When moving around a page via XENMAPSPACE_gmfn_range, deferring the TLB
flush for the "from" GFN range requires that the page remains allocated
to the guest until the TLB flush has actually occurred. Otherwise a
parallel hypercall to remove the page would only flush the TLB for the
GFN it has been moved to, but not the one is was mapped at originally.
This is part of XSA-346.
Fixes: cf95b2a9fd ("iommu: Introduce per cpu flag (iommu_dont_flush_iotlb) to avoid unnecessary iotlb... ")
Reported-by: Julien Grall <jgrall@amazon.com>
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Acked-by: Julien Grall <jgrall@amazon.com>
master commit: 5777a3742d
master date: 2020-10-20 14:21:32 +0200
Deferring flushes to a single, wide range one - as is done when
handling XENMAPSPACE_gmfn_range - is okay only as long as
pages don't get freed ahead of the eventual flush. While the only
function setting the flag (xenmem_add_to_physmap()) suggests by its name
that it's only mapping new entries, in reality the way
xenmem_add_to_physmap_one() works means an unmap would happen not only
for the page being moved (but not freed) but, if the destination GFN is
populated, also for the page being displaced from that GFN. Collapsing
the two flushes for this GFN into just one (end even more so deferring
it to a batched invocation) is not correct.
This is part of XSA-346.
Fixes: cf95b2a9fd ("iommu: Introduce per cpu flag (iommu_dont_flush_iotlb) to avoid unnecessary iotlb... ")
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Reviewed-by: Paul Durrant <paul@xen.org>
Acked-by: Julien Grall <jgrall@amazon.com>
master commit: dea460d869
master date: 2020-10-20 14:21:09 +0200
map_pages_to_xen will attempt to coalesce mappings into 2MiB and 1GiB
superpages if possible, to maximize TLB efficiency. This means both
replacing superpage entries with smaller entries, and replacing
smaller entries with superpages.
Unfortunately, while some potential races are handled correctly,
others are not. These include:
1. When one processor modifies a sub-superpage mapping while another
processor replaces the entire range with a superpage.
Take the following example:
Suppose L3[N] points to L2. And suppose we have two processors, A and
B.
* A walks the pagetables, get a pointer to L2.
* B replaces L3[N] with a 1GiB mapping.
* B Frees L2
* A writes L2[M] #
This is race exacerbated by the fact that virt_to_xen_l[21]e doesn't
handle higher-level superpages properly: If you call virt_xen_to_l2e
on a virtual address within an L3 superpage, you'll either hit a BUG()
(most likely), or get a pointer into the middle of a data page; same
with virt_xen_to_l1 on a virtual address within either an L3 or L2
superpage.
So take the following example:
* A reads pl3e and discovers it to point to an L2.
* B replaces L3[N] with a 1GiB mapping
* A calls virt_to_xen_l2e() and hits the BUG_ON() #
2. When two processors simultaneously try to replace a sub-superpage
mapping with a superpage mapping.
Take the following example:
Suppose L3[N] points to L2. And suppose we have two processors, A and B,
both trying to replace L3[N] with a superpage.
* A walks the pagetables, get a pointer to pl3e, and takes a copy ol3e pointing to L2.
* B walks the pagetables, gets a pointre to pl3e, and takes a copy ol3e pointing to L2.
* A writes the new value into L3[N]
* B writes the new value into L3[N]
* A recursively frees all the L1's under L2, then frees L2
* B recursively double-frees all the L1's under L2, then double-frees L2 #
Fix this by grabbing a lock for the entirety of the mapping update
operation.
Rather than grabbing map_pgdir_lock for the entire operation, however,
repurpose the PGT_locked bit from L3's page->type_info as a lock.
This means that rather than locking the entire address space, we
"only" lock a single 512GiB chunk of hypervisor address space at a
time.
There was a proposal for a lock-and-reverify approach, where we walk
the pagetables to the point where we decide what to do; then grab the
map_pgdir_lock, re-verify the information we collected without the
lock, and finally make the change (starting over again if anything had
changed). Without being able to guarantee that the L2 table wasn't
freed, however, that means every read would need to be considered
potentially unsafe. Thinking carefully about that is probably
something that wants to be done on public, not under time pressure.
This is part of XSA-345.
Reported-by: Hongyan Xia <hongyxia@amazon.com>
Signed-off-by: Hongyan Xia <hongyxia@amazon.com>
Signed-off-by: George Dunlap <george.dunlap@citrix.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
master commit: 1ce75e99d7
master date: 2020-10-20 14:20:19 +0200
xmalloc() & Co may not be called with IRQs off, or else check_lock()
will have its assertion trigger about locks getting acquired
inconsistently. Re-arranging the locking in evtchn_send() doesn't seem
very reasonable, especially since the per-channel lock was introduced to
avoid acquiring the per-domain event lock on the send paths. Issue a
second call to xsm_evtchn_send() instead, before acquiring the lock, to
give XSM / Flask a chance to pre-allocate whatever it may need.
As these nodes are used merely for caching earlier decisions' results,
allocate just one node in AVC code despite two potentially being needed.
Things will merely be not as performant if a second allocation was
wanted, just like when the pre-allocation fails.
Fixes: c0ddc86348 ("evtchn: convert per-channel lock to be IRQ-safe")
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Tested-by: Jason Andryuk <jandryuk@gmail.com>
Acked-by: Julien Grall <jgrall@amazon.com>
Reviewed-by: Jason Andryuk <jandryuk@gmail.com>
master commit: 52e1fc47ab
master date: 2020-10-02 08:36:21 +0200
Doing this just in hvm_emulate_one_insn() is not enough.
hvm_ud_intercept() and hvm_emulate_one_vm_event() can get invoked for
insns requiring one or more continuations, and at least in principle
hvm_emulate_one_mmio() could, too. Without proper setting of the field,
handle_hvm_io_completion() will do nothing completion-wise, and in
particular the missing re-invocation of the insn emulation paths will
lead to emulation caching not getting disabled in due course, causing
the ASSERT() in {svm,vmx}_vmenter_helper() to trigger.
Reported-by: Don Slutz <don.slutz@gmail.com>
Similar considerations go for the clearing of vio->mmio_access, which
gets moved as well.
Additionally all updating of vio->mmio_* now gets done dependent upon
the new completion value, rather than hvm_ioreq_needs_completion()'s
return value. This is because it is the completion chosen which controls
what path will be taken when handling the completion, not the simple
boolean return value. In particular, PIO completion doesn't involve
going through the insn emulator, and hence emulator state ought to get
cleared early (or it won't get cleared at all).
The new logic, besides allowing for a caller override for the
continuation type to be set (for VMX real mode emulation), will also
avoid setting an MMIO completion when a simpler PIO one will do. This
is a minor optimization only as a side effect - the behavior is strictly
needed at least for hvm_ud_intercept(), as only memory accesses can
successfully complete through handle_mmio(). Care of course needs to be
taken to correctly deal with "mixed" insns (doing both MMIO and PIO at
the same time, i.e. INS/OUTS). For this, hvmemul_validate() now latches
whether the insn being emulated is a memory access, as this information
is no longer easily available at the point where we want to consume it.
Note that the presence of non-NULL .validate fields in the two ops
structures in hvm_emulate_one_mmio() was really necessary even before
the changes here: Without this, passing non-NULL as middle argument to
hvm_emulate_init_once() is meaningless.
The restrictions on when the #UD intercept gets actually enabled are why
it was decided that this is not a security issue:
- the "hvm_fep" option to enable its use is a debugging option only,
- for the cross-vendor case is considered experimental, even if
unfortunately SUPPORT.md doesn't have an explicit statement about
this.
The other two affected functions are
- hvm_emulate_one_vm_event(), used for introspection,
- hvm_emulate_one_mmio(), used for Dom0 only,
which aren't qualifying this as needing an XSA either.
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Tested-by: Don Slutz <don.slutz@gmail.com>
Reviewed-by: Paul Durrant <paul@xen.org>
Reviewed-by: Kevin Tian <kevin.tian@intel.com>
master commit: b807cfe954
master date: 2020-09-15 10:19:33 +0200
Guest kernel does need to know in some cases where the tables are located
to treat these regions properly. One example is kexec process where
the first kernel needs to pass ACPI region locations to the second
kernel which is now a requirement in Linux after 02a3e3cdb7f12 ("x86/boot:
Parse SRAT table and count immovable memory regions") in order for kexec
transition to actually work.
That commit introduced accesses to XSDT and SRAT while the second kernel
is still using kexec transition tables. The transition tables do not have
e820 "reserved" regions mapped where those tables are located currently
in a Xen guest. Instead "ACPI data" regions are mapped with the transition
tables that was introduced by the following commit 6bbeb276b7 ("x86/kexec:
Add the EFI system tables and ACPI tables to the ident map").
Reserve 1MB (out of 16MB currently available) right after ACPI info page for
ACPI tables exclusively but populate this region on demand and only indicate
populated memory as "ACPI data" since according to ACPI spec that memory is
reclaimable by the guest if necessary. That is close to how we treat
the same ACPI data in PVH guests. 1MB should be enough for now but could be
later extended if required.
Signed-off-by: Igor Druzhinin <igor.druzhinin@citrix.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
master commit: 8efa46516c
master date: 2020-09-09 17:56:13 +0200
Like for evtchn_destroy() looping over all possible event channels to
close them can take a significant amount of time. Unlike done there, we
can't alter domain properties (i.e. d->valid_evtchns) here. Borrow, in a
lightweight form, the paging domctl continuation concept, redirecting
the continuations to different sub-ops. Just like there this is to be
able to allow for predictable overall results of the involved sub-ops:
Racing requests should either complete or be refused.
Note that a domain can't interfere with an already started (by a remote
domain) reset, due to being paused. It can prevent a remote reset from
happening by leaving a reset unfinished, but that's only going to affect
itself.
This is part of XSA-344.
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Acked-by: Julien Grall <jgrall@amazon.com>
Reviewed-by: Stefano Stabellini <sstabellini@kernel.org>
Especially closing of fully established interdomain channels can take
quite some time, due to the locking involved. Therefore we shouldn't
assume we can clean up still active ports all in one go. Besides adding
the necessary preemption check, also avoid pointlessly starting from
(or now really ending at) 0; 1 is the lowest numbered port which may
need closing.
Since we're now reducing ->valid_evtchns, free_xen_event_channel(),
and (at least to be on the safe side) notify_via_xen_event_channel()
need to cope with attempts to close / unbind from / send through already
closed (and no longer valid, as per port_is_valid()) ports.
This is part of XSA-344.
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Acked-by: Julien Grall <jgrall@amazon.com>
Reviewed-by: Stefano Stabellini <sstabellini@kernel.org>
Neither d->evtchn_port_ops nor max_evtchns(d) may be used in an entirely
lock-less manner, as both may change by a racing evtchn_reset(). In the
common case, at least one of the domain's event lock or the per-channel
lock needs to be held. In the specific case of the inter-domain sending
by evtchn_send() and notify_via_xen_event_channel() holding the other
side's per-channel lock is sufficient, as the channel can't change state
without both per-channel locks held. Without such a channel changing
state, evtchn_reset() can't complete successfully.
Lock-free accesses continue to be permitted for the shim (calling some
otherwise internal event channel functions), as this happens while the
domain is in effectively single-threaded mode. Special care also needs
taking for the shim's marking of in-use ports as ECS_RESERVED (allowing
use of such ports in the shim case is okay because switching into and
hence also out of FIFO mode is impossihble there).
As a side effect, certain operations on Xen bound event channels which
were mistakenly permitted so far (e.g. unmask or poll) will be refused
now.
This is part of XSA-343.
Reported-by: Julien Grall <jgrall@amazon.com>
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Acked-by: Julien Grall <jgrall@amazon.com>
... in order for send_guest_{global,vcpu}_virq() to be able to make use
of it.
This is part of XSA-343.
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Acked-by: Julien Grall <jgrall@amazon.com>
While the function closes all ports, it does so without holding any
lock, and hence racing requests may be issued causing new ports to get
opened. This would have been problematic in particular if such a newly
opened port had a port number above the new implementation limit (i.e.
when switching from FIFO to 2-level) after the reset, as prior to
"evtchn: relax port_is_valid()" this could have led to e.g.
evtchn_close()'s "BUG_ON(!port_is_valid(d2, port2))" to trigger.
Introduce a counter of active ports and check that it's (still) no
larger then the number of Xen internally used ones after obtaining the
necessary lock in evtchn_reset().
As to the access model of the new {active,xen}_evtchns fields - while
all writes get done using write_atomic(), reads ought to use
read_atomic() only when outside of a suitably locked region.
Note that as of now evtchn_bind_virq() and evtchn_bind_ipi() don't have
a need to call check_free_port().
This is part of XSA-343.
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Reviewed-by: Stefano Stabellini <sstabellini@kernel.org>
Reviewed-by: Julien Grall <jgrall@amazon.com>
The recording of d->max_evtchns in evtchn_2l_init(), in particular with
the limited set of callers of the function, is insufficient. Neither for
PV nor for HVM guests the bitness is known at domain_create() time, yet
the upper bound in 2-level mode depends upon guest bitness. Recording
too high a limit "allows" x86 32-bit domains to open not properly usable
event channels, management of which (inside Xen) would then result in
corruption of the shared info and vCPU info structures.
Keep the upper limit dynamic for the 2-level case, introducing a helper
function to retrieve the effective limit. This helper is now supposed to
be private to the event channel code. The used in do_poll() and
domain_dump_evtchn_info() weren't consistent with port uses elsewhere
and hence get switched to port_is_valid().
Furthermore FIFO mode's setup_ports() gets adjusted to loop only up to
the prior ABI limit, rather than all the way up to the new one.
Finally a word on the change to do_poll(): Accessing ->max_evtchns
without holding a suitable lock was never safe, as it as well as
->evtchn_port_ops may change behind do_poll()'s back. Using
port_is_valid() instead widens some the window for potential abuse,
until we've dealt with the race altogether (see XSA-343).
This is XSA-342.
Reported-by: Julien Grall <jgrall@amazon.com>
Fixes: 48974e6ce5 ("evtchn: use a per-domain variable for the max number of event channels")
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Reviewed-by: Stefano Stabellini <sstabellini@kernel.org>
Reviewed-by: Julien Grall <jgrall@amazon.com>
While the allocation of a bucket is always performed with the per-domain
lock, the bucket may be accessed without the lock taken (for instance, see
evtchn_send()).
Instead such sites relies on port_is_valid() to return a non-zero value
when the port has a struct evtchn associated to it. The function will
mostly check whether the port is less than d->valid_evtchns as all the
buckets/event channels should be allocated up to that point.
Unfortunately a compiler is free to re-order the assignment in
evtchn_allocate_port() so it would be possible to have d->valid_evtchns
updated before the new bucket has finish to allocate.
Additionally on Arm, even if this was compiled "correctly", the
processor can still re-order the memory access.
Add a write memory barrier in the allocation side and a read memory
barrier when the port is valid to prevent any re-ordering issue.
This is XSA-340.
Reported-by: Julien Grall <jgrall@amazon.com>
Signed-off-by: Julien Grall <jgrall@amazon.com>
Reviewed-by: Stefano Stabellini <sstabellini@kernel.org>
There is at least one path (SYSENTER with NT set, Xen converts to #GP) which
ends up injecting the #GP fault twice, first in compat_sysenter(), and then a
second time in compat_test_all_events(), due to the stale TBF_EXCEPTION left
in TRAPBOUNCE_flags.
The guest kernel sees the second fault first, which is a kernel level #GP
pointing at the head of the #GP handler, and is therefore a userspace
trigger-able DoS.
This particular bug has bitten us several times before, so rearrange
{compat_,}create_bounce_frame() to clobber TRAPBOUNCE on success, rather than
leaving this task to one area of code which isn't used uniformly.
Other scenarios which might result in a double injection (e.g. two calls
directly to compat_create_bounce_frame) will now crash the guest, which is far
more obvious than letting the kernel run with corrupt state.
This is XSA-339
Fixes: fdac951560 ("x86: clear EFLAGS.NT in SYSENTER entry path")
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
To avoid ports potentially becoming invalid behind the back of certain
other functions (due to ->max_evtchn shrinking) because of
- a guest invoking evtchn_reset() and from a 2nd vCPU opening new
channels in parallel (see also XSA-343),
- alloc_unbound_xen_event_channel() produced channels living above the
2-level range (see also XSA-342),
drop the max_evtchns check from port_is_valid(). For a port for which
the function once returned "true", the returned value may not turn into
"false" later on. The function's result may only depend on bounds which
can only ever grow (which is the case for d->valid_evtchns).
This also eliminates a false sense of safety, utilized by some of the
users (see again XSA-343): Without a suitable lock held, d->max_evtchns
may change at any time, and hence deducing that certain other operations
are safe when port_is_valid() returned true is not legitimate. The
opportunities to abuse this may get widened by the change here
(depending on guest and host configuration), but will be taken care of
by the other XSA.
This is XSA-338.
Fixes: 48974e6ce5 ("evtchn: use a per-domain variable for the max number of event channels")
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Reviewed-by: Stefano Stabellini <sstabellini@kernel.org>
Reviewed-by: Julien Grall <jgrall@amazon.com>
When assigned to less trusted or un-trusted guests, devices may change
state behind our backs (they may e.g. get reset by means we may not know
about). Therefore we should avoid reading BARs from hardware once a
device is no longer owned by Dom0. Furthermore when we can't read a BAR,
or when we read zero, we shouldn't instead use the caller provided
address unless that caller can be trusted.
Re-arrange the logic in msix_capability_init() such that only Dom0 (and
only if the device isn't DomU-owned yet) or calls through
PHYSDEVOP_prepare_msix will actually result in the reading of the
respective BAR register(s). Additionally do so only as long as in-use
table entries are known (note that invocation of PHYSDEVOP_prepare_msix
counts as a "pseudo" entry). In all other uses the value already
recorded will get used instead.
Clear the recorded values in _pci_cleanup_msix() as well as on the one
affected error path. (Adjust this error path to also avoid blindly
disabling MSI-X when it was enabled on entry to the function.)
While moving around variable declarations (in many cases to reduce their
scopes), also adjust some of their types.
This is part of XSA-337.
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Reviewed-by: Roger Pau Monné <roger.pau@citrix.com>
It's safer and faster to just use the cached last written
(untranslated) MSI message stored in msi_desc for the single user that
calls read_msi_msg.
This also prevents relying on the data read from the device MSI
registers in order to figure out the index into the IOMMU interrupt
remapping table, which is not safe.
This is part of XSA-337.
Reported-by: Andrew Cooper <andrew.cooper3@citrix.com>
Requested-by: Andrew Cooper <andrew.cooper3@citrix.com>
Signed-off-by: Roger Pau Monné <roger.pau@citrix.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
The current vPT code will migrate the emulated timers between vCPUs
(change the pt->vcpu field) while just holding the destination lock,
either from create_periodic_time or pt_adjust_global_vcpu_target if
the global target is adjusted. Changing the periodic_timer vCPU field
in this way creates a race where a third party could grab the lock in
the unlocked region of pt_adjust_global_vcpu_target (or before
create_periodic_time performs the vcpu change) and then release the
lock from a different vCPU, creating a locking imbalance.
Introduce a per-domain rwlock in order to protect periodic_time
migration between vCPU lists. Taking the lock in read mode prevents
any timer from being migrated to a different vCPU, while taking it in
write mode allows performing migration of timers across vCPUs. The
per-vcpu locks are still used to protect all the other fields from the
periodic_timer struct.
Note that such migration shouldn't happen frequently, and hence
there's no performance drop as a result of such locking.
This is XSA-336.
Reported-by: Igor Druzhinin <igor.druzhinin@citrix.com>
Tested-by: Igor Druzhinin <igor.druzhinin@citrix.com>
Signed-off-by: Roger Pau Monné <roger.pau@citrix.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
In the case that an HVM Stubdomain makes an XENMEM_acquire_resource hypercall,
the FIXME path will bypass rcu_unlock_domain() on the way out of the function.
Move the check to the start of the function. This does change the behaviour
of the get-size path for HVM Stubdomains, but that functionality is currently
broken and unused anyway, as well as being quite useless to entities which
can't actually map the resource anyway.
This is XSA-334.
Fixes: 83fa6552ce ("common: add a new mappable resource type: XENMEM_resource_grant_table")
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
This MSR doesn't exist on AMD hardware, and switching away from the safe
functions in the common MSR path was an erroneous change.
Partially revert the change.
This is XSA-333.
Fixes: 4fdc932b3c ("x86/Intel: drop another 32-bit leftover")
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
Reviewed-by: Wei Liu <wl@xen.org>
The function __cmpxchg_mb_timeout() was intended to have the same
semantics as __cmpxchg_mb(). Unfortunately, the memory barriers were
not added when first implemented.
There is no known issue with the existing callers, but the barriers are
added given this is the expected semantics in Xen.
The issue was introduced by XSA-295.
Backport: 4.8+
Fixes: 86b0bc9583 ("xen/arm: cmpxchg: Provide a new helper that can timeout")
Signed-off-by: Julien Grall <jgrall@amazon.com>
Reviewed-by: Stefano Stabellini <sstabellini@kernel.org>
Reviewed-by: Bertrand Marquis <bertrand.marquis@arm.com>
(cherry picked from commit d501ef90ae)
Xen has cpu_has_fp/cpu_has_simd to detect whether the CPU supports
FP/SIMD or not. But currently, these two MACROs only consider value 0
of ID_AA64PFR0_EL1.FP/SIMD as FP/SIMD features enabled. But for CPUs
that support FP/SIMD and half-precision floating-point arithmetic, the
ID_AA64PFR0_EL1.FP/SIMD are 1 (see Arm ARM DDI0487F.b, D13.2.64).
For these CPUs, xen will treat them as no FP/SIMD support, the
vfp_save/restore_state will not take effect.
From the TRM documents of Cortex-A75/A76/N1, we know these CPUs support
basic Advanced SIMD/FP and half-precision floating-point arithmetic. In
this case, on N1/A76/A75 platforms, Xen will always miss the floating
pointer registers save/restore. If different vCPUs are running on the
same pCPU, the floating pointer registers will be corrupted randomly.
This patch fixes Xen on these new cores.
Signed-off-by: Wei Chen <wei.chen@arm.com>
Reviewed-by: Bertrand Marquis <bertrand.marquis@arm.com>
Reviewed-by: Julien Grall <jgrall@amazon.com>
(cherry picked from commit 968bb86d04)
Commit 858c0be8c2 "xen/arm: Enable CPU Erratum 1165522 for Neoverse"
added a new erratum but forgot to update silicon-errata.txt.
Update the file accordingly to keep track of errata workaround in Xen.
Signed-off-by: Julien Grall <jgrall@amazon.com>
Reviewed-by: Bertrand Marquis <bertrand.marquis@arm.com>
Acked-by: Stefano Stabellini <sstabellini@kernel.org>
(cherry picked from commit 1814a626fb)
Enable CPU erratum of Speculative AT on the Neoverse N1 processor
versions r0p0 to r2p0.
Also Fix Cortex A76 Erratum string which had a wrong errata number.
Signed-off-by: Bertrand Marquis <bertrand.marquis@arm.com>
Acked-by: Julien Grall <jgrall@amazon.com>
(cherry picked from commit 858c0be8c2)
There are multiple bugs with the existing implementation.
On AMD CPUs prior to Zen2, loading a NUL segment selector doesn't clear the
segment base, which is a problem for 64bit code which typically expects to use
a NUL %fs/%gs selector.
On a context switch from any PV vcpu, to a 64bit PV vcpu with an %fs/%gs
selector which faults, the fixup logic loads NUL, and the guest is entered at
the failsafe callback with the stale base.
Alternatively, a PV context switch sequence of 64 (NUL, non-zero base) =>
32 (NUL) => 64 (NUL, zero base) will similarly cause Xen to enter the guest
with a stale base.
Both of these corner cases manifest as state corruption in the final vcpu.
However, damage is limited to to 64bit code expecting to use Thread Local
Storage with a base pointer of 0, which doesn't occur by default.
The context switch logic is extremely complicated, and is attempting to
optimise away loading a NUL selector (which is fast), or writing a 64bit base
of 0 (which is rare). Furthermore, it fails to respect Linux's ABI with
userspace, which manifests as userspace state corruption as far as Linux is
concerned.
Always restore all selector and base state, in all cases.
Leave a large comment explaining hardware behaviour, and the new ABI
expectations. Update the comments in the public headers.
Drop all "segment preloading" to handle the AMD corner case. It was never
anything but a waste of time for %ds/%es, and isn't needed now that %fs/%gs
bases are unconditionally written for 64bit PV guests. In load_segments(),
store the result of is_pv_32bit_vcpu() as it is an expensive predicate now,
and not used in a way which impacts speculative safety.
Reported-by: Andy Lutomirski <luto@kernel.org>
Reported-by: Sarah Newman <srn@prgmr.com>
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
x86/pv: Fix assertions in svm_load_segs()
OSSTest has shown an assertion failure:
http://logs.test-lab.xenproject.org/osstest/logs/153906/test-xtf-amd64-amd64-1/serial-rimava1.log
This is because we pass a non-NUL selector into svm_load_segs(), which is
something we must not do, as this path does not load the attributes/limits
from the GDT/LDT.
Drop the {fs,gs}_sel parameters from svm_load_segs() and use 0 instead. This
is acceptable even for non-zero NUL segments, as it is how the IRET
instruction behaves in all CPUs.
Only use the svm_load_segs() path when both FS and GS are NUL, which is the
common case when scheduling a 64bit vcpu with 64bit userspace in context.
Fixes: ad0fd291c5 ("x86/pv: Rewrite segment context switching from scratch")
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
master commit: ad0fd291c5
master date: 2020-09-07 11:32:34 +0100
master commit: 1e2d3be2e5
master date: 2020-09-08 16:46:31 +0100
The comments in save_segments(), _toggle_guest_pt() and write_cr() are false.
The %fs and %gs bases can be updated at any time by the guest.
As a consequence, Xen's fs_base/etc tracking state is always stale when the
vcpu is in context, and must not be used to complete MSR_{FS,GS}_BASE reads, etc.
In particular, a sequence such as:
wrmsr(MSR_FS_BASE, 0x1ull << 32);
write_fs(__USER_DS);
base = rdmsr(MSR_FS_BASE);
will return the stale base, not the new base. This may cause guest a guest
kernel's context switching of userspace to malfunction.
Therefore:
* Update save_segments(), _toggle_guest_pt() and read_msr() to always read
the segment bases from hardware.
* Update write_cr(), write_msr() and do_set_segment_base() to not not waste
time caching data which is instantly going to become stale again.
* Provide comments explaining when the tracking state is and isn't stale.
This bug has been present for 14 years, but several bugfixes since have built
on and extended the original flawed logic.
Fixes: ba9adb737b ("Apply stricter checking to RDMSR/WRMSR emulations.")
Fixes: c42494acb2 ("x86: fix FS/GS base handling when using the fsgsbase feature")
Fixed: eccc170053 ("x86/pv: Don't have %cr4.fsgsbase active behind a guest kernels back")
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
master commit: a5eaac9245
master date: 2020-09-07 11:32:34 +0100
The logic takes the segment selector unmodified from guest context. This
allowed the guest to load DPL0 descriptors into %gs. Fix up the RPL for
non-NUL selectors to be 3.
Xen's context switch logic skips saving the inactive %gs base, as it cannot be
modified by the guest behind Xen's back. This depends on Xen caching updates
to the inactive base, which is was missing from this path.
The consequence is that, following SEGBASE_GS_USER_SEL, the next context
switch will restore the stale inactive %gs base, and corrupt vcpu state.
Rework the hypercall to update the cached idea of gs_base_user, and fix the
behaviour in the case of the AMD NUL selector bug to always zero the segment
base.
Reported-by: Andy Lutomirski <luto@kernel.org>
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
master commit: afe018e041
master date: 2020-08-31 14:21:46 +0100
The overhead of (the lack of) MDS_NO alone has been measured at 30% on some
workloads. While we're not in a position yet to offer MSR_ARCH_CAPS generally
to guests, dom0 doesn't migrate, so we can pass a subset of hardware values
straight through.
This will cause PVH dom0's not to use KPTI by default, and all dom0's not to
use VERW flushing by default, and to use eIBRS in preference to retpoline on
recent Intel CPUs.
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
master commit: e46474278a
master date: 2020-08-31 13:43:26 +0100
... including serialisation/deserialisation logic and unit tests.
There is no current way to configure this MSR correctly for guests.
The toolstack side this logic needs building, which is far easier to
do with it in place.
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
master commit: e32605b07e
master date: 2020-08-27 12:48:46 +0100
LLVM 11 complains with:
<instantiation>:1:1: error: changed section flags for .init.rodata, expected: 0x2
.pushsection .init.rodata
^
<instantiation>:30:9: note: while in macro instantiation
entrypoint 0
^
entry.S:979:9: note: while in macro instantiation
.rept 256
^
And:
entry.S:1015:9: error: changed section flags for .init.rodata, expected: 0x2
.section .init.rodata
^
Fix it by explicitly using the same flags and type in all the
instances.
Signed-off-by: Roger Pau Monné <roger.pau@citrix.com>
Acked-by: Jan Beulich <jbeulich@suse.com>
master commit: d2770047a2
master date: 2020-08-27 09:53:46 +0200
In the case that bad_ioapic_register() fails, the current position of idx++
means that clear_fixmap(idx) will be called with the wrong index, and not
clean up the mapping just created.
Increment idx as part of the loop, rather than midway through the loop body.
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
master commit: b4175c6693
master date: 2020-08-05 17:35:11 +0100
All non-MMIO ranges (i.e those not mapping real device MMIO regions) that
map valid MFNs are normally marked MTRR_TYPE_WRBACK and 'ipat' is set. Hence
when PV drivers running in a guest populate the BAR space of the Xen Platform
PCI Device with pages such as the Shared Info page or Grant Table pages,
accesses to these pages will be cachable.
However, should IOMMU mappings be enabled be enabled for the guest then these
accesses become uncachable. This has a substantial negative effect on I/O
throughput of PV devices. Arguably PV drivers should bot be using BAR space to
host the Shared Info and Grant Table pages but it is currently commonplace for
them to do this and so this problem needs mitigation. Hence this patch makes
sure the 'ipat' bit is set for any special page regardless of where in GFN
space it is mapped.
NOTE: Clearly this mitigation only applies to Intel EPT. It is not obvious
that there is any similar mitigation possible for AMD NPT. Downstreams
such as Citrix XenServer have been carrying a patch similar to this for
several releases though.
Signed-off-by: Paul Durrant <pdurrant@amazon.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
master commit: ca24b2ffdb
master date: 2020-07-31 17:42:47 +0200
Displacement values can be negative, hence we shouldn't left-shift them.
Or else we get
(XEN) UBSAN: Undefined behaviour in x86_emulate/x86_emulate.c:3482:55
(XEN) left shift of negative value -2
While auditing shifts, I noticed a pair of missing parentheses, which
also get added right here.
Reported-by: Andrew Cooper <andrew.cooper3@citrix.com>
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Acked-by: Andrew Cooper <andrew.cooper3@citrix.com>
Tested-by: Andrew Cooper <andrew.cooper3@citrix.com>
x86emul: replace further UB shifts
I have no explanation how I managed to overlook these while putting
together what is now b6a907f8c8 ("x86emul: replace UB shifts").
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Acked-by: Andrew Cooper <andrew.cooper3@citrix.com>
master commit: b6a907f8c8
master date: 2020-07-31 17:41:58 +0200
master commit: 21de9680eb
master date: 2020-08-05 10:19:29 +0200
The bug is obvious here, other places in this function used
"cpufeat_mask" correctly.
Fixed: b648feff8e ("xen/x86: Improvements to in-hypervisor cpuid sanity checks")
Signed-off-by: Fam Zheng <famzheng@amazon.com>
Reviewed-by: Roger Pau Monné <roger.pau@citrix.com>
Reviewed-by: Andrew Cooper <andrew.cooper3@citrix.com>
master commit: 64219fa179
master date: 2020-07-29 19:03:41 +0100
wakeup_32 sets %ds and %es to BOOT_DS, while leaving %fs at what
wakeup_start did set it to, and %gs at whatever BIOS did load into it.
All of this may end up confusing the first load_segments() to run on
the BSP after resume, in particular allowing a non-nul selector value
to be left in %fs.
Alongside %ss, also put all other data segment registers into the same
state that the boot and CPU bringup paths put them in.
Reported-by: M. Vefa Bicakci <m.v.b@runbox.com>
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: 55f8c389d4
master date: 2020-07-24 10:17:26 +0200
This was lost when making the logic accessible to PVH Dom0.
While doing so make the access to the global function pointer safe
against races (as noticed by Roger): The only current user wants to be
invoked just once (but can tolerate to be invoked multiple times),
zapping the pointer at that point.
Fixes: 835d8d69d9 ("x86/rtc: provide mediated access to RTC for PVH dom0")
Signed-off-by: Jan Beulich <jbeulich@suse.com>
Reviewed-by: Roger Pau Monné <roger.pau@citrix.com>
master commit: f8fe3c0736
master date: 2020-07-15 15:46:30 +0200
Some x86 CPUs speculatively execute beyond indirect CALL/JMP instructions.
With CONFIG_INDIRECT_THUNK / Retpolines, indirect CALL/JMP instructions are
converted to direct CALL/JMP's to __x86_indirect_thunk_REG(), leaving just a
handful of indirect JMPs implementing those stubs.
There is no architectrual execution beyond an indirect JMP, so use INT3 as
recommended by vendors to halt speculative execution. This is shorter than
LFENCE (which would also work fine), but also shows up in logs if we do
unexpected execute them.
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
master commit: 3b7dab93f2
master date: 2020-07-01 17:01:24 +0100
XENMEM_acquire_resource and it's related structure is currently inside
a __XEN__ or __XEN_TOOLS__ guarded section to limit it's scope to the
hypervisor or the toolstack only. This is wrong as the hypercall is
already being used by the Linux kernel at least, and as such needs to
be public.
Also switch the usage of uint64_aligned_t to plain uint64_t, as
uint64_aligned_t is only to be used by the toolstack. Doing such
change will reduce the size of the structure on 32bit x86 by 4bytes,
since there will be no padding added after the frame_list handle.
This is fine, as users of the previous layout will allocate 4bytes of
padding that won't be read by Xen, and users of the new layout won't
allocate those, which is also fine since Xen won't try to access them.
Note that the structure already has compat handling, and such handling
will take care of copying the right size (ie: minus the padding) when
called from a 32bit x86 context. This is true for the compat code both
before and after this patch, since the structures in the memory.h
compat header are subject to a pragma pack(4), which already removed
the trailing padding that would otherwise be introduced by the
alignment of the frame field to 8 bytes.
Fixes: 3f8f12281d ('x86/mm: add HYPERVISOR_memory_op to acquire guest resources')
Signed-off-by: Roger Pau Monné <roger.pau@citrix.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
master commit: 0e2e54966a
master date: 2020-06-29 18:03:49 +0200
We do not expose the feature to guests, so should disallow access to the
respective MSRs. For simplicity, drop the entire block of MSRs, not just the
subset which have been specified thus far.
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
Reviewed-by: Wei Liu <wl@xen.org>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
master commit: bcdfbb70fc
master date: 2020-06-26 16:34:02 +0100
On some computers the bit width of the PM Timer as reported
by ACPI is 32 bits when in fact the FADT flags report correctly
that the timer is 24 bits wide. On affected machines such as the
ASUS FX504GM and never gaming laptops this results in the inability
to resume the machine from suspend. Without this patch suspend is
broken on affected machines and even if a machine manages to resume
correctly then the kernel time and xen timers are trashed.
Signed-off-by: Grzegorz Uriasz <gorbak25@gmail.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
master commit: f325d2477e
master date: 2020-06-25 09:11:09 +0200
While forking VMs running a small RTOS system (Zephyr) a Xen crash has been
observed due to a mm-lock order violation while copying the HVM CPU context
from the parent. This issue has been identified to be due to
hap_update_paging_modes first getting a lock on the gfn using get_gfn. This
call also creates a shared entry in the fork's memory map for the cr3 gfn. The
function later calls hap_update_cr3 while holding the paging_lock, which
results in the lock-order violation in vmx_load_pdptrs when it tries to unshare
the above entry when it grabs the page with the P2M_UNSHARE flag set.
Since vmx_load_pdptrs only reads from the page its usage of P2M_UNSHARE was
unnecessary to start with. Using P2M_ALLOC is the appropriate flag to ensure
the p2m is properly populated.
Note that the lock order violation is avoided because before the paging_lock is
taken a lookup is performed with P2M_ALLOC that forks the page, thus the second
lookup in vmx_load_pdptrs succeeds without having to perform the fork. We keep
P2M_ALLOC in vmx_load_pdptrs because there are code-paths leading up to it
which don't take the paging_lock and that have no previous lookup. Currently no
other code-path exists leading there with the paging_lock taken, thus no
further adjustments are necessary.
Signed-off-by: Tamas K Lengyel <tamas.lengyel@intel.com>
Reviewed-by: Roger Pau Monné <roger.pau@citrix.com>
Reviewed-by: Kevin Tian <kevin.tian@intel.com>
master commit: b67e859d08
master date: 2020-06-19 15:24:55 +0200
Currently a guest is able to register any guest physical address to use
for the vcpu_info structure as long as the structure can fits in the
rest of the frame.
This means a guest can provide an address that is not aligned to the
natural alignment of the structure.
On Arm 32-bit, unaligned access are completely forbidden by the
hypervisor. This will result to a data abort which is fatal.
On Arm 64-bit, unaligned access are only forbidden when used for atomic
access. As the structure contains fields (such as evtchn_pending_self)
that are updated using atomic operations, any unaligned access will be
fatal as well.
While the misalignment is only fatal on Arm, a generic check is added
as an x86 guest shouldn't sensibly pass an unaligned address (this
would result to a split lock).
This is XSA-327.
Reported-by: Julien Grall <jgrall@amazon.com>
Signed-off-by: Julien Grall <jgrall@amazon.com>
Reviewed-by: Andrew Cooper <andrew.cooper3@citrix.com>
Reviewed-by: Stefano Stabellini <sstabellini@kernel.org>
master commit: 3fdc211b01
master date: 2020-07-07 14:41:00 +0200
Modifications made to the page tables by EPT code need to be written
to memory when the page tables are shared with the IOMMU, as Intel
IOMMUs can be non-coherent and thus require changes to be written to
memory in order to be visible to the IOMMU.
In order to achieve this make sure data is written back to memory
after writing an EPT entry when the recalc bit is not set in
atomic_write_ept_entry. If such bit is set, the entry will be
adjusted and atomic_write_ept_entry will be called a second time
without the recalc bit set. Note that when splitting a super page the
new tables resulting of the split should also be written back.
Failure to do so can allow devices behind the IOMMU access to the
stale super page, or cause coherency issues as changes made by the
processor to the page tables are not visible to the IOMMU.
This allows to remove the VT-d specific iommu_pte_flush helper, since
the cache write back is now performed by atomic_write_ept_entry, and
hence iommu_iotlb_flush can be used to flush the IOMMU TLB. The newly
used method (iommu_iotlb_flush) can result in less flushes, since it
might sometimes be called rightly with 0 flags, in which case it
becomes a no-op.
This is part of XSA-321.
Signed-off-by: Roger Pau Monné <roger.pau@citrix.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
master commit: c23274fd04
master date: 2020-07-07 14:40:11 +0200
Some VT-d IOMMUs are non-coherent, which requires a cache write back
in order for the changes made by the CPU to be visible to the IOMMU.
This cache write back was unconditionally done using clflush, but there are
other more efficient instructions to do so, hence implement support
for them using the alternative framework.
This is part of XSA-321.
Signed-off-by: Roger Pau Monné <roger.pau@citrix.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
master commit: a64ea16522
master date: 2020-07-07 14:39:54 +0200
It's based on alternative_io_2 without inputs or outputs but with an
added memory clobber.
This is part of XSA-321.
Signed-off-by: Roger Pau Monné <roger.pau@citrix.com>
Acked-by: Jan Beulich <jbeulich@suse.com>
master commit: 23570bce00
master date: 2020-07-07 14:39:25 +0200
Current code in sync_cache assume that the address passed in is
aligned to a cache line size. Fix the code to support passing in
arbitrary addresses not necessarily aligned to a cache line size.
This is part of XSA-321.
Reported-by: Jan Beulich <jbeulich@suse.com>
Signed-off-by: Roger Pau Monné <roger.pau@citrix.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
master commit: b6d9398144
master date: 2020-07-07 14:39:05 +0200
The hook is only implemented for VT-d and it uses the already existing
iommu_sync_cache function present in VT-d code. The new hook is
added so that the cache can be flushed by code outside of VT-d when
using shared page tables.
Note that alloc_pgtable_maddr must use the now locally defined
sync_cache function, because IOMMU ops are not yet setup the first
time the function gets called during IOMMU initialization.
No functional change intended.
This is part of XSA-321.
Signed-off-by: Roger Pau Monné <roger.pau@citrix.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
master commit: 91526b460e
master date: 2020-07-07 14:38:34 +0200
Jan Beulich
Hongyan Xia
Wei Liu
Igor Druzhinin
Julien Grall
Andrew Cooper
Roger Pau Monné
Wei Chen
Bertrand Marquis
Paul Durrant
Fam Zheng
Grzegorz Uriasz
Tamas K Lengyel