way of doing business- modulo some startup spasms and peculiarities
of the way kthreads are started (*after* configuration, weird) and
some strangeness with the freeze/thaw code, what now happens is
that any of Loop Down, LIP, Loop Reset or Port Datbase or Name
Server Database Changed ASYNC events cause the queues to freeze
for this channel. The arrival of a Loop UP is not relevant.
What *is* relevant is that the Port Datbase or Name Server Changed
async event indicate that it's okay to go and (re)evaluate the
state of the FC link and (re)probe local loop and fabric membership.
We have a kthread do this because it's *sooooo* much nicer to be
able to sleep while doing the 130-250 mailbox commands it'll take
to re-evaluate things.
When the state is well known again, we can unfreeze the channel
queues. Then, as commands start arriving, we simply can start them
or bounce them with XS_SELTIMEOUT (if the device in question has
gone away). Previously, we did lazy evaluation, which meant that
if a change occurred, we would wait until the very *next* command
to go rebuild stuff.
The reason this is not sensible is:
a) Even with sleeping, you can hang up your system because you might be
making some poor stat(2) call pay the price of re-evaluating the whole
fabric.
b) If we ever really want to get to dynamic attachment/detachment, we
should find out sooner, rather than later, where things get to.
Split off ispminphys_1020 from ispminphys- a 1020 has a 24 bit limit-
not anything newer.
Re-enable LIPs and Loop Resets as async events- this allows the outer
layer to set policy about them.
Roll platform major && minor. Remove bogus waitq (no longer used).
Remove callout entry in softc (no longer used). Define some shorthands
for channels. Clean up a variety of cruft left over from the
thorpej_scsipi changeover.
Add SD_IO_TIMEOUT (an option in opt_scsi.h) also, which defaults to the
normal 60 * 1000 timeout for normal read/write commamds. This allows you
to shorten or lengthen the timeout as needed.
- Add some additional config block bits for the i82558/i82559.
- Change the config block template to only fill in the must-be-one
reserved bits, leave fxp_init() to fill in all the important things.
- On the i82558/i82559, we can use "Receive Long Frames" rather than
"Save Bad Frames" to support the VLAN MTU.
- Use 802.3x flow control on the i82558/i82559. This is all handled
transparently by the hardware. When in promiscuous mode, allow
wire-watchers to see the flow control frames.
- Use the Extended TxCB format on the i82558/i82559. This places two
Transmit Buffer Descriptors directly in the TxCB, which should cover
the vast majority of packets transmitted. This saves PCI transactions
that would otherwise be required to fetch the TBD list.
With the tansmit queue length changes from yesterday, this gets us up to
11.5MB/s TCP transmit speed, out of an absolute maximum of 12MB/s possible
on a 100Mb/s link.
some devices can't handle tag number larger than some values and always
reject commands with QUEUE FULL if the tag number is larger than this.
Under some circonstances the scsipi system may not decrease periph_openings
(as a workaround of other odd behavior) and we may end up requeuing the command
with a hight tag value again, and the driver could loop on this.
Workaround: in case of queue full, decrease periph_openings to
min(periph_active - 1, periph_openings - 1) so that, after some iteration,
periph_openings is less than the max tag value acceptable by the device.
Solves the problem with tagged queuing on ncr53c9x for me.
SIP_EVENT_COUNTERS.
- Decrease SIP_NTXSEGS to 8 and increase SIP_TXQUEUELEN to 256.
This gives us 256 transmit jobs with only twice the descriptor
memory footprint. This prevents stalling transmissions due to
lack of transmit jobs under heavy load.
- Increase SIP_NRXDESC to 128 so we don't run out of receive descriptors
under extreme load.
- As recommended by the manual, leave one descriptor in the transmit ring
with OWN clear to prevent wrap-around.
- Even though the manual claims that driver software must initialize
TXDP if there are no pending transmissions when a new transmission
is to be initialized, doing so causes serious performance degredation
on the DP83820 under heavy load. Not initializing TXDP also increases
performance on the DP83815.
All of this gets us up to ~90Mb/s on both the DP83820 and DP83815 when
connected to a 100base network.
This fixes a `hang' on shutdown when the `Synchronise Cache' command
is sent to a disk. (It uses polled mode because it is called from inside
a shutdown hook).
interrupt priority, to avoid allocating additional mbufs on
splsched(). Move some other misc bits out of splched()-guarded
section. This should make the code a bit more friendly wrt e.g.
clock interrupts. Sligh performance improvement was observed with
this change (like 1-2% on slow machine), even.
This also fixes ancient kern/2478 by Bill Sommerfeld.
