Hi All,
I've been struggling a bit deciding how to model Xen's max_event_channels and
e820_host xl.cfg(5) settings. For max_event_channels the man page says:
Limit the guest to using at most N event channels (PV interrupts). Guests
use hypervisor resources for each event channel they use.
The default of 1023 should be sufficient for typical guests. The maximum
value depends on what the guest supports. Guests supporting the FIFO-based
event channel ABI support up to 131,071 event channels. Other guests are
limited to 4095 (64-bit x86 and ARM) or 1023 (32-bit x86).
So event channels are PV interrupts that are used by PV devices and timers, and
as inter-processor interrupts. My initial thought was to add a maxEventChannels
attribute to the xenbus controller, similar maxGrantFrames. The latter was added
to allow specifying the max shared pages a guest could use, which in conjunction
with event channels provides data transfer mechanism between front and backends
for PV drivers. My only hesitation is that event channels are also used for
inter-processor interrupts, which is a bit outside the notion of a virtual
controller for Xen PV devices. However, the first sentence in the xenbus wiki
[1] states:
In practice, the bus is used for configuration negotiation, leaving most
data transfer to be done via an interdomain channel composed of a shared
page and an event channel.
So perhaps it is tolerable to consider max_event_channels an attribute of xenbus
that is specified at domain creation, to then be used by the guest however it
pleases up to the max value?
e820_host is a bit trickier. For this setting, which is PV-specific, the man
page says:
Selects whether to expose the host e820 (memory map) to the guest via the
virtual e820. When this option is false (0) the guest pseudo-physical
address space consists of a single contiguous RAM region. When this option
is specified the virtual e820 instead reflects the host e820 and contains
the same PCI holes. The total amount of RAM represented by the memory map
is always the same, this option configures only how it is laid out.
Exposing the host e820 to the guest gives the guest kernel the opportunity
to set aside the required part of its pseudo-physical address space in order
to provide address space to map passedthrough PCI devices. It is guest
Operating System dependent whether this option is required, specifically it
is required when using a mainline Linux ("pvops") kernel. This option
defaults to true (1) if any PCI passthrough devices are configured and
false (0) otherwise. If you do not configure any passthrough devices at
domain creation time but expect to hotplug devices later then you should
set this option. Conversely if your particular guest kernel does not
require this behavior then it is safe to allow this to be enabled but
you may wish to disable it anyway.
I'm tempted to unconditionally enable this setting. It is required for pvops
kernels and apparently harmless for other PV kernels. I asked one of the Xen
devs about any downsides to always enabling e820_host, to which he replied
"Scattered memory blocks inside the guest, possibly leading to slightly higher
overhead. But nothing really severe afaics.".
If there is a strong desire to model this setting it might best be placed under
hypervisor features, e.g.
<features>
<xen>
<hoste820 state='on'/>
</xen>
</features>
Opinions and/or comments are much appreciated.
Regards,
Jim
[1] https://wiki.xen.org/wiki/XenBus
