by default, and can be enabled by adding the SOSEND_LOAN option to your
kernel config. The SOSEND_COUNTERS option can be used to provide some
instrumentation.
Use of this option, combined with an application that does large enough
writes, gets us zero-copy on the TCP and UDP transmit path.
userlevel; this is necessary due to the 601, unlike other 6xx, having
no concept of separated Valid_user vs. Valid_supervisor for BATs.
* When crossing the kernel/userlevel boundary, have platform-provided
hooks set up the two fixed BAT entries, and possibly additional
segment registers to redeem the 601's BAT limitations.
Both of the above are only built if the $MACHINE provides these hooks,
sparing others the pain.
transmit and receive descriptor rings is limited to 256 descriptors.
So, set the if_snd queue length to 256 to let the upper layers queue
lots of packets, and let the driver handle up to 32 of them at a time.
(We should probably make this change to most Ethernet drivers, since
it actually saves some resources.)
* Increase the number of Tx DMA segments from 8 to 16.
* Clean up the way we count "how many times did I get a packet with N
DMA segments".
* Add a missing htole32() in wm_tx_cksum().
* Don't set both RS and RPS in the last Tx descriptor of a packet; just
use RS.
* Add some more information to the watchdog message.
so that they're more useful for arbitrary types of external storage:
* Add an "mbuf *" argument to (*ext_free)(). If non-NULL, (*ext_free)()
is expected to free the mbuf itself. This allows (*ext_free)() to use
the mbuf for bookkeeping (e.g. deferring the work to a helper thread).
If the "mbuf *" argument is NULL, we are assumed to be in a context
which is safe for performing the destructor operation *now*.
* Adjust MEXTREMOVE() and MFREE() routines for above change.
* Update "ade" and "ti" drivers for new semantics.
eliminate a race where another processor could grab the outgoing
process before we were done saving our state into it, with predictable
results.
Bug spotted on i386 by Frank van der Linden <fvdl@wasabisystems.com>.
It is unclear if this realy is needed and if, on which type of cards. I
haven't run accross a card that needs it yet. This may have been just
a copy & pasto from the ISAC interrupt handler carried over to IPAC.
Remove the clear-the-irq-after-enabling it dance (which had bad side
effects on some cards). Instead disble the ISAC receiver when we have
interrupts disabled. Adjust the interrupt handler to properly deal with
subtle differences of the ISAC implementation in IPAC chips.
generally translates to "high interrupt load") -- the old code re-enabled
interrupts in the machine-specific hardware interrupt handler causing the
handler to be re-entered, possible multiple times. Could lead to kernel
stack overflows, and all sorts of mysterious crashes/hangs as a result.
While here, fix up the IP32 interrupt handler code to also not re-enable
interrupts.
Thanks for ideas/comments go to Chuq and Stephen Ma.
* make kloader optional.
* default value of kernel path is KLOADER_KERNEL_PATH.
* options KLOADER_KERNEL_PATH can override kernel path.
* when bootstr is passed (RB_STRING), use it as kernel path.
Note:
(1) Character device major number chages to 49 from 48(in my original
code). So it is necessary to rebuild device files.
(2) Must use Booter 2.0.0a10 or later.
(3) Default mode is cpu busy wait. It is defined by flags 0x1000.
It will be more slow than before.
No flags means hardware interrupt mode. But it might be able to
get no interrupts.
Support machines:
(1) Quadra 630 series
(2) PowerBook 150
Non tested machine:
(1) PowerBook 190 series
Approved by: briggs
if error occurs after status is collected) race condition
when using the status byte to detect completed commands (a command descriptor
could be recycled before the device disconnected), and make the
interrupt routine handling completed commands more efficient (no need to
scan target * lun * tag array any more).
* default value of kernel path is KLOADER_KERNEL_PATH.
* options KLOADER_KERNEL_PATH can override kernel path.
* when bootstr is passed (RB_STRING), use it as kernel path.
* Remove the code that allocates a cluster if the packet would
fit in one; it totally defeats doing references to M_EXT mbufs
in the socket buffer. This drastically reduces the number of
data copies in the tcp_output() path for applications which use
large writes. Kudos to Matt Thomas for pointing me in the right
direction.
request (not always the passed in DMA tag if we try direct-map
and then fall back to sgmap-mapped). Use the actual window
when performing dmamap_sync and dmamap_unload operations.
Fixes DMA resource leak on systems with 2G+ RAM. Thanks to
Matt Thomas for help debugging this.
there may be tagged commands still running when we queue a request sense
command.
Solve this by using 2 DSA entry per LUN
- Now that we have the command DSA before select, we can load T/L/Q in
SCRATCHC. This makes the selection timeout handler simpler.
- Avoid a race condition when setting the free flag in the cmd ring (see
comment in the script)
- don't forget to update the ID in the head of LUN table after a sync/wide
negotiation. This fixes the command timeout at the first data command
after negotiation (the bus reset handler did update the ID properly,
so subsequent commands were OK).
- for DMA interrupts, clear fifo if it's not empty. Leaving the fifo dirty
would prevent subsequent interrupts from coming in.
- Various improvements in debug messages
- misc cleanups.
* pmap_protect(): write back the range when doing a r/w -> r/o
transition. (Still leave the block concerned with this in
pmap_clean_page() disabled, for now.)
* pmap_pte_init_xscale(): Disable read/write-allocate for now, until
we figure out why sometimes cache lines of NULs get deposited into
file data. Also, make sure ECC protection of page table access is
disabled for now.
* xscale_setup_minidata(): Make sure the mini-data cache is configured
write-back with read/write-allocate.
scheduler slot. This costs a few more instructions but divide the size of the
scheduler ring by 2, saving 1k of onboard RAM (a bus with 15 devices would
overflow the on-board RAM by 128 bytes).
make -V FILES
from being useful (and given that every other variable can be
extracted using make -V, the behaviour was unusually inconsistent
given that the original reason for clearing it doesn't seem to be
relevant anymore)
- use <bsd.prog.mk> instead of directly including <bsd.files.mk>
(and possibly <bsd.man.mk> or <bsd.own.mk>)
- remove obsolete NOPROG
- Add support for DT transfers (aka Ultra/160) in esiop
Note that DT transfers are not enabled for 53c1010-33 rev 0 yet; if I trust
FreeBSD it has a bug which prevent them to do DT properly.
From the same source there may be issues with some revs of 53c1010-66.
So check for version 3, not 4 when looking for DT support.
This should be safe as these bits are reserved for older devices, they
should be set to 0 when not supported.
closed, open those fds to /dev/null.
XXX: This needs to be fixed in a better way. The kernel should not need to
know about /dev/null or special case 0, 1, 2.
taken from OpenBSD. Test hardware kindly provided by Intel. This still needs
management bits, and doesn't support older controllers, but that shouldn't
be hard to fix.
after the selection timeout is posted but the number executed isn't
reliable. So wait for MSG_OUT rigth after the select so that the state of
the script when the interruption is handled is known.
For this add another indirecton: the DSA in the LUN table points to
a table of DSA indexed by the tag number when tagged command is in use.
For non tagged command, the LUN DSA still points to the tables describing the
xfer directly.
- initial support for MAP_AUTOGROW flag. When mapping beyond the end of file is
requested with MAP_AUTOGROW, if pages beyond the end of file are touched, the
file should be resized. We are not able to emulate this yet, so we immediatly
resize the file to fit the whole mapping.
- implements mmap64
yet.
If is restricted to SIOP which implement the load/store instruction, and
has 10 scratch registers (basically, 825 and newer, possibly 770).
It implements a different interface between host and script, using a real
ring for command starts, and improved support for reconnect which will allow
256 tag per device. It uses interrupt on the fly to signal complete command,
which allows several commands to be serviced per interrupt and doesn't require
the script to stop to signal command completion.
