PV console doesn't require Xen to be running as a shim and only
requires CONFIG_XEN_GUEST and running as a Xen guest. Update the
documentation accordingly.
Fixes: 4f6609d6a6 ("x86/guest: use PV console for Xen/Dom0 I/O")
Signed-off-by: Teddy Astie <teddy.astie@vates.tech>
Reviewed-by: Andrew Cooper <andrew.cooper3@citrix.com>
master commit: f47c2e96ab
master date: 2026-05-21 09:13:42 +0200
After initial publication, the SB-7033 / CVE-2024-36347 bulletin was updated
to list Zen5 CPUs as vulnerable. Use Fam1ah as an upper bound, and adjust the
command line documentation.
When the Zen6 (also Fam1ah processors) model numbers are known, they'll want
excluding from the range.
Fixes: 630e8875ab ("x86/ucode: Perform extra SHA2 checks on AMD Fam17h/19h microcode")
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Acked-by: Jan Beulich <jbeulich@suse.com>
Release-Acked-By: Oleksii Kurochko <oleksii.kurochko@gmail.com>
(cherry picked from commit c2529496d0)
When building a PVH dom0 Xen attempts to map all (relevant) MMIO regions
into the p2m for dom0 access. However the information Xen has about the
host memory map is limited. Xen doesn't have access to any resources
described in ACPI dynamic tables, and hence the p2m mappings provided might
not be complete.
PV doesn't suffer from this issue because a PV dom0 is capable of mapping
into it's page-tables any address not explicitly banned in d->iomem_caps.
Introduce a new command line options that allows Xen to attempt to fixup
the p2m page-faults, by creating p2m identity maps in response to p2m
page-faults.
This is aimed as a workaround to small ACPI regions Xen doesn't know about.
Note that missing large MMIO regions mapped in this way will lead to
slowness due to the VM exit processing, plus the mappings will always use
small pages.
The ultimate aim is to attempt to bring better parity with a classic PV
dom0.
Note such fixup rely on the CPU doing the access to the unpopulated
address. If the access is attempted from a device instead there's no
possible way to fixup, as IOMMU page-fault are asynchronous.
Signed-off-by: Roger Pau Monné <roger.pau@citrix.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
Acked-by: Oleksii Kurochko <oleksii.kurochko@gmail.com>
master commit: 104591f5dd
master date: 2025-03-05 10:26:46 +0100
Allow setting the used wallclock from the command line. When the option is set
to a value different than `auto` the probing is bypassed and the selected
implementation is used (as long as it's available).
The `xen` and `efi` options require being booted as a Xen guest (with Xen guest
supported built-in) or from UEFI firmware respectively.
Signed-off-by: Roger Pau Monné <roger.pau@citrix.com>
Reviewed-by: Marek Marczykowski-Górecki <marmarek@invisiblethingslab.com>
Release-Acked-by: Oleksii Kurochko<oleksii.kurochko@gmail.com>
AMD have updated the SRSO whitepaper[1] with further information. These
features exist on AMD Zen5 CPUs and are necessary for Xen to use.
The two features are in principle unrelated:
* SRSO_U/S_NO is an enumeration saying that SRSO attacks can't cross the
User(CPL3) / Supervisor(CPL<3) boundary. i.e. Xen don't need to use
IBPB-on-entry for PV64. PV32 guests are explicitly unsupported for
speculative issues, and excluded from consideration for simplicity.
* SRSO_MSR_FIX is an enumeration identifying that the BP_SPEC_REDUCE bit is
available in MSR_BP_CFG. When set, SRSO attacks can't cross the host/guest
boundary. i.e. Xen don't need to use IBPB-on-entry for HVM.
Extend ibpb_calculations() to account for these when calculating
opt_ibpb_entry_{pv,hvm} defaults. Add a `bp-spec-reduce=<bool>` option to
control the use of BP_SPEC_REDUCE, with it active by default.
Because MSR_BP_CFG is core-scoped with a race condition updating it, repurpose
amd_check_erratum_1485() into amd_check_bp_cfg() and calculate all updates at
once.
Xen also needs to to advertise SRSO_U/S_NO to guests to allow the guest kernel
to skip SRSO mitigations too:
* This is trivial for HVM guests. It is also is accurate for PV32 guests
too, but we have already excluded them from consideration, and do so again
here to simplify the policy logic.
* As written, SRSO_U/S_NO does not help for the PV64 user->kernel boundary.
However, after discussing with AMD, an implementation detail of having
BP_SPEC_REDUCE active causes the PV64 user->kernel boundary to have the
property described by SRSO_U/S_NO, so we can advertise SRSO_U/S_NO to
guests when the BP_SPEC_REDUCE precondition is met.
Finally, fix a typo in the SRSO_NO's comment.
[1] https://www.amd.com/content/dam/amd/en/documents/corporate/cr/speculative-return-stack-overflow-whitepaper.pdf
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Reviewed-by: Roger Pau Monné <roger.pau@citrix.com>
Add a new command line parameter to configure Xen cache colors.
These colors are dumped together with other coloring info.
Benchmarking the VM interrupt response time provides an estimation of
LLC usage by Xen's most latency-critical runtime task. Results on Arm
Cortex-A53 on Xilinx Zynq UltraScale+ XCZU9EG show that one color, which
reserves 64 KiB of L2, is enough to attain best responsiveness:
- Xen 1 color latency: 3.1 us
- Xen 2 color latency: 3.1 us
Since this is the most common target for Arm cache coloring, the default
amount of Xen colors is set to one.
More colors are instead very likely to be needed on processors whose L1
cache is physically-indexed and physically-tagged, such as Cortex-A57.
In such cases, coloring applies to L1 also, and there typically are two
distinct L1-colors. Therefore, reserving only one color for Xen would
senselessly partitions a cache memory that is already private, i.e.
underutilize it.
Signed-off-by: Luca Miccio <lucmiccio@gmail.com>
Signed-off-by: Marco Solieri <marco.solieri@minervasys.tech>
Signed-off-by: Carlo Nonato <carlo.nonato@minervasys.tech>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
Add a new memory page allocator that implements the cache coloring mechanism.
The allocation algorithm enforces equal frequency distribution of cache
partitions, following the coloring configuration of a domain. This allows
for an even utilization of cache sets for every domain.
Pages are stored in a color-indexed array of lists. Those lists are filled
by a simple init function which computes the color of each page.
When a domain requests a page, the allocator extracts the page from the list
with the maximum number of free pages among those that the domain can access,
given its coloring configuration.
The allocator can only handle requests of order-0 pages. This allows for
easier implementation and since cache coloring targets only embedded systems,
it's assumed not to be a major problem.
The buddy allocator must coexist with the colored one because the Xen heap
isn't colored. For this reason a new Kconfig option and a command line
parameter are added to let the user set the amount of memory reserved for
the buddy allocator. Even when cache coloring is enabled, this memory
isn't managed by the colored allocator.
Colored heap information is dumped in the dump_heap() debug-key function.
Based on original work from: Luca Miccio <lucmiccio@gmail.com>
Signed-off-by: Marco Solieri <marco.solieri@minervasys.tech>
Signed-off-by: Carlo Nonato <carlo.nonato@minervasys.tech>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
Acked-by: Michal Orzel <michal.orzel@amd.com>
Add a command line parameter to allow the user to set the coloring
configuration for Dom0.
A common configuration syntax for cache colors is introduced and
documented.
Take the opportunity to also add:
- default configuration notion.
- function to check well-formed configurations.
Direct mapping Dom0 isn't possible when coloring is enabled, so
CDF_directmap flag is removed when creating it.
Based on original work from: Luca Miccio <lucmiccio@gmail.com>
Signed-off-by: Carlo Nonato <carlo.nonato@minervasys.tech>
Signed-off-by: Marco Solieri <marco.solieri@minervasys.tech>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
Reviewed-by: Michal Orzel <michal.orzel@amd.com>
LLC coloring needs to know the last level cache layout in order to make the
best use of it. This can be probed by inspecting the CLIDR_EL1 register,
so the Last Level is defined as the last level visible by this register.
Note that this excludes system caches in some platforms.
Static memory allocation and cache coloring are incompatible because static
memory can't be guaranteed to use only colors assigned to the domain.
Panic during DomUs creation when both are enabled.
Based on original work from: Luca Miccio <lucmiccio@gmail.com>
Signed-off-by: Carlo Nonato <carlo.nonato@minervasys.tech>
Signed-off-by: Marco Solieri <marco.solieri@minervasys.tech>
Reviewed-by: Michal Orzel <michal.orzel@amd.com>
Acked-by: Jan Beulich <jbeulich@suse.com>
Last Level Cache (LLC) coloring allows to partition the cache in smaller
chunks called cache colors.
Since not all architectures can actually implement it, add a HAS_LLC_COLORING
Kconfig option.
LLC_COLORS_ORDER Kconfig option has a range maximum of 10 (2^10 = 1024)
because that's the number of colors that fit in a 4 KiB page when integers
are 4 bytes long.
LLC colors are a property of the domain, so struct domain has to be extended.
Based on original work from: Luca Miccio <lucmiccio@gmail.com>
Signed-off-by: Carlo Nonato <carlo.nonato@minervasys.tech>
Signed-off-by: Marco Solieri <marco.solieri@minervasys.tech>
Acked-by: Michal Orzel <michal.orzel@amd.com>
... as a shorthand for setting both suboptions at once. Currently, an admin
needs to pass cet=no-shstk,no-ibt to turn both off, where cet=0 is a better
option.
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Reviewed-by: Roger Pau Monné <roger.pau@citrix.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
With the introduction of mixed x2APIC mode (using cluster addressing for
IPIs and physical for external interrupts) the use of pure cluster mode
doesn't have any benefit.
Remove the mode itself, leaving only the code required for logical
addressing when sending IPIs.
Resolves: https://gitlab.com/xen-project/xen/-/issues/189
Signed-off-by: Matthew Barnes <matthew.barnes@cloud.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
Acked-by: Oleksii Kurochko <oleksii.kurochko@gmail.com>
All versions of Cortex-A53 cores are affected by the speculative
AT instruction erratum, as mentioned in the Cortex-A53 Revision r0
SDEN v21 documentation.
Enabled ARM64_WORKAROUND_AT_SPECULATE for all versions of Cortex-A53
cores, to avoid corrupting the TLB if performing a speculative AT
instruction during a guest context switch.
Signed-off-by: Andrei Cherechesu <andrei.cherechesu@nxp.com>
Acked-by: Julien Grall <jgrall@amazon.com>
Pass xen-ucode flags to do low level checks on microcode version and utilize
it to allow for microcode downgrade or reapply the same version of the
microcode. ucode_force is required to be passed to a low level Intel and AMD
for version checks to be done.
While adding ucode_force, opt_ucode_allow_same was removed. Remove
opt_ucode_allow_same from documentation.
Update CHANGELOG.md for opt_ucode_allow_same removal.
Signed-off-by: Fouad Hilly <fouad.hilly@cloud.com>
Acked-by: Andrew Cooper <andrew.cooper3@citrix.com>
Transactional Synchronization Extensions are supported on certain Intel's
CPUs only, hence can be put under CONFIG_INTEL build option.
The whole TSX support, even if supported by CPU, may need to be disabled via
options, by microcode or through spec-ctrl, depending on a set of specific
conditions. To make sure nothing gets accidentally runtime-broken all
modifications of global TSX configuration variables is secured by #ifdef's,
while variables themselves redefined to 0, so that ones can't mistakenly be
written to.
Signed-off-by: Sergiy Kibrik <Sergiy_Kibrik@epam.com>
Acked-by: Jan Beulich <jbeulich@suse.com>
To change the cd-rom medium, libxl will:
- QMP eject the medium from QEMU
- block-detach the old PV disk
- block-attach the new PV disk
- QMP change the medium to the new PV disk by fdset-id
The QMP code is reused, and remove and attach are implemented here.
The stubdom must internally handle adding /dev/xvdc to the appropriate
fdset. libxl in dom0 doesn't see the result of adding to the fdset as
that is internal to the stubdom, but the fdset's opaque fields will be
set to stub-devid:$devid, so libxl can identify it. $devid is common
between the stubdom and libxl, so it can be identified on both side.
The stubdom will name the device xvdY regardless of the guest name hdY,
sdY, or xvdY, but the stubdom will be assigned the same devid
facilitating lookup. Because the stubdom add-fd call is asynchronous,
libxl needs to poll query-fdsets to identify when add-fd has completed.
For cd-eject, we still need to attach the empty vbd. This is necessary
since xenstore is used to determine that hdc exists. Otherwise after
eject, hdc would be gone and the cd-insert would fail to find the drive
to insert new media.
Signed-off-by: Jason Andryuk <jandryuk@gmail.com>
Signed-off-by: Jason Andryuk <jason.andryuk@amd.com>
Reviewed-by: Anthony PERARD <anthony.perard@vates.tech>
PVH in particular has no (externally visible) notion of pIRQ-s. Mention
that in the description of the respective command line option and have
arch_hwdom_irqs() also reflect this (thus suppressing the log message
there as well, as being pretty meaningless in this case anyway).
Suggested-by: Roger Pau Monné <roger.pau@citrix.com>
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Reviewed-by: Roger Pau Monné <roger.pau@citrix.com>
Release-Acked-by: Oleksii Kurochko <oleksii.kurochko@gmail.com>
PHYSDEVOP_pirq_eoi_gmfn_v<N> accepting just a single GFN implies that no
more than 32k pIRQ-s can be used by a domain on x86. Document this upper
bound.
To also enforce the limit, (ab)use both arch_hwdom_irqs() (changing its
parameter type) and setup_system_domains(). This is primarily to avoid
exposing the two static variables or introducing yet further arch hooks.
While touching arch_hwdom_irqs() also mark it hwdom-init.
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Acked-by: Roger Pau Monné <roger.pau@citrix.com>
Release-Acked-by: Oleksii Kurochko <oleksii.kurochko@gmail.com>
Currently, users are allowed to map static shared memory in a
non-direct-mapped way for direct-mapped domains. This can lead to
clashing of guest memory spaces. Also, the current extended region
finding logic only removes the host physical addresses of the
static shared memory areas for direct-mapped domains, which may be
inconsistent with the guest memory map if users map the static
shared memory in a non-direct-mapped way. This will lead to incorrect
extended region calculation results.
To make things easier, add restriction that static shared memory
should also be direct-mapped for direct-mapped domains. Check the
host physical address to be matched with guest physical address when
parsing the device tree. Document this restriction in the doc.
Signed-off-by: Henry Wang <xin.wang2@amd.com>
Signed-off-by: Stefano Stabellini <stefano.stabellini@amd.com>
Acked-by: Michal Orzel <michal.orzel@amd.com>
This commit describe the new scenario where host address is not provided
in "xen,shared-mem" property and a new example is added to the page to
explain in details.
Take the occasion to fix some typos in the page.
Signed-off-by: Penny Zheng <penny.zheng@arm.com>
Signed-off-by: Luca Fancellu <luca.fancellu@arm.com>
Reviewed-by: Michal Orzel <michal.orzel@amd.com>
There are some use cases in which the dom0less domUs need to have
the XEN_DOMCTL_CDF_iommu set at the domain construction time. For
example, the dynamic dtbo feature allows the domain to be assigned
a device that is behind the IOMMU at runtime. For these use cases,
we need to have a way to specify the domain will need the IOMMU
mapping at domain construction time.
Introduce a "passthrough" DT property for Dom0less DomUs following
the same entry as the xl.cfg. Currently only provide two options,
i.e. "enable" and "disable". Set the XEN_DOMCTL_CDF_iommu at domain
construction time based on the property.
Signed-off-by: Henry Wang <xin.wang2@amd.com>
Reviewed-by: Julien Grall <jgrall@amazon.com>
By default there's already no use for this when we run in shim mode.
Plus there may also be a need to suppress the probing in case of issues
with it. Before introducing further port alias probing, introduce a
command line option allowing to bypass it, default it to on when in shim
mode, and gate RTC/CMOS port alias probing on it.
Requested-by: Roger Pau Monné <roger.pau@citrix.com>
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Reviewed-by: Jason Andryuk <jason.andryuk@amd.com>
Since commit 6cd046c501 ("xen/arm: Enlarge identity map space to 10TB")
Xen can be loaded below 10 TiB. Update docs accordingly.
Take the opportunity to update stale links to Linux docs.
Signed-off-by: Michal Orzel <michal.orzel@amd.com>
Reviewed-by: Luca Fancellu <luca.fancellu@arm.com>
Out of an abudnance of caution, implement the long loop too, and allowing for
it to be opted in to.
This is part of XSA-456 / CVE-2024-2201.
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Acked-by: Roger Pau Monné <roger.pau@citrix.com>
In the absence of BHI_DIS_S, mitigating Native-BHI requires the use of a
software sequence.
Introduce a new bhb-seq= option to select between avaialble sequences and
bhb-entry= to control the per-PV/HVM actions like we have for other blocks.
Activate the short sequence by default for PV and HVM guests on affected
hardware if BHI_DIS_S isn't present.
This is part of XSA-456 / CVE-2024-2201.
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Acked-by: Roger Pau Monné <roger.pau@citrix.com>
Introduce a "bhi-dis-s" boolean to match the other options we have for
MSR_SPEC_CTRL values. Also introduce bhi_calculations().
Use BHI_DIS_S whenever possible.
Guests which are levelled to be migration compatible with older CPUs can't see
BHI_DIS_S, and Xen must fill in the difference to make the guest safe. Use
the virt MSR_SPEC_CTRL infrastructure to force BHI_DIS_S behind the guest's
back.
This is part of XSA-456 / CVE-2024-2201.
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Acked-by: Roger Pau Monné <roger.pau@citrix.com>
Currently, RISC-V requires two extensions: _zbb and _zihintpause.
This patch introduces a compiler check to check if these extensions
are supported.
Additionally, it introduces the riscv/booting.txt file, which contains
information about the extensions that should be supported by the platform.
In the future, a feature will be introduced to check whether an extension
is supported at runtime.
However, this feature requires functionality for parsing device tree
source (DTS), which is not yet available.
Signed-off-by: Oleksii Kurochko <oleksii.kurochko@gmail.com>
Acked-by: Jan Beulich <jbeulich@suse.com>
There can be situations when the registers cannot be emulated to their full
functionality. This can be due to the complexity involved. In such cases, one
can emulate those registers as RAZ/WI for example. We call them as partial
emulation.
Some registers are non-optional and as such there is nothing preventing an OS
from accessing them.
Instead of injecting undefined exception (which might result in crashing a
guest), one may want to prefer a partial emulation to let the guest running
(in some cases accepting the fact that it might result in unwanted behavior).
A suitable example of this (as seen in subsequent patches) is emulation of
DBGDTRTX_EL0 (on Arm64) and DBGDTRTXINT(on Arm32). These non-optional
registers can be emulated as RAZ/WI and they can be enclosed within
CONFIG_PARTIAL_EMULATION.
Further, "partial-emulation" command line option allows us to
enable/disable partial emulation at run time. While CONFIG_PARTIAL_EMULATION
enables support for partial emulation at compile time (i.e. adds code for
partial emulation), this option may be enabled or disabled by Yocto or other
build systems. However if the build system turns this option on, users
can use scripts like Imagebuilder to generate uboot-script which will append
"partial-emulation=true" to xen command line to turn on the partial emulation.
Thus, it helps to avoid rebuilding xen.
By default, "CONFIG_PARTIAL_EMULATION=y" and "partial-emulation=false".
This is done so that Xen supports partial emulation. However, customers are
fully aware when they enable partial emulation. It's important to note that
enabling such support might result in unwanted/non-spec compliant behavior.
Added a note in SUPPORT.md to clarify the security support for partial
emulation.
Signed-off-by: Ayan Kumar Halder <ayan.kumar.halder@amd.com>
Signed-off-by: Michal Orzel <michal.orzel@amd.com>
Acked-by: Julien Grall <jgrall@amazon.com>
... by way of a new arch-selectable Kconfig control.
Note that some smaller pieces of code are left without #ifdef, to keep
things better readable. Hence items like ECS_PIRQ, nr_static_irqs, or
domain_pirq_to_irq() remain uniformly.
As to XEN_DOMCTL_irq_permission - this, despite having a uniformly
available wrapper in libxc, is unused on Arm: libxl bypasses those
calls, and the Python and OCamL bindings have no users at all.
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Acked-by: Julien Grall <jgrall@amazon.com>
Introduce a new Kconfig option to block speculation into lock protected
critical regions. The Kconfig option is enabled by default, but the mitigation
won't be engaged unless it's explicitly enabled in the command line using
`spec-ctrl=lock-harden`.
Convert the spinlock acquire macros into always-inline functions, and introduce
a speculation barrier after the lock has been taken. Note the speculation
barrier is not placed inside the implementation of the spin lock functions, as
to prevent speculation from falling through the call to the lock functions
resulting in the barrier also being skipped.
trylock variants are protected using a construct akin to the existing
evaluate_nospec().
This patch only implements the speculation barrier for x86.
Note spin locks are the only locking primitive taken care in this change,
further locking primitives will be adjusted by separate changes.
This is part of XSA-453 / CVE-2024-2193
Signed-off-by: Roger Pau Monné <roger.pau@citrix.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
VERW is going to be used for a 3rd purpose, and the existing nomenclature
didn't survive the Stale MMIO issues terribly well.
Rename the command line option from `md-clear=` to `verw=`. This is more
consistent with other options which tend to be named based on what they're
doing, not which feature enumeration they use behind the scenes. Retain
`md-clear=` as a deprecated alias.
Rename opt_md_clear_{pv,hvm} and opt_fb_clear_mmio to opt_verw_{pv,hvm,mmio},
which has a side effect of making spec_ctrl_init_domain() rather clearer to
follow.
No functional change.
This is part of XSA-452 / CVE-2023-28746.
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
Attempt to provide a more helpful error message when the user attempts to set
spec-ctrl=bti-thunk option but the support is build-time disabled.
While there also adjust the command line documentation to mention
CONFIG_INDIRECT_THUNK instead of INDIRECT_THUNK.
Reported-by: Andrew Cooper <andrew.cooper3@citrix.com>
Signed-off-by: Roger Pau Monné <roger.pau@citrix.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
In 13y of working on Xen, I've never seen seen it used. The implementation
was introduced (commit b69f92f301, Jul 28 2004) with known issues such as:
/* Resuming after we've stopped used to work, but more through luck
than any actual intention. It doesn't at the moment. */
which appear to have gone unfixed for the 20 years since.
Nowadays there are more robust ways of inspecting crashed state, such as a
kexec crash kernel, or running Xen in a VM.
This will allow us to clean up some hooks around the codebase which are
proving awkward for other tasks.
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Acked-by: Jan Beulich <jbeulich@suse.com>
These 3 Kconfig docs were imported from Linux erroneously. They are
GPL-2.0-only in Linux, but have no SPDX tag and were placed in such a way to
be included by the blanket statement saying that all RST files are CC-BY-4.0.
We should not be carrying a shadow copy of these docs. They aren't even wired
into our Sphinx docs, and anyone wanting to refer to Kconfig docs is going to
look at the Linux docs anyway. These, and more docs can be found at:
https://www.kernel.org/doc/html/latest/kbuild/
which also have corrections vs the snapshot we took.
Fixes: f80fe2b34f ("xen: Update Kconfig to Linux v5.4")
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Acked-by: Jan Beulich <jbeulich@suse.com>
Acked-by: Stefano Stabellini <sstabellini@kernel.org>
Release-acked-by: Henry Wang <Henry.Wang@arm.com>
The current implementation of x2APIC requires to either use Cluster Logical or
Physical mode for all interrupts. However the selection of Physical vs Logical
is not done at APIC setup, an APIC can be addressed both in Physical or Logical
destination modes concurrently.
Introduce a new x2APIC mode called Mixed, which uses Logical Cluster mode for
IPIs, and Physical mode for external interrupts, thus attempting to use the
best method for each interrupt type.
Using Physical mode for external interrupts allows more vectors to be used, and
interrupt balancing to be more accurate.
Using Logical Cluster mode for IPIs allows fewer accesses to the ICR register
when sending those, as multiple CPUs can be targeted with a single ICR register
write.
A simple test calling flush_tlb_all() 10000 times on a tight loop on AMD EPYC
9754 with 512 CPUs gives the following figures in nano seconds:
x mixed
+ phys
* cluster
N Min Max Median Avg Stddev
x 25 3.5131328e+08 3.5716441e+08 3.5410987e+08 3.5432659e+08 1566737.4
+ 12 1.231082e+09 1.238824e+09 1.2370528e+09 1.2357981e+09 2853892.9
Difference at 95.0% confidence
8.81472e+08 +/- 1.46849e+06
248.774% +/- 0.96566%
(Student's t, pooled s = 2.05985e+06)
* 11 3.5099276e+08 3.5561459e+08 3.5461234e+08 3.5415668e+08 1415071.9
No difference proven at 95.0% confidence
So Mixed has no difference when compared to Cluster mode, and Physical mode is
248% slower when compared to either Mixed or Cluster modes with a 95%
confidence.
Note that Xen uses Cluster mode by default, and hence is already using the
fastest way for IPI delivery at the cost of reducing the amount of vectors
available system-wide.
Make the newly introduced mode the default one.
Note the printing of the APIC addressing mode done in connect_bsp_APIC() has
been removed, as with the newly introduced mixed mode this would require more
fine grained printing, or else would be incorrect. The addressing mode can
already be derived from the APIC driver in use, which is printed by different
helpers.
Suggested-by: Andrew Cooper <andrew.cooper3@citrix.com>
Signed-off-by: Roger Pau Monné <roger.pau@citrix.com>
Reviewed-by: Andrew Cooper <andrew.cooper3@citrix.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
Acked-by: Henry Wang <Henry.Wang@arm.com>
Make the command line docs match the actual implementation, and state that the
default behaviour is selected at compile time.
Fixes: 980d6acf15 ("IOMMU: make DMA containment of quarantined devices optional")
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Reviewed-by: Roger Pau Monné <roger.pau@citrix.com>
Release-acked-by: Henry Wang <Henry.Wang@arm.com>
In commit 9d267c049d ("xen/arm64: Rework the memory layout"),
we decided to require Xen to be loaded below 2 TiB to simplify
the logic to enable the MMU. The limit was decided based on
how known platform boot plus some slack.
We had a recent report that this is not sufficient on the AVA
platform with a old firmware [1]. But the restriction is not
going to change in Xen 4.18. So document the limit clearly
in docs/misc/arm/booting.txt.
[1] https://lore.kernel.org/20231013122658.1270506-3-leo.yan@linaro.org
Signed-off-by: Julien Grall <jgrall@amazon.com>
Reviewed-by: Michal Orzel <michal.orzel@amd.com>
Reviewed-by: Bertrand Marquis <bertrand.marquis@arm.com>
Release-acked-by: Henry Wang <Henry.Wang@arm.com>
With the addition of per-device quarantine page tables the sink page is now
exclusive for each device, and thus writable. Update the documentation to
reflect the current implementation.
Fixes: 14dd241aad ('IOMMU/x86: use per-device page tables for quarantining')
Signed-off-by: Roger Pau Monné <roger.pau@citrix.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
Release-acked-by: Henry Wang <Henry.Wang@arm.com>
The credit scheduler tries as hard as it can to ensure that it always
runs scheduling units with positive credit (PRI_TS_UNDER) before
running those with negative credit (PRI_TS_OVER). If the next
runnable scheduling unit is of priority OVER, it will always run the
load balancer, which will scour the system looking for another
scheduling unit of the UNDER priority.
Unfortunately, as the number of cores on a system has grown, the cost
of the work-stealing algorithm has dramatically increased; a recent
trace on a system with 128 cores showed this taking over 50
microseconds.
Add a parameter, load_balance_ratelimit, to limit the frequency of
load balance operations on a given pcpu. Default this to 1
millisecond.
Invert the load balancing conditional to make it more clear, and line
up more closely with the comment above it.
Overall it might be cleaner to have the last_load_balance checking
happen inside csched_load_balance(), but that would require either
passing both now and spc into the function, or looking them up again;
both of which seemed to be worse than simply checking and setting the
values before calling it.
On a system with a vcpu:pcpu ratio of 2:1, running Windows guests
(which will end up calling YIELD during spinlock contention), this
patch increased performance significantly.
Signed-off-by: George Dunlap <george.dunlap@cloud.com>
Reviewed-by: Juergen Gross <jgross@suse.com>
Release-acked-by: Henry Wang <Henry.Wang@arm.com>
In the Zen1 microarchitecure, there is one divider in the pipeline which
services uops from both threads. In the case of #DE, the latched result from
the previous DIV to execute will be forwarded speculatively.
This is an interesting covert channel that allows two threads to communicate
without any system calls. In also allows userspace to obtain the result of
the most recent DIV instruction executed (even speculatively) in the core,
which can be from a higher privilege context.
Scrub the result from the divider by executing a non-faulting divide. This
needs performing on the exit-to-guest paths, and ist_exit-to-Xen.
Alternatives in IST context is believed safe now that it's done in NMI
context.
This is XSA-439 / CVE-2023-20588.
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
Teddy Astie