>One control block per target is insufficient if you have a full complement
>of targets attached and access those simultaneously (like in a ccd(4) array).
>We (now) allocate (somewhat arbitrarily) three per target.
>Noticed by Marshall Midden.
(1) after removing a shutdown hook (in shutdownhook_disestablish()),
free it. We created it, we have to free it. Without this,
shutdownhook_disestablish() leaks memory.
(2) in doshutdownhooks(), before running each hook, remove it from the
shutdown hook list. This makes sure that every hook is tried once
(because doshutdownhooks() is called from before rebooting, and
a fault in a shutdown hook will cause doshutdownhooks() to be called
again), but prevents the hooks from potentially being run infinitely
(as used to be possible, in the above-mentioned situation).
(e.g. from 0x3bc to 0x3bf, for example). Others may require more,
but until there's some special handling for them, 4 should be returned
so that on systems with I/O port accounting, the unit at 0x3bf can be
properly mapped, etc. (OK'd by mycroft.)
that their child busses can be attached after the PCI bus
autoconfiguration for their parent bus is done.
This works because:
(1) there can be at most one ISA/EISA bridge per PCI bus, and
(2) any ISA/EISA bridges must be attached to primary PCI
busses (i.e. bus zero).
That boils down to: there can only be one of these outstanding
at a time, it is cleared when configuring PCI bus 0 before any
subdevices have been found, and it is run after all subdevices
of PCI bus 0 have been found.
This (or something like it) is needed because there are some (legacy)
PCI devices which can show up as ISA/EISA devices as well (the prime
example of which are VGA controllers). If you attach ISA from a
PCI-ISA/EISA bridge, and the bridge is seen before the video board is,
the board can show up as an ISA device, and that can (bogusly)
complicate the PCI device's attach code, or make the PCI device not be
properly attached at all.
This could be done with machine-dependent code, but as more ports
add support for PCI (and PCI-ISA/EISA bridges) more will need it.
The i386 port could (perhaps should) be converted to use it as well.
devices actually do make sense on indirect-config busses, because you
might be able to have more than one of the busses! In addition, they're
useful because they don't require unit numbers to be wired down, so you
could e.g. have vga* at indirect? and vga* at direct?, and have the first
one found be unit number zero. Finally, devices which can divine their
own ports numbers, etc., actually should be cloning, even if you know you'll
only have one bus that they can live on.
don't machine check when a PCI Master Abort is signalled. This can
happen, for instance, when configuration space for a device that isn't
present is examined. When this is detected, act like we normally would
when machine checks are posted while examining nonexistant devices.
enabled (from the attach routine), and add comments as to why.
Some PALcode apparently 'saves' a clock interrupt for the kernel,
and if the clock interrupt handler is enabled at attach time, it
will be run when that interrupt hits, i.e. right after the spl0()
at the end of autoconfiguration. That would cause hardclock to be
run, but proc0's p_stats isn't set up by then, which would cause
hardclock to crash.
rather than and-ing 16G-1. That just strips the k0seg bits, rather
than making the false assumption that the physical address is going
to be in the lower 16G. That doesn't apply for CIA device-space
addresses, for instance.
even if PCI and the IDs are right), just for sanity, before declaring
success. Split the single 0x3b0 -> 0x3df allocation into three seperate
ones: 0x3b0 -> 0x3bc (leaving the 4 ports available for lpt),
0x3c0 -> 0x3cf, and 0x3d0 -> 0x3df. The former chunk has to be split
off if the lpt can exist there, and it's sort-of pretty to have each
group (based on second hex digit) have its own handle.
These alternative macros have a workaround for the STM^ bug in revision < 3
StrongARM CPU's that causes incorrect register saving if a cache line fill
is in progress during the STM.
a podulebus.
Make sure the podulebus driver conforms to the Acorn expansion card
specification:
- Probe the podule bus using sync access cycles rather than slow access
cycles.
- Read the podulebus header/ROM using sync access cycles rather than slow
access cycles
of targets attached and access those simultaneously (like in a ccd(4) array).
We (now) allocate (somewhat arbitrarily) three per target.
Noticed by Marshall Midden.
1. fix possible hang in en_txlaunch(). when attempting to extend
the length of an mbuf to avoid a flush we should extend it
by cnt [which is ((need - len) % 4)] rather than 4 - cnt.
also, add an EN_DEBUG printf() when we pad/FLUSH a buffer
to help with debugging/understanding what the driver is up to.
2. use interface packet counters
3. when turning off a recv VCI we recompute the new mode. make sure
we don't include the "in service" bit in the new mode, otherwise
a VCI may appear "hung" if you turn it off while a service
interrupt is pending.
4. when shutting down a VCI that is still receiving data, don't bother
going into "drain mode" if only the hardware in service bit is
set (otherwise the VCI may get "hung" in drain mode).
as a result of this we may get "unexpected rx interrupt" messages
which are not really an error, so put this printf in EN_DEBUG.
5. be sure to zero txspeed[lcv] when enabling a VCI (start at full
speed). (hooks for setting txspeed[] are currently not in
the driver, but we are playing with it locally).
credits:
#1: Detected by: Zdenek Salvet <salvet@horn.ics.muni.cz>, fix by me.
#2: Contributed by: Zdenek Salvet <salvet@horn.ics.muni.cz>
#3,#4,#5: Detected by: Milind M. Buddhikot <milind@dworkin.wustl.edu>,
fixed by me.
If not compiled with -D_KERNEL, include different includes and
do so macro magic so that this will fit sanely into test harnesses.
When used in user-land, this should be compiled with -D_EXTENT_TESTING.
Bug fixes:
(extent_insert_and_optimize) You can't do things like:
LIST_REMOVE(elem->...le_next, ...);
free(elem->...le_next, ...);
They just don't work (and will corrupt your list and/or malloc free list).
(extent_alloc_region_descriptor) Unless you wait, malloc can fail.
Don't accidentally deref a potentially-NULL pointer.
2841, plus some fixes to make the patches work on the Alpha. Seems to
improve the NCR driver a lot. We probably should try to incorporate
any updates that have happened since, too.
Interrupts for irqs with handlers are counted as normal, even if none of
the handlers claim the interrupt.
Should all be redone with hierarchical event counters. that count
hardware events and invocations of each handler separately.
* Increment cnt.v_intr for normal interrupts (bugfix).
to initialize conspa was invalid on non-040 machines, since we don't
know which MMU we're using on those, yet. On the other hand, by the time
we get to check_video() on the 040 machines, the MMU has been disabled.
Compromise: get the video PA in setmachdep() on the 040, and in
check_video() on the others. (This is not optimal. *sigh*)
ioctl and mmap routines take a void *, rather than a struct device *,
so that they can be set up to work more easily when using a 'struct device *'
isn't appropriate. Add a cookie (void *) to be passed to the mmap and
ioctl routines. Rename a few struct members, and shuffle them into
a more sensible order.
front-ends. Unfortunately, because of the way ISA and PCI are currently
probed, if you have a PCI VGA board in your machine and both drivers
in your kernel, the ISA VGA driver may accidentally match the PCI board.
For now, the only solution to this is to not put both drivers in
the GENERIC kernels.
Earlier, we re-complained about excessive token losses about once a minute.
However, on small ARCnet networks, the token will also be lost if only one
station (us) is active (ifconfig'd down interfaces don't take place in the
token exchange), and our syslog would be filled with repeated messages about
this condition.
Our new code only complains once, starts a timeout() each time a token
regeneration is reported by the chip, and generates a log message about
the regenerated token if it was stable for 15 seconds.
buffer out of the hardware (it is invalid!).
This fixes hangups due to spurious rx interupts.
XXX Maybe I should completely reset the hardware in this case?
- New variables: biosextmem, biosbasemem, nkpde
- Above can be set by using the options BIOSEXTMEM, BIOSBASEMEM,
and NKPDE respectively (EXTMEM_SIZE is now called BIOSEXTMEM).
When preset this way, they won't be filled in / calculated.
- Readable by sysctl using machdep.nkpde, machdep.biosbasemem
and machdep.biosextmem.
- nkpde is calculated as:
min(NKPDE_MAX, NKPDE_BASE + (biosextmem >> 10) * NKPDE_SCALE)
Where NKPDE_MAX is 31, NKPDE_BASE is 4, NKPDE_SCALE = 1.
* Add a slot in the softc, ep_pktlenshift, to indicate that
a board implements ``large packets'' (FDDI-sized frames) and
therefore implicitly shifts the argument to packet-length commands
by two bits to fit a 13-bit length in an 11-bit command argument field.
* Change the reset of station-mask filterto be independent of bus-type
(e.g.,. for 3c515)
* Replace the "2044" constant used to disable TX threshold interrupts
on 3c509-generation cards with a #define. Use 2047 instead,
since that works on the large-packet cards (Vortex, Demon, Boomerang)
and 2044 doesn't.
* Add code to epconfing to probe for large-packet support,
set the softc appropriately, and shift packet sizes in commands
by ep_pktlenshift to compensate for the implicit shift.
* Tweak the epconfig() code to avoid hanging on startup.
Set the Tx threshold to interrupt when there's space for
a maximal-sized ether frame, then do epinit() then epstop(),
instead of just the Tx and Rx reset. Required on at least a 3c595.
Tested on a 3c595, but not on 3c509-generation hardware.
We test this condition AFTER the interupt condition, because it is only
disabled during autoconfiguration.
Without this patch, if autoconfiguration takes long enough, eventually a
timer interupt will call hardclock() and panic the machine.
the Digital Semiconductor AlphaPC 164 Motherboard Technical Reference
Manual. This may not work, but it's the best i can do with the
documentation I have.
is set in the RPB's rpb_variation field. This fixes a bug where machines
(e.g. the 3000/900) would see that they could touch memory where a built-in
PMAGB-BA, assume that it was there, and panic later because the memory they
were accessing didn't look like PMAGB-BA registers (because it wasn't).
gcc thinks that the 'q' modifier describes a "long long", and so -Wformat
whines when printing with 'q' on the alpha, since int64_t-sized types are
done with variations on "long" rather than "long long".
(This fixes problems with the printf format fixes i checked in yesterday.
ptrdiff_t is an 'int' on the i386 but a 'long' on the alpha, so the cast
really is necessary... *sigh*)
- The functions that implement them and the argument names are
prepended with "sys_".
- Optional systems calls are "UNIMPL" if the support is not being
compiled into the kernel.
-Werror -Wall -Wstrict-prototypes -Wmissing-prototypes
-Wcast-qual -Wno-format
(if not already set). -Wno-format is necessary because of the few
bits of code in the tree that use NetBSD's special kernel printf()
formats, and because there are a few bits of code that try to use the 'q'
modifier to print int64_t's (but int64_t's are only 'long' on the
alpha, not 'long long').
representing the names of those bits, prints them into a buffer
provided by the caller, and returns a pointer to that buffer.
Functionality is identical to that of the (non-standard) `%b' printf()
format, which will be deprecated.
Rename the non-exported function ksprintn() to ksnprintn(), and change
it to use a buffer provided by the caller, rather than at static
buffer.
On sun4/sun4c with a virtual address hole (starting at 512MB), this seems
like a reasonable compromise: about 196MB left for shared libs and sysv-style
shared memory segments.
On the sun4m the limits can easily be made larger: consider turning
MAXTSIZ and MAXDSIZ into tunable variables..
deleting (and cleaning up) N pages of #ifdefs in machdep.c. While we're
at it, make the cpu type options generate flags, and check them instead of
kernel Makefile defines.
Fixes profiling for non-underscore-prepending toolchains
(elf, e.g., shared libs), and breaks a.out/ecoff toolchains.
May break mips kernel profiling too. Needs more thought, since the
original intent of __mcount vs ___mcount on mips date back to pre-1.0 days.
21164 + 21172 box, with Yet Another Way of doing interrupts. The
interrupt mapping and handling code hasn't been written yet, and none
of this code has been tested. (Checkpoint of work in progress.)
* handle interpreters with nonzero virtual address of entry-point:
subtract p_vaddr from computed entrypoint, as the mips elf exec did.
* Add #ifdef ELF_INTERP_NON_RELOCATABLE/#endif around the code
that tries to choose a `good' address at which to load an interpreter,
if none was set by the emul probe function.
(the address chosen could be improved to avoid fragmenting the
process virtual address space).
* define ELF_INTERP_NON_RELOCATABLE in machine/elf_machdep.h for mips CPUs,
which currently use a GNU-derived ld.so.
ELF_INTERP_NON_RELOCATABLE is not necessary for native NetBSD/alpha ELF
binaries. It may be required for GNU-derived ELF dynamic loaders (Linux/i386?)
- Add reset attribute sequences for bold, underline, and reverse
- Add insert/delete line sequences
- Recognize but ignore G0/G1 character set sequences
Performance:
- Speed up clear_line() when clearing the entire line
Also, some stylistic changes for checking/setting scroll region bounds
and in putc_getpars().
These changes make the ITE work properly with the 1.2 termcap file.
Keep queue of pending sockets in a double linked list. Previously,
a singly linked list was used, giving O(N) insertion/deletion times,
and was a major time consumer for sockets with large pending queues.
The double linked list give O(C) insertion/deletion times with only
a small cost in complexity.
Since a socket can be on, at most, one queue at a time, both so_q and
so_q0 can safely be used as (forward and backward, respectively) queue
pointers.
Submitted my Matt Thomas <matt@3am-software.com>, a long time ago.
(Geez, I've been running with this patch for _months_, and had completely
forgotten about it!)
- The boundary argument to bus_space_alloc() should be a bus_size_t, not
a bus_addr_t.
- The buffer arguments in the "multiple write" methods should have
const qualifiers.
And one from me:
- Make bus_space_barrier() eat up the arguments passed to it so that
the compiler doesn't needlessly whine.
the variable `dumpsize' (used by savecore(8)) and that the CPU specific data
gets dumped first.
Also, squash a typo that prevented the crash dump to be shifted towards the
end of the swap partition. In fact, the entire dumpconf() routine has been
redone, mostly by pasting large chunks from the alpha port.
have an Apple Sound Chip. Make sure that we have configured the device
before allowing accesses to ASC memory. Among other things, this
prevents the 660AV and 840AV from getting a kernel bus error when trying
to beep on the console.
everything before the beginning of the "a" partition on a disk
is read-only except when accessed via RAW_PART. (a problem if
"a" isn't at the front of the disk). pointed out by pk back
in july.
matthias