1.. SPDX-License-Identifier: GPL-2.0
2
3=================
4KVM VCPU Requests
5=================
6
7Overview
8========
9
10KVM supports an internal API enabling threads to request a VCPU thread to
11perform some activity.  For example, a thread may request a VCPU to flush
12its TLB with a VCPU request.  The API consists of the following functions::
13
14  /* Check if any requests are pending for VCPU @vcpu. */
15  bool kvm_request_pending(struct kvm_vcpu *vcpu);
16
17  /* Check if VCPU @vcpu has request @req pending. */
18  bool kvm_test_request(int req, struct kvm_vcpu *vcpu);
19
20  /* Clear request @req for VCPU @vcpu. */
21  void kvm_clear_request(int req, struct kvm_vcpu *vcpu);
22
23  /*
24   * Check if VCPU @vcpu has request @req pending. When the request is
25   * pending it will be cleared and a memory barrier, which pairs with
26   * another in kvm_make_request(), will be issued.
27   */
28  bool kvm_check_request(int req, struct kvm_vcpu *vcpu);
29
30  /*
31   * Make request @req of VCPU @vcpu. Issues a memory barrier, which pairs
32   * with another in kvm_check_request(), prior to setting the request.
33   */
34  void kvm_make_request(int req, struct kvm_vcpu *vcpu);
35
36  /* Make request @req of all VCPUs of the VM with struct kvm @kvm. */
37  bool kvm_make_all_cpus_request(struct kvm *kvm, unsigned int req);
38
39Typically a requester wants the VCPU to perform the activity as soon
40as possible after making the request.  This means most requests
41(kvm_make_request() calls) are followed by a call to kvm_vcpu_kick(),
42and kvm_make_all_cpus_request() has the kicking of all VCPUs built
43into it.
44
45VCPU Kicks
46----------
47
48The goal of a VCPU kick is to bring a VCPU thread out of guest mode in
49order to perform some KVM maintenance.  To do so, an IPI is sent, forcing
50a guest mode exit.  However, a VCPU thread may not be in guest mode at the
51time of the kick.  Therefore, depending on the mode and state of the VCPU
52thread, there are two other actions a kick may take.  All three actions
53are listed below:
54
551) Send an IPI.  This forces a guest mode exit.
562) Waking a sleeping VCPU.  Sleeping VCPUs are VCPU threads outside guest
57   mode that wait on waitqueues.  Waking them removes the threads from
58   the waitqueues, allowing the threads to run again.  This behavior
59   may be suppressed, see KVM_REQUEST_NO_WAKEUP below.
603) Nothing.  When the VCPU is not in guest mode and the VCPU thread is not
61   sleeping, then there is nothing to do.
62
63VCPU Mode
64---------
65
66VCPUs have a mode state, ``vcpu->mode``, that is used to track whether the
67guest is running in guest mode or not, as well as some specific
68outside guest mode states.  The architecture may use ``vcpu->mode`` to
69ensure VCPU requests are seen by VCPUs (see "Ensuring Requests Are Seen"),
70as well as to avoid sending unnecessary IPIs (see "IPI Reduction"), and
71even to ensure IPI acknowledgements are waited upon (see "Waiting for
72Acknowledgements").  The following modes are defined:
73
74OUTSIDE_GUEST_MODE
75
76  The VCPU thread is outside guest mode.
77
78IN_GUEST_MODE
79
80  The VCPU thread is in guest mode.
81
82EXITING_GUEST_MODE
83
84  The VCPU thread is transitioning from IN_GUEST_MODE to
85  OUTSIDE_GUEST_MODE.
86
87READING_SHADOW_PAGE_TABLES
88
89  The VCPU thread is outside guest mode, but it wants the sender of
90  certain VCPU requests, namely KVM_REQ_TLB_FLUSH, to wait until the VCPU
91  thread is done reading the page tables.
92
93VCPU Request Internals
94======================
95
96VCPU requests are simply bit indices of the ``vcpu->requests`` bitmap.
97This means general bitops, like those documented in [atomic-ops]_ could
98also be used, e.g. ::
99
100  clear_bit(KVM_REQ_UNBLOCK & KVM_REQUEST_MASK, &vcpu->requests);
101
102However, VCPU request users should refrain from doing so, as it would
103break the abstraction.  The first 8 bits are reserved for architecture
104independent requests, all additional bits are available for architecture
105dependent requests.
106
107Architecture Independent Requests
108---------------------------------
109
110KVM_REQ_TLB_FLUSH
111
112  KVM's common MMU notifier may need to flush all of a guest's TLB
113  entries, calling kvm_flush_remote_tlbs() to do so.  Architectures that
114  choose to use the common kvm_flush_remote_tlbs() implementation will
115  need to handle this VCPU request.
116
117KVM_REQ_VM_DEAD
118
119  This request informs all VCPUs that the VM is dead and unusable, e.g. due to
120  fatal error or because the VM's state has been intentionally destroyed.
121
122KVM_REQ_UNBLOCK
123
124  This request informs the vCPU to exit kvm_vcpu_block.  It is used for
125  example from timer handlers that run on the host on behalf of a vCPU,
126  or in order to update the interrupt routing and ensure that assigned
127  devices will wake up the vCPU.
128
129KVM_REQ_OUTSIDE_GUEST_MODE
130
131  This "request" ensures the target vCPU has exited guest mode prior to the
132  sender of the request continuing on.  No action needs be taken by the target,
133  and so no request is actually logged for the target.  This request is similar
134  to a "kick", but unlike a kick it guarantees the vCPU has actually exited
135  guest mode.  A kick only guarantees the vCPU will exit at some point in the
136  future, e.g. a previous kick may have started the process, but there's no
137  guarantee the to-be-kicked vCPU has fully exited guest mode.
138
139KVM_REQUEST_MASK
140----------------
141
142VCPU requests should be masked by KVM_REQUEST_MASK before using them with
143bitops.  This is because only the lower 8 bits are used to represent the
144request's number.  The upper bits are used as flags.  Currently only two
145flags are defined.
146
147VCPU Request Flags
148------------------
149
150KVM_REQUEST_NO_WAKEUP
151
152  This flag is applied to requests that only need immediate attention
153  from VCPUs running in guest mode.  That is, sleeping VCPUs do not need
154  to be awaken for these requests.  Sleeping VCPUs will handle the
155  requests when they are awaken later for some other reason.
156
157KVM_REQUEST_WAIT
158
159  When requests with this flag are made with kvm_make_all_cpus_request(),
160  then the caller will wait for each VCPU to acknowledge its IPI before
161  proceeding.  This flag only applies to VCPUs that would receive IPIs.
162  If, for example, the VCPU is sleeping, so no IPI is necessary, then
163  the requesting thread does not wait.  This means that this flag may be
164  safely combined with KVM_REQUEST_NO_WAKEUP.  See "Waiting for
165  Acknowledgements" for more information about requests with
166  KVM_REQUEST_WAIT.
167
168VCPU Requests with Associated State
169===================================
170
171Requesters that want the receiving VCPU to handle new state need to ensure
172the newly written state is observable to the receiving VCPU thread's CPU
173by the time it observes the request.  This means a write memory barrier
174must be inserted after writing the new state and before setting the VCPU
175request bit.  Additionally, on the receiving VCPU thread's side, a
176corresponding read barrier must be inserted after reading the request bit
177and before proceeding to read the new state associated with it.  See
178scenario 3, Message and Flag, of [lwn-mb]_ and the kernel documentation
179[memory-barriers]_.
180
181The pair of functions, kvm_check_request() and kvm_make_request(), provide
182the memory barriers, allowing this requirement to be handled internally by
183the API.
184
185Ensuring Requests Are Seen
186==========================
187
188When making requests to VCPUs, we want to avoid the receiving VCPU
189executing in guest mode for an arbitrary long time without handling the
190request.  We can be sure this won't happen as long as we ensure the VCPU
191thread checks kvm_request_pending() before entering guest mode and that a
192kick will send an IPI to force an exit from guest mode when necessary.
193Extra care must be taken to cover the period after the VCPU thread's last
194kvm_request_pending() check and before it has entered guest mode, as kick
195IPIs will only trigger guest mode exits for VCPU threads that are in guest
196mode or at least have already disabled interrupts in order to prepare to
197enter guest mode.  This means that an optimized implementation (see "IPI
198Reduction") must be certain when it's safe to not send the IPI.  One
199solution, which all architectures except s390 apply, is to:
200
201- set ``vcpu->mode`` to IN_GUEST_MODE between disabling the interrupts and
202  the last kvm_request_pending() check;
203- enable interrupts atomically when entering the guest.
204
205This solution also requires memory barriers to be placed carefully in both
206the requesting thread and the receiving VCPU.  With the memory barriers we
207can exclude the possibility of a VCPU thread observing
208!kvm_request_pending() on its last check and then not receiving an IPI for
209the next request made of it, even if the request is made immediately after
210the check.  This is done by way of the Dekker memory barrier pattern
211(scenario 10 of [lwn-mb]_).  As the Dekker pattern requires two variables,
212this solution pairs ``vcpu->mode`` with ``vcpu->requests``.  Substituting
213them into the pattern gives::
214
215  CPU1                                    CPU2
216  =================                       =================
217  local_irq_disable();
218  WRITE_ONCE(vcpu->mode, IN_GUEST_MODE);  kvm_make_request(REQ, vcpu);
219  smp_mb();                               smp_mb();
220  if (kvm_request_pending(vcpu)) {        if (READ_ONCE(vcpu->mode) ==
221                                              IN_GUEST_MODE) {
222      ...abort guest entry...                 ...send IPI...
223  }                                       }
224
225As stated above, the IPI is only useful for VCPU threads in guest mode or
226that have already disabled interrupts.  This is why this specific case of
227the Dekker pattern has been extended to disable interrupts before setting
228``vcpu->mode`` to IN_GUEST_MODE.  WRITE_ONCE() and READ_ONCE() are used to
229pedantically implement the memory barrier pattern, guaranteeing the
230compiler doesn't interfere with ``vcpu->mode``'s carefully planned
231accesses.
232
233IPI Reduction
234-------------
235
236As only one IPI is needed to get a VCPU to check for any/all requests,
237then they may be coalesced.  This is easily done by having the first IPI
238sending kick also change the VCPU mode to something !IN_GUEST_MODE.  The
239transitional state, EXITING_GUEST_MODE, is used for this purpose.
240
241Waiting for Acknowledgements
242----------------------------
243
244Some requests, those with the KVM_REQUEST_WAIT flag set, require IPIs to
245be sent, and the acknowledgements to be waited upon, even when the target
246VCPU threads are in modes other than IN_GUEST_MODE.  For example, one case
247is when a target VCPU thread is in READING_SHADOW_PAGE_TABLES mode, which
248is set after disabling interrupts.  To support these cases, the
249KVM_REQUEST_WAIT flag changes the condition for sending an IPI from
250checking that the VCPU is IN_GUEST_MODE to checking that it is not
251OUTSIDE_GUEST_MODE.
252
253Request-less VCPU Kicks
254-----------------------
255
256As the determination of whether or not to send an IPI depends on the
257two-variable Dekker memory barrier pattern, then it's clear that
258request-less VCPU kicks are almost never correct.  Without the assurance
259that a non-IPI generating kick will still result in an action by the
260receiving VCPU, as the final kvm_request_pending() check does for
261request-accompanying kicks, then the kick may not do anything useful at
262all.  If, for instance, a request-less kick was made to a VCPU that was
263just about to set its mode to IN_GUEST_MODE, meaning no IPI is sent, then
264the VCPU thread may continue its entry without actually having done
265whatever it was the kick was meant to initiate.
266
267One exception is x86's posted interrupt mechanism.  In this case, however,
268even the request-less VCPU kick is coupled with the same
269local_irq_disable() + smp_mb() pattern described above; the ON bit
270(Outstanding Notification) in the posted interrupt descriptor takes the
271role of ``vcpu->requests``.  When sending a posted interrupt, PIR.ON is
272set before reading ``vcpu->mode``; dually, in the VCPU thread,
273vmx_sync_pir_to_irr() reads PIR after setting ``vcpu->mode`` to
274IN_GUEST_MODE.
275
276Additional Considerations
277=========================
278
279Sleeping VCPUs
280--------------
281
282VCPU threads may need to consider requests before and/or after calling
283functions that may put them to sleep, e.g. kvm_vcpu_block().  Whether they
284do or not, and, if they do, which requests need consideration, is
285architecture dependent.  kvm_vcpu_block() calls kvm_arch_vcpu_runnable()
286to check if it should awaken.  One reason to do so is to provide
287architectures a function where requests may be checked if necessary.
288
289References
290==========
291
292.. [atomic-ops] Documentation/atomic_bitops.txt and Documentation/atomic_t.txt
293.. [memory-barriers] Documentation/memory-barriers.txt
294.. [lwn-mb] https://lwn.net/Articles/573436/
295