By Dave Scott - 2013-12-29
This article is part of a series documenting how MirageOS applications run under Xen. This article is about "events"; i.e. how can an app wait for input to arrive and tell someone that output is available?
A running virtual machine under Xen is known as a domain. A domain has a number of virtual CPUs (vCPUs) which run until the Xen scheduler decides to pre-empt them, or until they ask to block via a hypercall (a system call to the hypervisor). A typical domain has no direct hardware access and instead performs I/O by talking to other privileged driver domains (often domain 0) via Xen-specific disk and network protocols. These protocols use two primitives:
This article focuses on how events work; a future article will describe how shared memory works.
An event channel is a logical connection between (domain_1, port_1) and (domain_2, port_2) where port_1 and port_2 are integers, like TCP port numbers or Unix file descriptors. An event sent from one domain will cause the other domain to unblock (if it hasn't been "masked"). To understand how event channels are used, it's worth comparing I/O under Unix to I/O under Xen:
When a Unix process starts, it runs in a context with environment variables, pre-connected file descriptors and command-line arguments. When a Xen domain starts, it runs in a context with a start info page, pre-bound event channels and pre-shared memory for console and xenstore.
A Unix process which wants to perform network I/O will normally connect sockets (additional file descriptors) to network resources, and the kernel will take care of talking protocols like TCP/IP. A Xen domain which wants to perform network I/O will share memory with- and then bind event channels to- network driver domains, and then exchange raw ethernet frames. The Xen domain will contain its own TCP/IP stack (such as mirage-tcpip).
When a Unix process wants to read or write data via a file descriptor it can use select(2) to wait until data (or space) is available, and then use read(2) or write(2), passing pointers to data buffers as arguments. When a Xen domain wants to wait for data (or space) it will block until an event arrives, and then send an event to signal that data has been produced or consumed. Note that neither blocking nor sending take buffers as arguments since under Xen, data (or metadata) is placed into shared memory beforehand. The events are simply a way to say, "look at the shared buffers again".
Every domain maps a special shared info page which contains bitmaps representing the state of each event channel. This per-channel state consists of:
Every vCPU has a vcpu_info record in the shared info page, which stores two relevant domain-global (not per event channel) bits:
Note that all MirageOS guests are single vCPU and therefore we can simplify things by relying on the (single) per-vCPU evtchn_upcall_mask rather than the fine-grained evtchn_mask (normally a multi-vCPU guest would use the evtchn_upcall_mask to control reentrant execution and the evtchn_mask to coalesce event wakeups).
Note the shared info page is shared between the domain and the hypervisor
without any locks, so an architecture-specific protocol must be used to access
it (usually via C macros with names like
When a domain wants to transmit an event, it calls the calls the EVTCHNOP_send hypercall. Within Xen, this calls xen/common/event_channel.c:evtchn_set_pending which tests the evtchn_pending bit for this event channel. If it's already set then no further work is needed and so it returns. If the bit isn't already set, then it is set and then evtchn_mask is queried. The evtchn_mask is always clear for MirageOS guests, so control passes to xen/arch/x86/domain.c:vcpu_mark_events_pending which sets the per-vCPU evtchn_upcall_pending bit and then calls xen/arch/x86/domain.c:vcpu_kick which calls xen/common/schedule.c:vcpu_unblock which calls xen/common/schedule.c:vcpu_wake which finally sets the vCPU to a "runnable" state.
When a domain wishes to wait for an event, it can either call SCHEDOP_block to wait forever for any (unmasked) event, or call SCHEDOP_poll to wait for an event on a small set (specifically less than or equal to 128) of listed ports up to a timeout (like select(2)). Since we don't want to limit ourselves to 128 ports, MirageOS applications on Xen exclusively use SCHEDOP_block. The implementation of SCHEDOP_block simply calls xen/common/schedule.c:vcpu_block_enable_events which calls xen/include/asm-x86/event.h:local_event_delivery_enable to clear the evtchn_upcall_mask bit and then calls xen/common/schedule.c:vcpu_block which performs a final check for incoming events and takes the vCPU offline.
(Updated 2020-10-26. The following information is of historical interest, since MirageOS 3.9.0 our Xen backend has been revised, and only supports PVH mode and does not use mini-os anymore.)
MirageOS applications running on Xen are linked with a small C library derived from mini-os. This library takes care of initial boot: mapping the shared info page and initialising the event channel state. Once the domain state is setup, the OCaml runtime is initialised and the OCaml OS.Main.run callback is evaluated repeatedly until it returns false, signifying exit.
mirage-platform/xen/runtime/kernel/eventchn_stubs.c:evtchn_look_for_work contains mini-os boilerplate to safely interrogate the event channel bits in the shared info page, and copies them to a shadow array which is private to the domain. The function returns true if there is "work to do" i.e. some of the bits in the event channel bitmap were set.
Assuming there is "work to do", mirage-platform/xen/lib/activations.ml:run iterates over the shadow copy of the event channel bits and wakes up any Lwt threads which have registered themselves as waiters. Typically a MirageOS device driver will repeatedly call mirage-platform/xen/lib/activations.mli:after as follows:
let rec process_events channel last_event = Activations.after channel last_event >>= fun latest_event -> ... process_events channel latest_event in process_events channel Activations.program_start
The Activations module keeps a counter and a condition variable per event channel, using the condition variable to wake any threads which are already blocked and the counter to prevent a thread from blocking just after an event has been received.
If there is no "work to do", then control passes to mirage-platform/xen/runtime/kernel/main.c:caml_block_domain which sets a timer and calls SCHEDOP_block. When Xen wakes up the domain, control passes first to a global hypervisor callback which is where an OS would normally inspect the event channel bitmaps and call channel-specific interrupt handlers. In MirageOS's case all we do is clear the vCPU's evtchn_upcall_pending flag and return, safe in the knowledge that the SCHEDOP_block call will now return, and the main OCaml loop will be executed again.
Now that you understand how events work under Xen and how MirageOS uses them, what else do you need to know? Future articles in this series will answer the following questions: