cores/threads because some have CR4.MCE clear will shut down. Therefore
we want to keep CR4.MCE enabled when offlining a CPU, and we need to
bring up all CPUs in order to be able to set CR4.MCE in the first place.
The use of clear_in_cr4() in cpu_mcheck_disable() was ill advised
anyway, and to avoid future similar mistakes I'm removing clear_in_cr4()
altogether right here.
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Reviewed-by: Andrew Cooper <andrew.cooper3@citrix.com>
Reviewed-by: Roger Pau Monné <roger.pau@citrix.com>
Reviewed-by: Wei Liu <wei.liu2@citrix.com>
master commit: 8797d20a6e
master date: 2018-07-19 13:43:33 +0100
In order to be able to service #MC on offlined CPUs, the GDT, IDT,
stack, and per-CPU data (which includes the TSS) need to be kept
allocated. They should only be freed upon CPU removal (which we
currently don't support, so some code is becoming effectively dead for
the moment).
Note that for now park_offline_cpus doesn't get set to true anywhere -
this is going to be the subject of a subsequent patch.
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Reviewed-by: Wei Liu <wei.liu2@citrix.com>
Reviewed-by: Andrew Cooper <andrew.cooper3@citrix.com>
master commit: 2e6c8f182c
master date: 2018-07-19 13:43:33 +0100
Fam17 replaces CUs by HTs, which we should reflect accordingly, even if
the difference is not very big. The most relevant change (requiring some
code restructuring) is that the topoext feature no longer means there is
a valid CU ID.
Take the opportunity and convert wrongly plain int variables in
set_cpu_sibling_map() to unsigned int.
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Reviewed-by: Brian Woods <brian.woods@amd.com>
Reviewed-by: Roger Pau Monné <roger.pau@citrix.com>
Acked-by: Andrew Cooper <andrew.cooper3@citrix.com>
master commit: 9429b07a0a
master date: 2018-07-19 09:42:42 +0200
While I've run into the issue with further patches in place which no
longer guarantee the per-CPU area to start out as all zeros, the
CPU_DOWN_FAILED processing looks to have the same issue: By not zapping
the per-CPU cpupool pointer, cpupool_cpu_add()'s (indirect) invocation
of schedule_cpu_switch() will trigger the "c != old_pool" assertion
there.
Clearing the field during CPU_DOWN_PREPARE is too early (afaict this
should not happen before cpu_disable_scheduler()). Clearing it in
CPU_DEAD and CPU_DOWN_FAILED would be an option, but would take the same
piece of code twice. Since the field's value shouldn't matter while the
CPU is offline, simply clear it (implicitly) for CPU_ONLINE and
CPU_DOWN_FAILED, but only for other than the suspend/resume case (which
gets specially handled in cpupool_cpu_remove()).
By adjusting the conditional in cpupool_cpu_add() CPU_DOWN_FAILED
handling in the suspend case should now also be handled better.
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Reviewed-by: Juergen Gross <jgross@suse.com>
master commit: cb1ae9a278
master date: 2018-07-19 09:41:55 +0200
The function's use of the stop-machine logic has so far prevented its
use ahead of the processing of the "ordinary" initcalls. Since at this
early time we're in a controlled environment anyway, there's no need for
such a heavy tool. Additionally this ought to have less of a performance
impact especially on large systems, compared to the alternative of
making stop-machine functionality available earlier.
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Reviewed-by: Wei Liu <wei.liu2@citrix.com>
Reviewed-by: Andrew Cooper <andrew.cooper3@citrix.com>
master commit: 5894c0a2da
master date: 2018-07-16 15:15:12 +0200
This is not really very nice.
This line doesn't have much value in itself. The rest of this comment
block is pretty clear what it wants to convey. So delete it.
(While we are here, adopt the CODING_STYLE-mandated formatting.)
Signed-off-by: Ian Jackson <Ian.Jackson@eu.citrix.com>
Acked-by: Wei Liu <wei.liu2@citrix.com>
Acked-by: Lars Kurth <lars.kurth.xen@gmail.com>
Acked-by: George Dunlap <dunlapg@umich.edu
Acked-by: Jan Beulich <JBeulich@suse.com>
master commit: 41cb2db626
master date: 2018-07-12 16:38:30 +0100
For one, "no-xen" should not imply "no-eager-fpu", as "eager FPU" mode
is to guard guests, not Xen itself, which is also expressed so by
print_details().
And then opt_ssbd, despite being off by default, should also be cleared
by the "no" and "no-xen" sub-options.
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Reviewed-by: Andrew Cooper <andrew.cooper3@citrix.com>
master commit: ac3f9a7214
master date: 2018-07-10 12:22:31 +0200
In this case, just like xcr0_accum, nonlazy_xstate_used should always be
set to the intended new value, rather than possibly leaving the flag set
from a prior state load.
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Reviewed-by: Wei Liu <wei.liu2@citrix.com>
Acked-by: Andrew Cooper <andrew.cooper3@citrix.com>
master commit: f46bf0e101
master date: 2018-07-09 10:51:02 +0200
This is as the infrastructure appeared in Linux 4.17, adapted slightly for
Xen.
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Signed-off-by: Julien Grall <julien.grall@arm.com>
Acked-by: Jan Beulich <jbeulich@suse.com>
master commit: 2ddfae51d8
master date: 2018-07-06 16:49:57 +0100
Use the calculated lengths instead of pointers, as 'e' being NULL will
otherwise cause undue parsing failures.
Reported-by: Karl Johnson <karljohnson.it@gmail.com>
Signed-off-by: Jan Beulich <jbeulich@suse.com>
The changes for XSA-267 did not touch management of CR0.TS for HVM
guests. In fully eager mode this bit should never be set when
respective vCPU-s are active, or else hvmemul_get_fpu() might leave it
wrongly set, leading to #NM in hypervisor context.
{svm,vmx}_enter() and {svm,vmx}_fpu_dirty_intercept() become unreachable
this way. Explicit {svm,vmx}_fpu_leave() invocations need to be guarded
now.
With no CR0.TS management necessary in fully eager mode, there's also no
need anymore to intercept #NM.
Reported-by: Charles Arnold <carnold@suse.com>
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Reviewed-by: Andrew Cooper <andrew.cooper3@citrix.com>
master commit: 488efc29e4
master date: 2018-06-28 09:07:06 +0200
A read-only check was introduced for XSA-142, commit ef6cb76026 ("libxl:
relax readonly check introduced by XSA-142 fix") added the passing of
the extra setting, but commit dab0539568 ("Introduce COLO mode and
refactor relevant function") dropped the passing of the setting again,
quite likely due to improper re-basing.
Restore the readonly= parameter to SCSI disks. For IDE disks this is
supposed to be rejected; add an assert. And there is a bare ad-hoc
disk drive string in libxl__build_device_model_args_new, which we also
update.
This is XSA-266.
Reported-by: Andrew Reimers <andrew.reimers@orionvm.com>
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Signed-off-by: Ian Jackson <Ian.Jackson@eu.citrix.com>
master commit: dd64d3c41a
master date: 2018-06-28 09:05:06 +0200
The generated configurations are identical apart from, in some cases,
reordering of the id=%s element. So, overall, no functional change.
This is part of XSA-266.
Reported-by: Andrew Reimers <andrew.reimers@orionvm.com>
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Signed-off-by: Ian Jackson <Ian.Jackson@eu.citrix.com>
master commit: 724e5aa31b
master date: 2018-06-28 09:04:55 +0200
One of the fix for XSA-260 (c/s 75d6828bc2 "x86/traps: Fix handling of #DB
exceptions in hypervisor context") added some safety checks to help avoid
livelocks of #DB faults.
While a General Detect #DB exception does have fault semantics, hardware
clears %dr7.gd on entry to the handler, meaning that it is actually safe to
return to. Furthermore, %dr6.gd is guest controlled and sticky (never cleared
by hardware). A malicious PV guest can therefore trigger the fatal_trap() and
crash Xen.
Instruction breakpoints are more tricky. The breakpoint match bits in %dr6
are not sticky, but the Intel manual warns that they may be set for
non-enabled breakpoints, so add a breakpoint enabled check.
Beyond that, because of the restriction on the linear addresses PV guests can
set, and the fault (rather than trap) nature of instruction breakpoints
(i.e. can't be deferred by a MovSS shadow), there should be no way to
encounter an instruction breakpoint in Xen context. However, for extra
robustness, deal with this situation by clearing the breakpoint configuration,
rather than crashing.
This is XSA-265
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
master commit: 17bf512972
master date: 2018-06-28 09:04:20 +0200
While unlikely, it is not impossible for a multi-vCPU guest to leverage
bypasses of preemption checks to drive Xen into an unbounded loop.
This is XSA-264.
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Reviewed-by: Andrew Cooper <andrew.cooper3@citrix.com>
master commit: 17608703c6
master date: 2018-06-28 09:03:09 +0200
We must not leave a vCPU with CR0.TS clear when it is not in fully eager
mode and has not touched non-lazy state. Instead of adding a 3rd
invocation of stts() to vcpu_restore_fpu_eager(), consolidate all of
them into a single one done at the end of the function.
Rename the function at the same time to better reflect its purpose, as
the patches touches all of its occurences anyway.
The new function parameter is not really well named, but
"need_stts_if_not_fully_eager" seemed excessive to me.
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Reviewed-by: Andrew Cooper <andrew.cooper3@citrix.com>
Reviewed-by: Paul Durrant <paul.durrant@citrix.com>
master commit: 23839a0fa0
master date: 2018-06-26 15:23:08 +0200
There are two issues. First, the nonlazy xstates were never restored
after returning from the runtime call.
Secondly, with the fully_eager_fpu mitigation for XSA-267 / LazyFPU, the
unilateral stts() is no longer correct, and hits an assertion later when
a lazy state restore tries to occur for a fully eager vcpu.
Fix both of these issues by calling vcpu_restore_fpu_eager(). As EFI
runtime services can be used in the idle context, the idle assertion
needs to move until after the fully_eager_fpu check.
Introduce a "curr" local variable and replace other uses of "current"
at the same time.
Reported-by: Andrew Cooper <andrew.cooper3@citrix.com>
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Tested-by: Juergen Gross <jgross@suse.com>
master commit: 437211cb69
master date: 2018-06-21 11:35:46 +0200
Even with opt_msr_sc_{pv,hvm} both false we should set up the variable
as usual, to ensure proper one-time setup during boot and CPU bringup.
This then also brings the code in line with the comment immediately
ahead of the printk() being modified saying "irrespective of guests".
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Reviewed-by: Andrew Cooper <andrew.cooper3@citrix.com>
master commit: d6239f6471
master date: 2018-05-29 12:38:52 +0200
Just like for HVM the feature set should be used for EBX output, while
EAX should be restricted to the low 16 bits and ECX/EDX should be zero.
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Reviewed-by: Andrew Cooper <andrew.cooper3@citrix.com>
master commit: 849cc9ac56
master date: 2018-05-29 10:51:02 +0100
Intel Core processors since at least Nehalem speculate past #NM, which is the
mechanism by which lazy FPU context switching is implemented.
On affected processors, Xen must use fully eager FPU context switching to
prevent guests from being able to read FPU state (SSE/AVX/etc) from previously
scheduled vcpus.
This is part of XSA-267 / CVE-2018-3665
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
(cherry picked from commit 243435bf67)
This is controlled on a per-vcpu bases for flexibility.
This is part of XSA-267 / CVE-2018-3665
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
(cherry picked from commit 146dfe9277)
While the involved code (in pv_domain_initialise()) sits behind an
!is_idle_domain() check already in 4.10, we need to add one here.
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Reviewed-by: Andrew Cooper <andrew.cooper3@citrix.com>
Avoid flushing the complete TLB when switching %cr3 for mitigation of
Meltdown by using the PCID feature if available.
We are using 4 PCID values for a 64 bit pv domain subject to XPTI and
2 values for the non-XPTI case:
- guest active and in kernel mode
- guest active and in user mode
- hypervisor active and guest in user mode (XPTI only)
- hypervisor active and guest in kernel mode (XPTI only)
We use PCID only if PCID _and_ INVPCID are supported. With PCID in use
we disable global pages in cr4. A command line parameter controls in
which cases PCID is being used.
As the non-XPTI case has shown not to perform better with PCID at least
on some machines the default is to use PCID only for domains subject to
XPTI.
With PCID enabled we always disable global pages. This avoids having to
either flush the complete TLB or do a cycle through all PCID values
when invalidating a single global page.
Signed-off-by: Juergen Gross <jgross@suse.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
Add some helper macros to access the address and pcid parts of cr3.
Use those helpers where appropriate.
Signed-off-by: Juergen Gross <jgross@suse.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
pv_guest_cr4_to_real_cr4() is becoming more and more complex. Convert
it from a macro to an ordinary function.
Signed-off-by: Juergen Gross <jgross@suse.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
Today cpu_info->xen_cr3 is either 0 to indicate %cr3 doesn't need to
be switched on entry to Xen, or negative for keeping the value while
indicating not to restore %cr3, or positive in case %cr3 is to be
restored.
Switch to use a flag byte instead of a negative xen_cr3 value in order
to allow %cr3 values with the high bit set in case we want to keep TLB
entries when using the PCID feature.
This reduces the number of branches in interrupt handling and results
in better performance (e.g. parallel make of the Xen hypervisor on my
system was using about 3% less system time).
Signed-off-by: Juergen Gross <jgross@suse.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
Instead of flushing the TLB from global pages when switching address
spaces with XPTI being active just disable global pages via %cr4
completely when a domain subject to XPTI is active. This avoids the
need for extra TLB flushes as loading %cr3 will remove all TLB
entries.
In order to avoid states with cr3/cr4 having inconsistent values
(e.g. global pages being activated while cr3 already specifies a XPTI
address space) move loading of the new cr4 value to write_ptbase()
(actually to switch_cr3_cr4() called by write_ptbase()).
This requires to use switch_cr3_cr4() instead of write_ptbase() when
building dom0 in order to avoid setting cr4 with cr4.smap set.
Signed-off-by: Juergen Gross <jgross@suse.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
If possible use the INVPCID instruction for flushing the TLB instead of
toggling cr4.pge for that purpose.
While at it remove the dependency on cr4.pge being required for mtrr
loading, as this will be required later anyway.
Add a command line option "invpcid" for controlling the use of
INVPCID (default to true).
Signed-off-by: Juergen Gross <jgross@suse.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
Instead of switching XPTI globally on or off add a per-domain flag for
that purpose. This allows to modify the xpti boot parameter to support
running dom0 without Meltdown mitigations. Using "xpti=no-dom0" as boot
parameter will achieve that.
Move the xpti boot parameter handling to xen/arch/x86/pv/domain.c as
it is pv-domain specific.
Signed-off-by: Juergen Gross <jgross@suse.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
Instead of having multiple places with more or less identical asm
statements just have one function doing a write to cr3.
As this function should be named write_cr3() rename the current
write_cr3() function to switch_cr3().
Suggested-by: Andrew Copper <andrew.cooper3@citrix.com>
Signed-off-by: Juergen Gross <jgross@suse.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
For mitigation of Meltdown the current L4 page table is copied to the
cpu local root page table each time a 64 bit pv guest is entered.
Copying can be avoided in cases where the guest L4 page table hasn't
been modified while running the hypervisor, e.g. when handling
interrupts or any hypercall not modifying the L4 page table or %cr3.
So add a per-cpu flag indicating whether the copying should be
performed and set that flag only when loading a new %cr3 or modifying
the L4 page table. This includes synchronization of the cpu local
root page table with other cpus, so add a special synchronization flag
for that case.
A simple performance check (compiling the hypervisor via "make -j 4")
in dom0 with 4 vcpus shows a significant improvement:
- real time drops from 112 seconds to 103 seconds
- system time drops from 142 seconds to 131 seconds
Signed-off-by: Juergen Gross <jgross@suse.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
Their need is not tied to the actual flushing of TLBs, but the ticking
of the TLB clock. Make this more obvious by folding the two invocations
into a single one in pre_flush().
Also defer the latching of CR4 in write_cr3() until after pre_flush()
(and hence implicitly until after IRQs are off), making operation
sequence the same in both cases (eliminating the theoretical risk of
pre_flush() altering CR4). This then also improves register allocation,
as the compiler doesn't need to use a callee-saved register for "cr4"
anymore.
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Reviewed-by: Andrew Cooper <andrew.cooper3@citrix.com>
Andrew Cooper
Jan Beulich
Ian Jackson
Juergen Gross
Wei Liu