[Qemu-devel] [qemu-web PATCH] Add "Understanding QEMU devices" blog post

[Thomas Huth]

Last July, Eric Blake wrote a nice summary for newcomers about what
QEMU has to do to emulate devices for the guests. So far, we missed
to integrate this somewhere into the QEM web site or wiki, so let's
publish this now as a nice blog post for the users.

Signed-off-by: Thomas Huth <address@hidden>
---
 _posts/2018-01-26-understanding-qemu-devices.md | 139 ++++++++++++++++++++++++
 1 file changed, 139 insertions(+)
 create mode 100644 _posts/2018-01-26-understanding-qemu-devices.md

diff --git a/_posts/2018-01-26-understanding-qemu-devices.md 
b/_posts/2018-01-26-understanding-qemu-devices.md
new file mode 100644
index 0000000..b436ef0
--- /dev/null
+++ b/_posts/2018-01-26-understanding-qemu-devices.md
@@ -0,0 +1,139 @@
+---
+layout: post
+title:  "Understanding QEMU devices"
+date:   2018-01-26 10:00:00 +0100
+author: Eric Blake
+categories: blog
+---
+Here are some notes that may help newcomers understand what is actually
+happening with QEMU devices:
+
+With QEMU, one thing to remember is that we are trying to emulate what
+an OS would see on bare-metal hardware.  All bare-metal machines are
+basically giant memory maps, where software poking at a particular
+address will have a particular side effect (the most common side effect
+is, of course, accessing memory; but other common regions in memory
+include the register banks for controlling particular pieces of
+hardware, like the hard drive or a network card, or even the CPU
+itself).  The end-goal of emulation is to allow a user-space program,
+using only normal memory accesses, to manage all of the side-effects
+that a guest OS is expecting.
+
+As an implementation detail, some hardware, like x86, actually has two
+memory spaces, where I/O space uses different assembly codes than
+normal; QEMU has to emulate these alternative accesses.  Similarly, many
+modern hardware is so complex that the CPU itself provides both
+specialized assembly instructions and a bank of registers within the
+memory map (a classic example being the management of the MMU, or
+separation between Ring 0 kernel code and Ring 3 userspace code - if
+that's not crazy enough, there's nested virtualization).  With certain
+hardware, we have virtualization hooks where the CPU itself makes it
+easy to trap on just the problematic assembly instructions (those that
+access I/O space or CPU internal registers, and therefore require side
+effects different than a normal memory access), so that the guest just
+executes the same assembly sequence as on bare metal, but that execution
+then causes a trap to let user-space QEMU then react to the instructions
+using just its normal user-space memory accesses before returning
+control to the guest.  This is the kvm accelerator, and can let a guest
+run nearly as fast as bare metal, where the slowdowns are caused by each
+trap from guest back to QEMU (a vmexit) to handle a difficult assembly
+instruction or memory address.  QEMU also supports a TCG accelerator,
+which takes the guest assembly instructions and compiles it on the fly
+into comparable host instructions or calls to host helper routines (not
+as fast, but results in QEMU being able to do cross-hardware emulation).
+
+The next thing to realize is what is happening when an OS is accessing
+various hardware resources.  For example, most OS ship with a driver
+that knows how to manage an IDE disk - the driver is merely software
+that is programmed to make specific I/O requests to a specific subset of
+the memory map (wherever the IDE bus lives, as hard-coded by the
+hardware board designers), in order to make the disk drive hardware then
+obey commands to copy data from memory to persistent storage (writing to
+disk) or from persistent storage to memory (reading from the disk).
+When you first buy bare-metal hardware, your disk is uninitialized; you
+install the OS that uses the driver to make enough bare-metal accesses
+to the IDE hardware portion of the memory map to then turn the disk into
+a set of partitions and filesystems on top of those partitions.
+
+So, how does QEMU emulate this? In the big memory map it provides to the
+guest, it emulates an IDE disk at the same address as bare-metal would.
+When the guest OS driver issues particular memory writes to the IDE
+control registers in order to copy data from memory to persistent
+storage, QEMU traps on those writes (whether via kvm hypervisor assist,
+or by noticing during TCG translation that the addresses being accessed
+are special), and emulates the same side effects by issuing host
+commands to copy the specified guest memory into host storage.  On the
+host side, the easiest way to emulate persistent storage is via treating
+a file in the host filesystem as raw data (a 1:1 mapping of offsets in
+the host file to disk offsets being accessed by the guest driver), but
+QEMU actually has the ability to glue together a lot of different host
+formats (raw, qcow2, qed, vhdx, ...) and protocols (file system, block
+device, NBD, sheepdog, gluster, ...) where any combination of host
+format and protocol can serve as the backend that is then tied to the
+QEMU emulation providing the guest device.
+
+Thus, when you tell QEMU to use a host qcow2 file, the guest does not
+have to know qcow2, but merely has its normal driver make the same
+register reads and writes as it would on bare metal, which cause vmexits
+into QEMU code, then QEMU maps those accesses into reads and writes in
+the appropriate offsets of the qcow2 file.  When you first install the
+guest, all the guest sees is a blank uninitialized linear disk
+(regardless of whether that disk is linear in the host, as in raw
+format, or optimized for random access, as in the qcow2 format); it is
+up to the guest OS to decide how to partition its view of the hardware
+and install filesystems on top of that, and QEMU does not care what
+filesystems the guest is using, only what pattern of raw disk I/O
+register control sequences are issued.
+
+The next thing to realize is that emulating IDE is not always the most
+efficient.  Every time the guest writes to the control registers, it has
+to go through special handling, and vmexits slow down emulation.  One
+way to speed this up is through paravirtualization, or cooperation
+between the guest and host.  The QEMU developers have produced a
+specification for a set of hardware registers and the behavior for those
+registers which are designed to result in the minimum number of vmexits
+possible while still accomplishing what a hard disk must do, namely,
+transferring data between normal guest memory and persistent storage.
+This specification is called virtio; using it requires installation of a
+virtio driver in the guest.  While there is no known hardware that
+follows the same register layout as virtio, the concept is the same: a
+virtio disk behaves like a memory-mapped register bank, where the guest
+OS driver then knows what sequence of register commands to write into
+that bank to cause data to be copied in and out of other guest memory.
+Much of the speedups in virtio come by its design - the guest sets aside
+a portion of regular memory for the bulk of its command queue, and only
+has to kick a single register to then tell QEMU to read the command
+queue (fewer mapped register accesses mean fewer vmexits), coupled with
+handshaking guarantees that the guest driver won't be changing the
+normal memory while QEMU is acting on it.
+
+In a similar vein, many OS have support for a number of network cards, a
+common example being the e1000 card on the PCI bus.  On bare metal, an
+OS will probe PCI space, see that a bank of registers with the signature
+for e1000 is populated, and load the driver that then knows what
+register sequences to write in order to let the hardware card transfer
+network traffic in and out of the guest.  So QEMU has, as one of its
+many network card emulations, an e1000 device, which is mapped to the
+same guest memory region as a real one would live on bare metal.  And
+once again, the e1000 register layout tends to require a lot of register
+writes (and thus vmexits) for the amount of work the hardware performs,
+so the QEMU developers have added the virtio-net card (a PCI hardware
+specification, although no bare-metal hardware exists that actually
+implements it), such that installing a virtio-net driver in the guest OS
+can then minimize the number of vmexits while still getting the same
+side-effects of sending network traffic.  If you tell QEMU to start a
+guest with a virtio-net card, then the guest OS will probe PCI space and
+see a bank of registers with the virtio-net signature, and load the
+appropriate driver like it would for any other PCI hardware.
+
+In summary, even though QEMU was first written as a way of emulating
+hardware memory maps in order to virtualize a guest OS, it turns out
+that the fastest virtualization also depends on virtual hardware: a
+memory map of registers with particular documented side effects that has
+no bare-metal counterpart.  And at the end of the day, all
+virtualization really means is running a particular set of assembly
+instructions (the guest OS) to manipulate locations within a giant
+memory map for causing a particular set of side effects, where QEMU is
+just a user-space application providing a memory map and mimicking the
+same side effects you would get when executing those guest instructions
+on the appropriate bare metal hardware.
-- 
1.8.3.1