summaryrefslogtreecommitdiff
path: root/Documentation/virtual/kvm/vcpu-requests.rst
diff options
context:
space:
mode:
Diffstat (limited to 'Documentation/virtual/kvm/vcpu-requests.rst')
-rw-r--r--Documentation/virtual/kvm/vcpu-requests.rst307
1 files changed, 307 insertions, 0 deletions
diff --git a/Documentation/virtual/kvm/vcpu-requests.rst b/Documentation/virtual/kvm/vcpu-requests.rst
new file mode 100644
index 000000000000..5feb3706a7ae
--- /dev/null
+++ b/Documentation/virtual/kvm/vcpu-requests.rst
@@ -0,0 +1,307 @@
+=================
+KVM VCPU Requests
+=================
+
+Overview
+========
+
+KVM supports an internal API enabling threads to request a VCPU thread to
+perform some activity. For example, a thread may request a VCPU to flush
+its TLB with a VCPU request. The API consists of the following functions::
+
+ /* Check if any requests are pending for VCPU @vcpu. */
+ bool kvm_request_pending(struct kvm_vcpu *vcpu);
+
+ /* Check if VCPU @vcpu has request @req pending. */
+ bool kvm_test_request(int req, struct kvm_vcpu *vcpu);
+
+ /* Clear request @req for VCPU @vcpu. */
+ void kvm_clear_request(int req, struct kvm_vcpu *vcpu);
+
+ /*
+ * Check if VCPU @vcpu has request @req pending. When the request is
+ * pending it will be cleared and a memory barrier, which pairs with
+ * another in kvm_make_request(), will be issued.
+ */
+ bool kvm_check_request(int req, struct kvm_vcpu *vcpu);
+
+ /*
+ * Make request @req of VCPU @vcpu. Issues a memory barrier, which pairs
+ * with another in kvm_check_request(), prior to setting the request.
+ */
+ void kvm_make_request(int req, struct kvm_vcpu *vcpu);
+
+ /* Make request @req of all VCPUs of the VM with struct kvm @kvm. */
+ bool kvm_make_all_cpus_request(struct kvm *kvm, unsigned int req);
+
+Typically a requester wants the VCPU to perform the activity as soon
+as possible after making the request. This means most requests
+(kvm_make_request() calls) are followed by a call to kvm_vcpu_kick(),
+and kvm_make_all_cpus_request() has the kicking of all VCPUs built
+into it.
+
+VCPU Kicks
+----------
+
+The goal of a VCPU kick is to bring a VCPU thread out of guest mode in
+order to perform some KVM maintenance. To do so, an IPI is sent, forcing
+a guest mode exit. However, a VCPU thread may not be in guest mode at the
+time of the kick. Therefore, depending on the mode and state of the VCPU
+thread, there are two other actions a kick may take. All three actions
+are listed below:
+
+1) Send an IPI. This forces a guest mode exit.
+2) Waking a sleeping VCPU. Sleeping VCPUs are VCPU threads outside guest
+ mode that wait on waitqueues. Waking them removes the threads from
+ the waitqueues, allowing the threads to run again. This behavior
+ may be suppressed, see KVM_REQUEST_NO_WAKEUP below.
+3) Nothing. When the VCPU is not in guest mode and the VCPU thread is not
+ sleeping, then there is nothing to do.
+
+VCPU Mode
+---------
+
+VCPUs have a mode state, ``vcpu->mode``, that is used to track whether the
+guest is running in guest mode or not, as well as some specific
+outside guest mode states. The architecture may use ``vcpu->mode`` to
+ensure VCPU requests are seen by VCPUs (see "Ensuring Requests Are Seen"),
+as well as to avoid sending unnecessary IPIs (see "IPI Reduction"), and
+even to ensure IPI acknowledgements are waited upon (see "Waiting for
+Acknowledgements"). The following modes are defined:
+
+OUTSIDE_GUEST_MODE
+
+ The VCPU thread is outside guest mode.
+
+IN_GUEST_MODE
+
+ The VCPU thread is in guest mode.
+
+EXITING_GUEST_MODE
+
+ The VCPU thread is transitioning from IN_GUEST_MODE to
+ OUTSIDE_GUEST_MODE.
+
+READING_SHADOW_PAGE_TABLES
+
+ The VCPU thread is outside guest mode, but it wants the sender of
+ certain VCPU requests, namely KVM_REQ_TLB_FLUSH, to wait until the VCPU
+ thread is done reading the page tables.
+
+VCPU Request Internals
+======================
+
+VCPU requests are simply bit indices of the ``vcpu->requests`` bitmap.
+This means general bitops, like those documented in [atomic-ops]_ could
+also be used, e.g. ::
+
+ clear_bit(KVM_REQ_UNHALT & KVM_REQUEST_MASK, &vcpu->requests);
+
+However, VCPU request users should refrain from doing so, as it would
+break the abstraction. The first 8 bits are reserved for architecture
+independent requests, all additional bits are available for architecture
+dependent requests.
+
+Architecture Independent Requests
+---------------------------------
+
+KVM_REQ_TLB_FLUSH
+
+ KVM's common MMU notifier may need to flush all of a guest's TLB
+ entries, calling kvm_flush_remote_tlbs() to do so. Architectures that
+ choose to use the common kvm_flush_remote_tlbs() implementation will
+ need to handle this VCPU request.
+
+KVM_REQ_MMU_RELOAD
+
+ When shadow page tables are used and memory slots are removed it's
+ necessary to inform each VCPU to completely refresh the tables. This
+ request is used for that.
+
+KVM_REQ_PENDING_TIMER
+
+ This request may be made from a timer handler run on the host on behalf
+ of a VCPU. It informs the VCPU thread to inject a timer interrupt.
+
+KVM_REQ_UNHALT
+
+ This request may be made from the KVM common function kvm_vcpu_block(),
+ which is used to emulate an instruction that causes a CPU to halt until
+ one of an architectural specific set of events and/or interrupts is
+ received (determined by checking kvm_arch_vcpu_runnable()). When that
+ event or interrupt arrives kvm_vcpu_block() makes the request. This is
+ in contrast to when kvm_vcpu_block() returns due to any other reason,
+ such as a pending signal, which does not indicate the VCPU's halt
+ emulation should stop, and therefore does not make the request.
+
+KVM_REQUEST_MASK
+----------------
+
+VCPU requests should be masked by KVM_REQUEST_MASK before using them with
+bitops. This is because only the lower 8 bits are used to represent the
+request's number. The upper bits are used as flags. Currently only two
+flags are defined.
+
+VCPU Request Flags
+------------------
+
+KVM_REQUEST_NO_WAKEUP
+
+ This flag is applied to requests that only need immediate attention
+ from VCPUs running in guest mode. That is, sleeping VCPUs do not need
+ to be awaken for these requests. Sleeping VCPUs will handle the
+ requests when they are awaken later for some other reason.
+
+KVM_REQUEST_WAIT
+
+ When requests with this flag are made with kvm_make_all_cpus_request(),
+ then the caller will wait for each VCPU to acknowledge its IPI before
+ proceeding. This flag only applies to VCPUs that would receive IPIs.
+ If, for example, the VCPU is sleeping, so no IPI is necessary, then
+ the requesting thread does not wait. This means that this flag may be
+ safely combined with KVM_REQUEST_NO_WAKEUP. See "Waiting for
+ Acknowledgements" for more information about requests with
+ KVM_REQUEST_WAIT.
+
+VCPU Requests with Associated State
+===================================
+
+Requesters that want the receiving VCPU to handle new state need to ensure
+the newly written state is observable to the receiving VCPU thread's CPU
+by the time it observes the request. This means a write memory barrier
+must be inserted after writing the new state and before setting the VCPU
+request bit. Additionally, on the receiving VCPU thread's side, a
+corresponding read barrier must be inserted after reading the request bit
+and before proceeding to read the new state associated with it. See
+scenario 3, Message and Flag, of [lwn-mb]_ and the kernel documentation
+[memory-barriers]_.
+
+The pair of functions, kvm_check_request() and kvm_make_request(), provide
+the memory barriers, allowing this requirement to be handled internally by
+the API.
+
+Ensuring Requests Are Seen
+==========================
+
+When making requests to VCPUs, we want to avoid the receiving VCPU
+executing in guest mode for an arbitrary long time without handling the
+request. We can be sure this won't happen as long as we ensure the VCPU
+thread checks kvm_request_pending() before entering guest mode and that a
+kick will send an IPI to force an exit from guest mode when necessary.
+Extra care must be taken to cover the period after the VCPU thread's last
+kvm_request_pending() check and before it has entered guest mode, as kick
+IPIs will only trigger guest mode exits for VCPU threads that are in guest
+mode or at least have already disabled interrupts in order to prepare to
+enter guest mode. This means that an optimized implementation (see "IPI
+Reduction") must be certain when it's safe to not send the IPI. One
+solution, which all architectures except s390 apply, is to:
+
+- set ``vcpu->mode`` to IN_GUEST_MODE between disabling the interrupts and
+ the last kvm_request_pending() check;
+- enable interrupts atomically when entering the guest.
+
+This solution also requires memory barriers to be placed carefully in both
+the requesting thread and the receiving VCPU. With the memory barriers we
+can exclude the possibility of a VCPU thread observing
+!kvm_request_pending() on its last check and then not receiving an IPI for
+the next request made of it, even if the request is made immediately after
+the check. This is done by way of the Dekker memory barrier pattern
+(scenario 10 of [lwn-mb]_). As the Dekker pattern requires two variables,
+this solution pairs ``vcpu->mode`` with ``vcpu->requests``. Substituting
+them into the pattern gives::
+
+ CPU1 CPU2
+ ================= =================
+ local_irq_disable();
+ WRITE_ONCE(vcpu->mode, IN_GUEST_MODE); kvm_make_request(REQ, vcpu);
+ smp_mb(); smp_mb();
+ if (kvm_request_pending(vcpu)) { if (READ_ONCE(vcpu->mode) ==
+ IN_GUEST_MODE) {
+ ...abort guest entry... ...send IPI...
+ } }
+
+As stated above, the IPI is only useful for VCPU threads in guest mode or
+that have already disabled interrupts. This is why this specific case of
+the Dekker pattern has been extended to disable interrupts before setting
+``vcpu->mode`` to IN_GUEST_MODE. WRITE_ONCE() and READ_ONCE() are used to
+pedantically implement the memory barrier pattern, guaranteeing the
+compiler doesn't interfere with ``vcpu->mode``'s carefully planned
+accesses.
+
+IPI Reduction
+-------------
+
+As only one IPI is needed to get a VCPU to check for any/all requests,
+then they may be coalesced. This is easily done by having the first IPI
+sending kick also change the VCPU mode to something !IN_GUEST_MODE. The
+transitional state, EXITING_GUEST_MODE, is used for this purpose.
+
+Waiting for Acknowledgements
+----------------------------
+
+Some requests, those with the KVM_REQUEST_WAIT flag set, require IPIs to
+be sent, and the acknowledgements to be waited upon, even when the target
+VCPU threads are in modes other than IN_GUEST_MODE. For example, one case
+is when a target VCPU thread is in READING_SHADOW_PAGE_TABLES mode, which
+is set after disabling interrupts. To support these cases, the
+KVM_REQUEST_WAIT flag changes the condition for sending an IPI from
+checking that the VCPU is IN_GUEST_MODE to checking that it is not
+OUTSIDE_GUEST_MODE.
+
+Request-less VCPU Kicks
+-----------------------
+
+As the determination of whether or not to send an IPI depends on the
+two-variable Dekker memory barrier pattern, then it's clear that
+request-less VCPU kicks are almost never correct. Without the assurance
+that a non-IPI generating kick will still result in an action by the
+receiving VCPU, as the final kvm_request_pending() check does for
+request-accompanying kicks, then the kick may not do anything useful at
+all. If, for instance, a request-less kick was made to a VCPU that was
+just about to set its mode to IN_GUEST_MODE, meaning no IPI is sent, then
+the VCPU thread may continue its entry without actually having done
+whatever it was the kick was meant to initiate.
+
+One exception is x86's posted interrupt mechanism. In this case, however,
+even the request-less VCPU kick is coupled with the same
+local_irq_disable() + smp_mb() pattern described above; the ON bit
+(Outstanding Notification) in the posted interrupt descriptor takes the
+role of ``vcpu->requests``. When sending a posted interrupt, PIR.ON is
+set before reading ``vcpu->mode``; dually, in the VCPU thread,
+vmx_sync_pir_to_irr() reads PIR after setting ``vcpu->mode`` to
+IN_GUEST_MODE.
+
+Additional Considerations
+=========================
+
+Sleeping VCPUs
+--------------
+
+VCPU threads may need to consider requests before and/or after calling
+functions that may put them to sleep, e.g. kvm_vcpu_block(). Whether they
+do or not, and, if they do, which requests need consideration, is
+architecture dependent. kvm_vcpu_block() calls kvm_arch_vcpu_runnable()
+to check if it should awaken. One reason to do so is to provide
+architectures a function where requests may be checked if necessary.
+
+Clearing Requests
+-----------------
+
+Generally it only makes sense for the receiving VCPU thread to clear a
+request. However, in some circumstances, such as when the requesting
+thread and the receiving VCPU thread are executed serially, such as when
+they are the same thread, or when they are using some form of concurrency
+control to temporarily execute synchronously, then it's possible to know
+that the request may be cleared immediately, rather than waiting for the
+receiving VCPU thread to handle the request in VCPU RUN. The only current
+examples of this are kvm_vcpu_block() calls made by VCPUs to block
+themselves. A possible side-effect of that call is to make the
+KVM_REQ_UNHALT request, which may then be cleared immediately when the
+VCPU returns from the call.
+
+References
+==========
+
+.. [atomic-ops] Documentation/core-api/atomic_ops.rst
+.. [memory-barriers] Documentation/memory-barriers.txt
+.. [lwn-mb] https://lwn.net/Articles/573436/