
	Timekeeping Virtualization for X86-Based Architectures

	Zachary Amsden <zamsden@redhat.com>
	Copyright (c) 2010, Red Hat.  All rights reserved.

1) Overview
2) Timing Devices
3) TSC Hardware
4) Virtualization Problems

=========================================================================

1) Overview

One of the most complicated parts of the X86 platform, and specifically,
the virtualization of this platform is the plethora of timing devices available
and the complexity of emulating those devices.  In addition, virtualization of
time introduces a new set of challenges because it introduces a multiplexed
division of time beyond the control of the guest CPU.

First, we will describe the various timekeeping hardware available, then
present some of the problems which arise and solutions available, giving
specific recommendations for certain classes of KVM guests.

The purpose of this document is to collect data and information relevant to
timekeeping which may be difficult to find elsewhere, specifically,
information relevant to KVM and hardware-based virtualization.

=========================================================================

2) Timing Devices

First we discuss the basic hardware devices available.  TSC and the related
KVM clock are special enough to warrant a full exposition and are described in
the following section.

2.1) i8254 - PIT

One of the first timer devices available is the programmable interrupt timer,
or PIT.  The PIT has a fixed frequency 1.193182 MHz base clock and three
channels which can be programmed to deliver periodic or one-shot interrupts.
These three channels can be configured in different modes and have individual
counters.  Channel 1 and 2 were not available for general use in the original
IBM PC, and historically were connected to control RAM refresh and the PC
speaker.  Now the PIT is typically integrated as part of an emulated chipset
and a separate physical PIT is not used.

The PIT uses I/O ports 0x40 - 0x43.  Access to the 16-bit counters is done
using single or multiple byte access to the I/O ports.  There are 6 modes
available, but not all modes are available to all timers, as only timer 2
has a connected gate input, required for modes 1 and 5.  The gate line is
controlled by port 61h, bit 0, as illustrated in the following diagram.

 --------------             ----------------
|              |           |                |
|  1.1932 MHz  |---------->| CLOCK      OUT | ---------> IRQ 0
|    Clock     |   |       |                |
 --------------    |    +->| GATE  TIMER 0  |
                   |        ----------------
                   |
                   |        ----------------
                   |       |                |
                   |------>| CLOCK      OUT | ---------> 66.3 KHZ DRAM
                   |       |                |            (aka /dev/null)
                   |    +->| GATE  TIMER 1  |
                   |        ----------------
                   |
                   |        ----------------
                   |       |                |
                   |------>| CLOCK      OUT | ---------> Port 61h, bit 5
                           |                |      |
Port 61h, bit 0 ---------->| GATE  TIMER 2  |       \_.----   ____
                            ----------------         _|    )--|LPF|---Speaker
                                                    / *----   \___/
Port 61h, bit 1 -----------------------------------/

The timer modes are now described.

Mode 0: Single Timeout.   This is a one-shot software timeout that counts down
 when the gate is high (always true for timers 0 and 1).  When the count
 reaches zero, the output goes high.

Mode 1: Triggered One-shot.  The output is intially set high.  When the gate
 line is set high, a countdown is initiated (which does not stop if the gate is
 lowered), during which the output is set low.  When the count reaches zero,
 the output goes high.

Mode 2: Rate Generator.  The output is initially set high.  When the countdown
 reaches 1, the output goes low for one count and then returns high.  The value
 is reloaded and the countdown automatically resumes.  If the gate line goes
 low, the count is halted.  If the output is low when the gate is lowered, the
 output automatically goes high (this only affects timer 2).

Mode 3: Square Wave.   This generates a high / low square wave.  The count
 determines the length of the pulse, which alternates between high and low
 when zero is reached.  The count only proceeds when gate is high and is
 automatically reloaded on reaching zero.  The count is decremented twice at
 each clock to generate a full high / low cycle at the full periodic rate.
 If the count is even, the clock remains high for N/2 counts and low for N/2
 counts; if the clock is odd, the clock is high for (N+1)/2 counts and low
 for (N-1)/2 counts.  Only even values are latched by the counter, so odd
 values are not observed when reading.  This is the intended mode for timer 2,
 which generates sine-like tones by low-pass filtering the square wave output.

Mode 4: Software Strobe.  After programming this mode and loading the counter,
 the output remains high until the counter reaches zero.  Then the output
 goes low for 1 clock cycle and returns high.  The counter is not reloaded.
 Counting only occurs when gate is high.

Mode 5: Hardware Strobe.  After programming and loading the counter, the
 output remains high.  When the gate is raised, a countdown is initiated
 (which does not stop if the gate is lowered).  When the counter reaches zero,
 the output goes low for 1 clock cycle and then returns high.  The counter is
 not reloaded.

In addition to normal binary counting, the PIT supports BCD counting.  The
command port, 0x43 is used to set the counter and mode for each of the three
timers.

PIT commands, issued to port 0x43, using the following bit encoding:

Bit 7-4: Command (See table below)
Bit 3-1: Mode (000 = Mode 0, 101 = Mode 5, 11X = undefined)
Bit 0  : Binary (0) / BCD (1)

Command table:

0000 - Latch Timer 0 count for port 0x40
	sample and hold the count to be read in port 0x40;
	additional commands ignored until counter is read;
	mode bits ignored.

0001 - Set Timer 0 LSB mode for port 0x40
	set timer to read LSB only and force MSB to zero;
	mode bits set timer mode

0010 - Set Timer 0 MSB mode for port 0x40
	set timer to read MSB only and force LSB to zero;
	mode bits set timer mode

0011 - Set Timer 0 16-bit mode for port 0x40
	set timer to read / write LSB first, then MSB;
	mode bits set timer mode

0100 - Latch Timer 1 count for port 0x41 - as described above
0101 - Set Timer 1 LSB mode for port 0x41 - as described above
0110 - Set Timer 1 MSB mode for port 0x41 - as described above
0111 - Set Timer 1 16-bit mode for port 0x41 - as described above

1000 - Latch Timer 2 count for port 0x42 - as described above
1001 - Set Timer 2 LSB mode for port 0x42 - as described above
1010 - Set Timer 2 MSB mode for port 0x42 - as described above
1011 - Set Timer 2 16-bit mode for port 0x42 as described above

1101 - General counter latch
	Latch combination of counters into corresponding ports
	Bit 3 = Counter 2
	Bit 2 = Counter 1
	Bit 1 = Counter 0
	Bit 0 = Unused

1110 - Latch timer status
	Latch combination of counter mode into corresponding ports
	Bit 3 = Counter 2
	Bit 2 = Counter 1
	Bit 1 = Counter 0

	The output of ports 0x40-0x42 following this command will be:

	Bit 7 = Output pin
	Bit 6 = Count loaded (0 if timer has expired)
	Bit 5-4 = Read / Write mode
	    01 = MSB only
	    10 = LSB only
	    11 = LSB / MSB (16-bit)
	Bit 3-1 = Mode
	Bit 0 = Binary (0) / BCD mode (1)

2.2) RTC

The second device which was available in the original PC was the MC146818 real
time clock.  The original device is now obsolete, and usually emulated by the
system chipset, sometimes by an HPET and some frankenstein IRQ routing.

The RTC is accessed through CMOS variables, which uses an index register to
control which bytes are read.  Since there is only one index register, read
of the CMOS and read of the RTC require lock protection (in addition, it is
dangerous to allow userspace utilities such as hwclock to have direct RTC
access, as they could corrupt kernel reads and writes of CMOS memory).

The RTC generates an interrupt which is usually routed to IRQ 8.  The interrupt
can function as a periodic timer, an additional once a day alarm, and can issue
interrupts after an update of the CMOS registers by the MC146818 is complete.
The type of interrupt is signalled in the RTC status registers.

The RTC will update the current time fields by battery power even while the
system is off.  The current time fields should not be read while an update is
in progress, as indicated in the status register.

The clock uses a 32.768kHz crystal, so bits 6-4 of register A should be
programmed to a 32kHz divider if the RTC is to count seconds.

This is the RAM map originally used for the RTC/CMOS:

Location    Size    Description
------------------------------------------
00h         byte    Current second (BCD)
01h         byte    Seconds alarm (BCD)
02h         byte    Current minute (BCD)
03h         byte    Minutes alarm (BCD)
04h         byte    Current hour (BCD)
05h         byte    Hours alarm (BCD)
06h         byte    Current day of week (BCD)
07h         byte    Current day of month (BCD)
08h         byte    Current month (BCD)
09h         byte    Current year (BCD)
0Ah         byte    Register A
                       bit 7   = Update in progress
                       bit 6-4 = Divider for clock
                                  000 = 4.194 MHz
                                  001 = 1.049 MHz
                                  010 = 32 kHz
                                  10X = test modes
                                  110 = reset / disable
                                  111 = reset / disable
                       bit 3-0 = Rate selection for periodic interrupt
                                  000 = periodic timer disabled
                                  001 = 3.90625 uS
                                  010 = 7.8125 uS
                                  011 = .122070 mS
                                  100 = .244141 mS
                                     ...
                                 1101 = 125 mS
                                 1110 = 250 mS
                                 1111 = 500 mS
0Bh         byte    Register B
                       bit 7   = Run (0) / Halt (1)
                       bit 6   = Periodic interrupt enable
                       bit 5   = Alarm interrupt enable
                       bit 4   = Update-ended interrupt enable
                       bit 3   = Square wave interrupt enable
                       bit 2   = BCD calendar (0) / Binary (1)
                       bit 1   = 12-hour mode (0) / 24-hour mode (1)
                       bit 0   = 0 (DST off) / 1 (DST enabled)
OCh         byte    Register C (read only)
                       bit 7   = interrupt request flag (IRQF)
                       bit 6   = periodic interrupt flag (PF)
                       bit 5   = alarm interrupt flag (AF)
                       bit 4   = update interrupt flag (UF)
                       bit 3-0 = reserved
ODh         byte    Register D (read only)
                       bit 7   = RTC has power
                       bit 6-0 = reserved
32h         byte    Current century BCD (*)
  (*) location vendor specific and now determined from ACPI global tables

2.3) APIC

On Pentium and later processors, an on-board timer is available to each CPU
as part of the Advanced Programmable Interrupt Controller.  The APIC is
accessed through memory-mapped registers and provides interrupt service to each
CPU, used for IPIs and local timer interrupts.

Although in theory the APIC is a safe and stable source for local interrupts,
in practice, many bugs and glitches have occurred due to the special nature of
the APIC CPU-local memory-mapped hardware.  Beware that CPU errata may affect
the use of the APIC and that workarounds may be required.  In addition, some of
these workarounds pose unique constraints for virtualization - requiring either
extra overhead incurred from extra reads of memory-mapped I/O or additional
functionality that may be more computationally expensive to implement.

Since the APIC is documented quite well in the Intel and AMD manuals, we will
avoid repetition of the detail here.  It should be pointed out that the APIC
timer is programmed through the LVT (local vector timer) register, is capable
of one-shot or periodic operation, and is based on the bus clock divided down
by the programmable divider register.

2.4) HPET

HPET is quite complex, and was originally intended to replace the PIT / RTC
support of the X86 PC.  It remains to be seen whether that will be the case, as
the de facto standard of PC hardware is to emulate these older devices.  Some
systems designated as legacy free may support only the HPET as a hardware timer
device.

The HPET spec is rather loose and vague, requiring at least 3 hardware timers,
but allowing implementation freedom to support many more.  It also imposes no
fixed rate on the timer frequency, but does impose some extremal values on
frequency, error and slew.

In general, the HPET is recommended as a high precision (compared to PIT /RTC)
time source which is independent of local variation (as there is only one HPET
in any given system).  The HPET is also memory-mapped, and its presence is
indicated through ACPI tables by the BIOS.

Detailed specification of the HPET is beyond the current scope of this
document, as it is also very well documented elsewhere.

2.5) Offboard Timers

Several cards, both proprietary (watchdog boards) and commonplace (e1000) have
timing chips built into the cards which may have registers which are accessible
to kernel or user drivers.  To the author's knowledge, using these to generate
a clocksource for a Linux or other kernel has not yet been attempted and is in
general frowned upon as not playing by the agreed rules of the game.  Such a
timer device would require additional support to be virtualized properly and is
not considered important at this time as no known operating system does this.

=========================================================================

3) TSC Hardware

The TSC or time stamp counter is relatively simple in theory; it counts
instruction cycles issued by the processor, which can be used as a measure of
time.  In practice, due to a number of problems, it is the most complicated
timekeeping device to use.

The TSC is represented internally as a 64-bit MSR which can be read with the
RDMSR, RDTSC, or RDTSCP (when available) instructions.  In the past, hardware
limitations made it possible to write the TSC, but generally on old hardware it
was only possible to write the low 32-bits of the 64-bit counter, and the upper
32-bits of the counter were cleared.  Now, however, on Intel processors family
0Fh, for models 3, 4 and 6, and family 06h, models e and f, this restriction
has been lifted and all 64-bits are writable.  On AMD systems, the ability to
write the TSC MSR is not an architectural guarantee.

The TSC is accessible from CPL-0 and conditionally, for CPL > 0 software by
means of the CR4.TSD bit, which when enabled, disables CPL > 0 TSC access.

Some vendors have implemented an additional instruction, RDTSCP, which returns
atomically not just the TSC, but an indicator which corresponds to the
processor number.  This can be used to index into an array of TSC variables to
determine offset information in SMP systems where TSCs are not synchronized.
The presence of this instruction must be determined by consulting CPUID feature
bits.

Both VMX and SVM provide extension fields in the virtualization hardware which
allows the guest visible TSC to be offset by a constant.  Newer implementations
promise to allow the TSC to additionally be scaled, but this hardware is not
yet widely available.

3.1) TSC synchronization

The TSC is a CPU-local clock in most implementations.  This means, on SMP
platforms, the TSCs of different CPUs may start at different times depending
on when the CPUs are powered on.  Generally, CPUs on the same die will share
the same clock, however, this is not always the case.

The BIOS may attempt to resynchronize the TSCs during the poweron process and
the operating system or other system software may attempt to do this as well.
Several hardware limitations make the problem worse - if it is not possible to
write the full 64-bits of the TSC, it may be impossible to match the TSC in
newly arriving CPUs to that of the rest of the system, resulting in
unsynchronized TSCs.  This may be done by BIOS or system software, but in
practice, getting a perfectly synchronized TSC will not be possible unless all
values are read from the same clock, which generally only is possible on single
socket systems or those with special hardware support.

3.2) TSC and CPU hotplug

As touched on already, CPUs which arrive later than the boot time of the system
may not have a TSC value that is synchronized with the rest of the system.
Either system software, BIOS, or SMM code may actually try to establish the TSC
to a value matching the rest of the system, but a perfect match is usually not
a guarantee.  This can have the effect of bringing a system from a state where
TSC is synchronized back to a state where TSC synchronization flaws, however
small, may be exposed to the OS and any virtualization environment.

3.3) TSC and multi-socket / NUMA

Multi-socket systems, especially large multi-socket systems are likely to have
individual clocksources rather than a single, universally distributed clock.
Since these clocks are driven by different crystals, they will not have
perfectly matched frequency, and temperature and electrical variations will
cause the CPU clocks, and thus the TSCs to drift over time.  Depending on the
exact clock and bus design, the drift may or may not be fixed in absolute
error, and may accumulate over time.

In addition, very large systems may deliberately slew the clocks of individual
cores.  This technique, known as spread-spectrum clocking, reduces EMI at the
clock frequency and harmonics of it, which may be required to pass FCC
standards for telecommunications and computer equipment.

It is recommended not to trust the TSCs to remain synchronized on NUMA or
multiple socket systems for these reasons.

3.4) TSC and C-states

C-states, or idling states of the processor, especially C1E and deeper sleep
states may be problematic for TSC as well.  The TSC may stop advancing in such
a state, resulting in a TSC which is behind that of other CPUs when execution
is resumed.  Such CPUs must be detected and flagged by the operating system
based on CPU and chipset identifications.

The TSC in such a case may be corrected by catching it up to a known external
clocksource.

3.5) TSC frequency change / P-states

To make things slightly more interesting, some CPUs may change frequency.  They
may or may not run the TSC at the same rate, and because the frequency change
may be staggered or slewed, at some points in time, the TSC rate may not be
known other than falling within a range of values.  In this case, the TSC will
not be a stable time source, and must be calibrated against a known, stable,
external clock to be a usable source of time.

Whether the TSC runs at a constant rate or scales with the P-state is model
dependent and must be determined by inspecting CPUID, chipset or vendor
specific MSR fields.

In addition, some vendors have known bugs where the P-state is actually
compensated for properly during normal operation, but when the processor is
inactive, the P-state may be raised temporarily to service cache misses from
other processors.  In such cases, the TSC on halted CPUs could advance faster
than that of non-halted processors.  AMD Turion processors are known to have
this problem.

3.6) TSC and STPCLK / T-states

External signals given to the processor may also have the effect of stopping
the TSC.  This is typically done for thermal emergency power control to prevent
an overheating condition, and typically, there is no way to detect that this
condition has happened.

3.7) TSC virtualization - VMX

VMX provides conditional trapping of RDTSC, RDMSR, WRMSR and RDTSCP
instructions, which is enough for full virtualization of TSC in any manner.  In
addition, VMX allows passing through the host TSC plus an additional TSC_OFFSET
field specified in the VMCS.  Special instructions must be used to read and
write the VMCS field.

3.8) TSC virtualization - SVM

SVM provides conditional trapping of RDTSC, RDMSR, WRMSR and RDTSCP
instructions, which is enough for full virtualization of TSC in any manner.  In
addition, SVM allows passing through the host TSC plus an additional offset
field specified in the SVM control block.

3.9) TSC feature bits in Linux

In summary, there is no way to guarantee the TSC remains in perfect
synchronization unless it is explicitly guaranteed by the architecture.  Even
if so, the TSCs in multi-sockets or NUMA systems may still run independently
despite being locally consistent.

The following feature bits are used by Linux to signal various TSC attributes,
but they can only be taken to be meaningful for UP or single node systems.

X86_FEATURE_TSC 		: The TSC is available in hardware
X86_FEATURE_RDTSCP		: The RDTSCP instruction is available
X86_FEATURE_CONSTANT_TSC 	: The TSC rate is unchanged with P-states
X86_FEATURE_NONSTOP_TSC		: The TSC does not stop in C-states
X86_FEATURE_TSC_RELIABLE	: TSC sync checks are skipped (VMware)

4) Virtualization Problems

Timekeeping is especially problematic for virtualization because a number of
challenges arise.  The most obvious problem is that time is now shared between
the host and, potentially, a number of virtual machines.  Thus the virtual
operating system does not run with 100% usage of the CPU, despite the fact that
it may very well make that assumption.  It may expect it to remain true to very
exacting bounds when interrupt sources are disabled, but in reality only its
virtual interrupt sources are disabled, and the machine may still be preempted
at any time.  This causes problems as the passage of real time, the injection
of machine interrupts and the associated clock sources are no longer completely
synchronized with real time.

This same problem can occur on native harware to a degree, as SMM mode may
steal cycles from the naturally on X86 systems when SMM mode is used by the
BIOS, but not in such an extreme fashion.  However, the fact that SMM mode may
cause similar problems to virtualization makes it a good justification for
solving many of these problems on bare metal.

4.1) Interrupt clocking

One of the most immediate problems that occurs with legacy operating systems
is that the system timekeeping routines are often designed to keep track of
time by counting periodic interrupts.  These interrupts may come from the PIT
or the RTC, but the problem is the same: the host virtualization engine may not
be able to deliver the proper number of interrupts per second, and so guest
time may fall behind.  This is especially problematic if a high interrupt rate
is selected, such as 1000 HZ, which is unfortunately the default for many Linux
guests.

There are three approaches to solving this problem; first, it may be possible
to simply ignore it.  Guests which have a separate time source for tracking
'wall clock' or 'real time' may not need any adjustment of their interrupts to
maintain proper time.  If this is not sufficient, it may be necessary to inject
additional interrupts into the guest in order to increase the effective
interrupt rate.  This approach leads to complications in extreme conditions,
where host load or guest lag is too much to compensate for, and thus another
solution to the problem has risen: the guest may need to become aware of lost
ticks and compensate for them internally.  Although promising in theory, the
implementation of this policy in Linux has been extremely error prone, and a
number of buggy variants of lost tick compensation are distributed across
commonly used Linux systems.

Windows uses periodic RTC clocking as a means of keeping time internally, and
thus requires interrupt slewing to keep proper time.  It does use a low enough
rate (ed: is it 18.2 Hz?) however that it has not yet been a problem in
practice.

4.2) TSC sampling and serialization

As the highest precision time source available, the cycle counter of the CPU
has aroused much interest from developers.  As explained above, this timer has
many problems unique to its nature as a local, potentially unstable and
potentially unsynchronized source.  One issue which is not unique to the TSC,
but is highlighted because of its very precise nature is sampling delay.  By
definition, the counter, once read is already old.  However, it is also
possible for the counter to be read ahead of the actual use of the result.
This is a consequence of the superscalar execution of the instruction stream,
which may execute instructions out of order.  Such execution is called
non-serialized.  Forcing serialized execution is necessary for precise
measurement with the TSC, and requires a serializing instruction, such as CPUID
or an MSR read.

Since CPUID may actually be virtualized by a trap and emulate mechanism, this
serialization can pose a performance issue for hardware virtualization.  An
accurate time stamp counter reading may therefore not always be available, and
it may be necessary for an implementation to guard against "backwards" reads of
the TSC as seen from other CPUs, even in an otherwise perfectly synchronized
system.

4.3) Timespec aliasing

Additionally, this lack of serialization from the TSC poses another challenge
when using results of the TSC when measured against another time source.  As
the TSC is much higher precision, many possible values of the TSC may be read
while another clock is still expressing the same value.

That is, you may read (T,T+10) while external clock C maintains the same value.
Due to non-serialized reads, you may actually end up with a range which
fluctuates - from (T-1.. T+10).  Thus, any time calculated from a TSC, but
calibrated against an external value may have a range of valid values.
Re-calibrating this computation may actually cause time, as computed after the
calibration, to go backwards, compared with time computed before the
calibration.

This problem is particularly pronounced with an internal time source in Linux,
the kernel time, which is expressed in the theoretically high resolution
timespec - but which advances in much larger granularity intervals, sometimes
at the rate of jiffies, and possibly in catchup modes, at a much larger step.

This aliasing requires care in the computation and recalibration of kvmclock
and any other values derived from TSC computation (such as TSC virtualization
itself).

4.4) Migration

Migration of a virtual machine raises problems for timekeeping in two ways.
First, the migration itself may take time, during which interrupts cannot be
delivered, and after which, the guest time may need to be caught up.  NTP may
be able to help to some degree here, as the clock correction required is
typically small enough to fall in the NTP-correctable window.

An additional concern is that timers based off the TSC (or HPET, if the raw bus
clock is exposed) may now be running at different rates, requiring compensation
in some way in the hypervisor by virtualizing these timers.  In addition,
migrating to a faster machine may preclude the use of a passthrough TSC, as a
faster clock cannot be made visible to a guest without the potential of time
advancing faster than usual.  A slower clock is less of a problem, as it can
always be caught up to the original rate.  KVM clock avoids these problems by
simply storing multipliers and offsets against the TSC for the guest to convert
back into nanosecond resolution values.

4.5) Scheduling

Since scheduling may be based on precise timing and firing of interrupts, the
scheduling algorithms of an operating system may be adversely affected by
virtualization.  In theory, the effect is random and should be universally
distributed, but in contrived as well as real scenarios (guest device access,
causes of virtualization exits, possible context switch), this may not always
be the case.  The effect of this has not been well studied.

In an attempt to work around this, several implementations have provided a
paravirtualized scheduler clock, which reveals the true amount of CPU time for
which a virtual machine has been running.

4.6) Watchdogs

Watchdog timers, such as the lock detector in Linux may fire accidentally when
running under hardware virtualization due to timer interrupts being delayed or
misinterpretation of the passage of real time.  Usually, these warnings are
spurious and can be ignored, but in some circumstances it may be necessary to
disable such detection.

4.7) Delays and precision timing

Precise timing and delays may not be possible in a virtualized system.  This
can happen if the system is controlling physical hardware, or issues delays to
compensate for slower I/O to and from devices.  The first issue is not solvable
in general for a virtualized system; hardware control software can't be
adequately virtualized without a full real-time operating system, which would
require an RT aware virtualization platform.

The second issue may cause performance problems, but this is unlikely to be a
significant issue.  In many cases these delays may be eliminated through
configuration or paravirtualization.

4.8) Covert channels and leaks

In addition to the above problems, time information will inevitably leak to the
guest about the host in anything but a perfect implementation of virtualized
time.  This may allow the guest to infer the presence of a hypervisor (as in a
red-pill type detection), and it may allow information to leak between guests
by using CPU utilization itself as a signalling channel.  Preventing such
problems would require completely isolated virtual time which may not track
real time any longer.  This may be useful in certain security or QA contexts,
but in general isn't recommended for real-world deployment scenarios.