[1] add EN_ENIDMAFIX - the byte-aligner on the ENI version of the card
appears to lose under heavy load so avoid using it. see comment
in midway.c for full details.
note that the Adaptec version of the card works properly under
load.
detected by Kenjiro Cho <kjc@csl.sony.co.jp> [confirmed by chuck]
fix by chuck & kjc
[2] update some freebsd specific code [from kjc]
[3] for circular buffers: ensure there is always one free slot so
that we can easily tell the difference between a full and
empty list. re-structure a few loops to reflect this.
fixes a problem with mbufs being free'd while still in DMA,
and makes drive a bit more robust.
detected and fixed by kjc
'inline' as well as static. mark prototypes for static inline functions
as possibly unused (with __attribute__ ((unused))), to avoid generating
warnings when compiling without optimization but with most ports'
default warning flags. Clean up prototype list spacing, and make it more
consistent.
Anne Hutton <hutton@isi.edu>]:
- add support for Adaptec 155 PCI ATM cards (e.g. ANA-5940)
- add sc->is_adaptec to handle differences between cards.
- break out MID_MK_TXQ/MID_MK_RXQ seperate macros to handle
the new Adaptec format TXQ/RXQ.
- adjust en_dqneed to return 1 on ADP (since the Adaptec can
DMA anything in one DRQ/DTQ!)
- add hook for a bus specific reset function (adaptec has
a seperate reset register that needs to be hit when
resettting the midway).
- adjust DMA test to not worry about burst sizes on the
adaptec (since it handles it all for us!) and to handle
the new DTQ/DRQ format.
- add Adaptec DMA support to en_txlaunch() and en_service()
BUG FIXES:
- fixed receiver panic under heavy load ("lost mbuf in slot 0!").
when the reassembly buffer overflows, the T-bit is set in
the RDB and the data field is empty. en_service() sets up
a 4-byte (RDB size) dummy DMA without IF_ENQUEUE. but the
recv intr handling in en_intr() always does IF_DEQUEUE.
as a result, a successive recv intr loses its mbuf and
leads to a panic. the solution is to only IF_DEQUEUE if
the interrupt has non-zero length (indicating that there
is an mbuf to get). in order for this to work, EN_DQ_MK
must always be non-zero. we do this by or'ing in an unused
bit (0x80000).
reported by: Kenjiro Cho <kjc@csl.sony.co.jp>
- fix setting of transmit channel when txspeed[] is non-zero
(e.g. traffic shaping). the old scheme didn't work
properly (it allowed the same VCI to use multiple tx channels
thus defeating the txspeed[] parameter). the new scheme
statically assigns a VC to a channel when txspeed[] is set.
[note that the code to set txspeed[] isn't in the driver right
now since a MI interface to do this hasn't been made yet]
we add sc->txvc2slot[] and sc->txslot[n].nref for this.
reported by: Kenjiro Cho <kjc@csl.sony.co.jp>,
Milind M Buddihikot <milind@ccrc.wustl.edu>,
Dong Lin <dong@eecs.harvard.edu>
- when doing SRAM copies, be sure to round up the length to the next
largest word (otherwise the driver will try to do a byte clean
up DMA and then get an ID error interrupt).
MINOR CLEANUPS:
- clean up loops in DMA test
contributed by: Kenjiro Cho <kjc@csl.sony.co.jp>
- restructure and cleanup of en_read/en_write macros/inlines
- clean up some byte ordering stuff so that we are consistant throughout
the driver
- if aal5 frame has a CRC error then the length field in the aal5 trailer
may not be valid, so we can not use it [and we must dump the frame]
[Yuhang Sun <sunyh@dworkin.wustl.edu> & chuck@ccrc.wustl.edu]
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.
- No more distinction between i/o-mapped and memory-mapped
devices. It's all "bus space" now, and space tags
differentiate the space with finer grain than the
bus chipset tag.
- Add memory barrier methods.
- Implement space alloc/free methods.
- Implement region read/write methods (like memcpy to/from
bus space).
This interface provides a better abstraction for dealing with
machine-independent chipset drivers.
- adjust txhiwat and mindma params a bit
- fixed a couple of incorrectly labeled panic calls
- the "location" was being calculated incorrectly in some cases (forgot
to subtract off MID_RAMBASE). this only caused problem when trying
to change the size of the tx/rx buffers (e.g. to 64KB).
- fixed possible non-aligned DMA burst in the starting byte burst case.
(e.g. if we could DMA 3 bytes, but only have 2 it is not legal
for us to use MIDDMA_BYTE2 mode).
- opt: on tx: try and avoid flushing the internal buffer by padding out the
length of the last mbuf a bit (if possible)
- merged multiple DRQ/DTQ ADD macros into a single DRQ and a single DTQ
macro with a uniform interface to make the code simpler and easier to read.
- en_start: only update atm_flags if EN_MBUF_OPT is enabled (which it
should be)
- for alburst: make sure we don't DMA more bytes than we need (on both
tx and rx). if the alburst is larger than we need, drop to
MIDDMA_WORD mode.
- major change: enable the use of byte and 2 byte DMA on the trasmit side.
this allows us to DMA from non-word sized/aligned mbufs directly.
[the old code would always call en_mfix which would copy (or move) the
data in order to ensure proper alignment... it turns out TCP gives
us non-word sized/aligned mbufs when it is retransmitting, so we needed
to handle this case more efficiently.] the following functions
were changed to make this work:
- en_dqneed: add an arg to let us know if we are transmitting or not.
if we are TX, then we must take into account byte DMAs when
estimating the number of DTQs we will need for a buffer
- en_start: only mfix mbufs if DMA is disabled
- en_txdma: only set launch.nodma if we have en_mfix'd the mbuf chain
also, we may need a DTQ to flush the chip's internal byte buffer
- en_txlaunch: only attempt a copy if we have the proper alignment.
add byte dma code for the front and end of the buffer.
make sure the internal dma buffer is flushed out.
- stats: keep track of how many times we have to use byte sized DMA
midwayreg:
- add byte/2byte DMA defines
midwayvar:
- add new stat counter to monitor less-than-word lengthed DMA
- en_mfix shouldn't touch M_EXT mbufs. change to avoid this [at
the expense of allocating a new cluster mbuf to copy to].
XXX: en_mfix is getting called more often than I hoped it would
(seems to happen when TCP retransmits... we get all sorts of odd
sized, odd lengthed data mbufs, yuck). i may revise the xmit
DMA code to use non-word sized dma.
- don't try and add a PDU trailer (or TBD for that matter) to an M_EXT
mbuf. when M_EXT is set, the data area of the mbuf can be shared
with other mbufs (see m_copym()), and writing to one M_EXT mbuf may
invalidate another M_EXT mbuf (saw this with PDU trailers).
- be sure to validate the length value from the PDU trailer so that it
doesn't throw us out of sync with the chip (fixes yet another source
of the dread "en_service: id mismatch" panic).
- on recv, don't bother allocating an mbuf if mlen is zero since we are
only going to toss the data anyway.
to ignore inbound data sizes less than the threshold. otherwise,
when connected to a video aal0 input the driver gives us a massive
stream of 56 byte mbufs each with one aal0 cell in it and the
system just can't keep up with it, especially if the socket buffer
size is large [it hangs until you turn off the video source].
fixes: when turning off a vc, try and check the freshest copy of the mode
when seeing if we need to enter the "drain" state. also, don't
panic if we get unexpected rx interrupt on a VCI (instead make sure
the VC is off, print a warning, and move on!).
- change rxso into a void handle ... no need to know details at this low
level.
fix:
- forgot to init "vci" during ENOTHER_DRAIN case of RX DMA interrupt
[could cause a vm_fault on native mode connection close depending
on needing to drain out the VC and also the random value of vci]
- more stat counters (rxmbufout -- mbufs, ttrash -- trashed RBD)
- add swsl_size to keep track of size of software slist
- revise when we call en_service
- fix handling of circular lists (was getting full vs empty confused in
extreem cases...)
- native mode atm: remove sbappend code: can't do it here since
socket buffer is protected only by splsoftnet and that doesn't
mask enintr off. forces us to switch to a two level interrupt
scheme. revise call to atm_input to reflect that.
[1] rxslot "cur" pointer wasn't updated if we copied all data and
didn't have any fill. this could cause a service id mismatch
panic under load.
[2] recv of aal0 forgot to include the cell header in the data length
calc [add cell header length to midwayreg.h]
[3] fix control info "chop" for raw mode and aal0
[4] handle rxsb better for native atm connections
(XXX: still working on this ... not complete and needs revision)
thorpej