diff --git a/Documentation/ia64/paravirt_ops.txt b/Documentation/ia64/paravirt_ops.txt
new file mode 100644
index 0000000..39ded02
--- /dev/null
+++ b/Documentation/ia64/paravirt_ops.txt
@@ -0,0 +1,137 @@
+Paravirt_ops on IA64
+====================
+                          21 May 2008, Isaku Yamahata <yamahata@valinux.co.jp>
+
+
+Introduction
+------------
+The aim of this documentation is to help with maintainability and/or to
+encourage people to use paravirt_ops/IA64.
+
+paravirt_ops (pv_ops in short) is a way for virtualization support of
+Linux kernel on x86. Several ways for virtualization support were
+proposed, paravirt_ops is the winner.
+On the other hand, now there are also several IA64 virtualization
+technologies like kvm/IA64, xen/IA64 and many other academic IA64
+hypervisors so that it is good to add generic virtualization
+infrastructure on Linux/IA64.
+
+
+What is paravirt_ops?
+---------------------
+It has been developed on x86 as virtualization support via API, not ABI.
+It allows each hypervisor to override operations which are important for
+hypervisors at API level. And it allows a single kernel binary to run on
+all supported execution environments including native machine.
+Essentially paravirt_ops is a set of function pointers which represent
+operations corresponding to low level sensitive instructions and high
+level functionalities in various area. But one significant difference
+from usual function pointer table is that it allows optimization with
+binary patch. It is because some of these operations are very
+performance sensitive and indirect call overhead is not negligible.
+With binary patch, indirect C function call can be transformed into
+direct C function call or in-place execution to eliminate the overhead.
+
+Thus, operations of paravirt_ops are classified into three categories.
+- simple indirect call
+  These operations correspond to high level functionality so that the
+  overhead of indirect call isn't very important.
+
+- indirect call which allows optimization with binary patch
+  Usually these operations correspond to low level instructions. They
+  are called frequently and performance critical. So the overhead is
+  very important.
+
+- a set of macros for hand written assembly code
+  Hand written assembly codes (.S files) also need paravirtualization
+  because they include sensitive instructions or some of code paths in
+  them are very performance critical.
+
+
+The relation to the IA64 machine vector
+---------------------------------------
+Linux/IA64 has the IA64 machine vector functionality which allows the
+kernel to switch implementations (e.g. initialization, ipi, dma api...)
+depending on executing platform.
+We can replace some implementations very easily defining a new machine
+vector. Thus another approach for virtualization support would be
+enhancing the machine vector functionality.
+But paravirt_ops approach was taken because
+- virtualization support needs wider support than machine vector does.
+  e.g. low level instruction paravirtualization. It must be
+       initialized very early before platform detection.
+
+- virtualization support needs more functionality like binary patch.
+  Probably the calling overhead might not be very large compared to the
+  emulation overhead of virtualization. However in the native case, the
+  overhead should be eliminated completely.
+  A single kernel binary should run on each environment including native,
+  and the overhead of paravirt_ops on native environment should be as
+  small as possible.
+
+- for full virtualization technology, e.g. KVM/IA64 or
+  Xen/IA64 HVM domain, the result would be
+  (the emulated platform machine vector. probably dig) + (pv_ops).
+  This means that the virtualization support layer should be under
+  the machine vector layer.
+
+Possibly it might be better to move some function pointers from
+paravirt_ops to machine vector. In fact, Xen domU case utilizes both
+pv_ops and machine vector.
+
+
+IA64 paravirt_ops
+-----------------
+In this section, the concrete paravirt_ops will be discussed.
+Because of the architecture difference between ia64 and x86, the
+resulting set of functions is very different from x86 pv_ops.
+
+- C function pointer tables
+They are not very performance critical so that simple C indirect
+function call is acceptable. The following structures are defined at
+this moment. For details see linux/include/asm-ia64/paravirt.h
+  - struct pv_info
+    This structure describes the execution environment.
+  - struct pv_init_ops
+    This structure describes the various initialization hooks.
+  - struct pv_iosapic_ops
+    This structure describes hooks to iosapic operations.
+  - struct pv_irq_ops
+    This structure describes hooks to irq related operations
+  - struct pv_time_op
+    This structure describes hooks to steal time accounting.
+
+- a set of indirect calls which need optimization
+Currently this class of functions correspond to a subset of IA64
+intrinsics. At this moment the optimization with binary patch isn't
+implemented yet.
+struct pv_cpu_op is defined. For details see
+linux/include/asm-ia64/paravirt_privop.h
+Mostly they correspond to ia64 intrinsics 1-to-1.
+Caveat: Now they are defined as C indirect function pointers, but in
+order to support binary patch optimization, they will be changed
+using GCC extended inline assembly code.
+
+- a set of macros for hand written assembly code (.S files)
+For maintenance purpose, the taken approach for .S files is single
+source code and compile multiple times with different macros definitions.
+Each pv_ops instance must define those macros to compile.
+The important thing here is that sensitive, but non-privileged
+instructions must be paravirtualized and that some privileged
+instructions also need paravirtualization for reasonable performance.
+Developers who modify .S files must be aware of that. At this moment
+an easy checker is implemented to detect paravirtualization breakage.
+But it doesn't cover all the cases.
+
+Sometimes this set of macros is called pv_cpu_asm_op. But there is no
+corresponding structure in the source code.
+Those macros mostly 1:1 correspond to a subset of privileged
+instructions. See linux/include/asm-ia64/native/inst.h.
+And some functions written in assembly also need to be overrided so
+that each pv_ops instance have to define some macros. Again see
+linux/include/asm-ia64/native/inst.h.
+
+
+Those structures must be initialized very early before start_kernel.
+Probably initialized in head.S using multi entry point or some other trick.
+For native case implementation see linux/arch/ia64/kernel/paravirt.c.