Syntax is like the `file' keyword; e.g.:
object arch/i386/i386/mumble.o [mumble] [needs-flag]
Largely from Michael Richardson in PR 3833, with some changes by me.
new swap system. The dump specification syntax is now more flexible,
and supports constructs like the following:
config netbsd root on ? type ? dumps on ?
- wildcarded root, fstype, and dump device
config netbsd root on ? type ffs dumps on sd0b
- wildcarded root, ffs root fs, always dump on sd0b
config netbsd root on de0 type nfs dumps on wd0b
- mount an nfs root using de0, and write kernel crash dumps
to wd0b
Also, garbage-collect some now unused code, now that swap configuration
is no longer handled by config(8).
reported in PR #3668 from maximum entropy <entropy@tanstaafl.bernstein.com>.
The problem: The unit number is not being conveyed to the configuration
crosscheck with root is a network interface; the device number is set
to NODEV, which sets up the (correct) defaults for swap/dump devices.
We can't change NODEV to a unit number, as the PR's patch suggests, as
this breaks the swap/dump defaults. Rewriting that code is not really
justified, since it's soon to be obsoleted.
The kludge: add an "nv_ifunit" member to struct nvlist, which convey's
a network interface's unit number to the crosscheck routine. It's not
pretty, but it's not terribly invasive, and can be garbage-collected
easily when the swap/dump goo is obsoleted.
. link to the list of locator names from each entry in cfdata[] (for
future use by programs to print out the device tree nicely and/or
for boot-time configuration tools)
be included in object dependencies. config(8) is told to generate
a header for a particular option with the new "defopt" keyword, used
in the files.* system description files. Options that are placed in
header files are not given -D... cpp flags.
This approach allows options to be turned into headers incrementally,
rather than all at once, and allows for non-header options, as well.
and the root device:
- New "file-system" keyword is used to configure file systems into
the kernel.
- New way of specifying root device, which allows root file system
type to always be specified. Examples:
config gennetbsd swap generic
config sdnetbsd root on sd0a swap on sd0b
config nfsnetbsd root on nfs
are replaced by:
config gennetbsd root on ? type ?
config sdnetbsd root on sd0a type ffs swap on sd0b
config nfsnetbsd root on ? type nfs
config lenetbsd root on le0 type nfs
Note that specific network interfaces may now be specified as
the root device.
- swapgeneric.c is no longer used; generate a swap*.c file for each
"config" line in the kernel configuration file.
requires that some ordering requirements are checked by the
back-end C code instead of the parser (dirspecs, maxpart).
* Be more careful to require newline tokens in the grammer where
they are expected, and deal with blank lines, etc. This allows
elimination of a trailing context on newline in the scanner.
* Let the parser set values for "needs-count" and "needs-flag"
instead of making those special cases in the scanner.
* Get rid of '= ' preceeding actions (obsolete yacc syntax)
* Make the scanner not insert an extra newline after includes.
(It was just an accidental side-effect of the ENDFILE stuff.)
on tech-kern. (See man page.) Implementation by Greg Hudson.
Also, remove special case for i386 in vector handling, although this code isn't
actually used any more.
in XXX-marked comments in the recent attachment changes), this was a
long-standing bug in config.
The problem: If a device is attached to a device via an attribute exported
by that device (i.e. foo* at bar0, where 'bar' exports an attribute that
'foo' attaches to), but the device attached to is not present in the
kernel configuration file, AND another device which exports an attribute
that 'foo' attaches to _is_ present (e.g. a device baz0, if one could
specify 'foo0 at baz0'), then: the configuration file will (incorrectly)
be accepted by config, and the resulting ioconf.c will include a bogus
cfdata entry for the device (in the example, 'foo*'). This typically
causes the resulting kernel to crash during autoconfiguration.
The solution: Be much more careful about keeping track of where a device
was attached, and, in particular, if a device was attached to another device,
_always_ keep track of what device it was attached to. Then, when
cross-checking, if the attached-to device isn't present, give up and do not
check attributes. Also, document the process much more thoroughly.
used for checking device attachment names as are used for device name
checking, because device names can be used as attachment names. (Actually,
less strict rules could be used, but there's little point in that.) This
was not a mistake of design, this was just a mistake; i misunderstood
the devbase name checking code.
right/consistent. If you had something like:
file file.c foo bar baz needs-flag
and any one of foo, bar, or baz caused it to be brought into the compile,
in the header you'd end up with:
#define NFOO 1
#define NBAR 1
#define NBAZ 1
even if only one of them were selected. Other headers might have had a
different (inconsistent) set of definitions, depending on whether any of
their components were included, and any files necessary for the unspecified
options would not actually be present in the Makefile files list. The
correct behaviour for the example above if only 'foo' is selected by
the config file is:
#define NFOO 1
#define NBAR 0
#define NBAZ 0
which is what config now does. This bug has been present for a while.
(I don't know for sure that it was present in 4.4-Lite2, but from looking
at the Lite2 config sources, it appears to be there.)
- split softc size and match/attach out from cfdriver into
a new struct cfattach.
- new "attach" directive for files.*. May specify the name of
the cfattach structure, so that devices may be easily attached
to parents with different autoconfiguration semantics.
lists to not have a newline properly emitted. (It was emitting a
newline only if the line position was != 7. However, the only time the
line position was 7 was right after the initial variable assignment
string (e.g. "OBJS=\t") was printed.)
is different than the "machinename" internal variable, read
machinearch's files.${machinearch} and add it to the list of files
for the machine. Also, regardless of whether or not they're different,
create a ${machinearch} sylink (or directory) pointing to the machinearch
include files (or containing them).
wouldn't". this is overridden by (and is a subset of "device-driver".
if you have a "standard" file, you can also do: "standard foo needs-count"
to get a foo.h file. (for the hp300 stuff)
2. "irq ?" sets it to (u_short)-1
3. "irq #" sets it to (1<<#)
4. not specifying an interrupt sets it to 0.
Until someone else comes up with a better scheme, that's the way it is.
If you have a driver that turns the interrupt off, set it to ZERO.
If, after calling XXprobe(), id_irq is still (u_short)-1, that is the same
as if XXprobe() failed.
devices hanging off controllers, any kind of controller.
A device on isa0 is called at probe(self), then attach(self)
A controller on isa0 is called at probe(self), then all it's children with
defined unit numbers are initialized by calling attach(subdev); next all
subdevices with unit ? are initialized by calling attach(sundev).
Almost all device entry points is now like the vax/sun model (intr being
the weird one)
and files.kernel has now been completely replaced.
features supported: not nearly as broken as the stuff before
expression support for dependencies
support for 'requires'
no longer generates lots of stupid unnecessary .h files
broke lots of broken stuff, and forced fixing it.
(docs to arrive later)
added '-k' option for continue even after error
documented '-g'
This patch adds the symbol names to icu.s that vmstat expects
the interrupt counters to be called. It also adds code to config
so that the names of the interrupts are written at the end of vectors.s
so vmstat can report real device names. It also cleans up and enables
the logging of stray interrupts. The counters for false interrupts
are added but the fix for them is not (the fix I have is not done
very good.) A false interrupt is when a device asserts it's interrupt
signal, then removes it before the 8259 can latch it. This is the number
one cause of stray IRQ7's and IRQ15's.
Additional device probe information is now printed. This includes
ending I/O address (many drivers do not return the correct value from a
probe this still needs to be fixed), memory address and size, driver
flags passed in by config.
not identical to barsoom's patch, but I think it's an easier patch.
Check out compile/KERN/ioconf.c to see the effect. (On e kernel with wd and/or
fd of course)
lukem