Since KLD modules are not compiled with the same options headers that
the kernel is compiled with,
they must not be loaded into a kernel compiled with the
PAE
option.
Many devices or their device drivers are not capable of direct memory access
to physical addresses above 4 gigabytes.
In order to make use of direct memory access IO in a system with more than
4 gigabytes of memory when the
PAE
option is used,
these drivers must use a facility for remapping or substituting physical
memory which is not accessible to the device.
One such facility is provided by the
busdma
interface.
Device drivers which do not account for such devices will not work reliably
in a system with more than 4 gigabytes of memory when the
PAE
option is used,
and may cause data corruption.
The
PAE
kernel configuration file includes the
PAE
option, and explicitly excludes all device drivers which are known to not work
or have not been tested in a system with the
PAE
option and more than 4 gigabytes of memory.
Many parameters which determine how memory is used in the kernel are based on
the amount of physical memory.
The formulas used to determine the values of these parameters for specific
memory configurations may not take into account the fact there may be more
than 4 gigabytes of memory, and may not scale well to these memory
configurations.
In particular,
it may be necessary to increase the amount of virtual address space available
to the kernel,
or to reduce the amount of a specific resource that is heavily used,
in order to avoid running out of virtual address space.
The
KVA_PAGES
option may be used to increase the kernel virtual address space,
and the
kern.maxvnodes
sysctl(8)
may be used to decrease the number of vnodes allowed,
an example of a resource that the kernel is likely to overallocate in
large memory configurations.
For optimal performance and stability it may be necessary to consult the
tuning(7)
manual page, and make adjustments to the parameters documented there.
