checksum offoload by downloading AIC-6915 firmware. Changes are
o Header file cleanup.
o Simplified probe logic.
o s/u_int{8,16,32}_t/uint{8,16,32}_t/g
o K&R -> ANSI C.
o In register access function, added support both memory mapped and
IO space register acccess. The function will dynamically detect
which method would be choosed.
o sf_setperf() was modified to support strict-alignment
architectures.
o Use SF_MII_DATAPORT instead of hardcoded value 0xffff.
o Added link state/speed, duplex changes handling task q. The task q
is also responsible for flow control settings.
o Always hornor link up/down state reported by mii layers. The link
state information is used in sf_start() to determine whether we
got a valid link.
o Added experimental flow-control setup. It was commented out but
will be activated once we have flow-cotrol infrastructure in mii
layer.
o Simplify IFF_UP/IFCAP_POLLING and IFF_PROMISC handling logic. Rx
filter always honors promiscuous mode.
o Implemented suspend/resume methods.
o Reorganized Rx filter routine so promiscuous mode changes doesn't
require interface re-initialization.
o Reimplemnted driver probe routine such that it looks for matching
device from supported hardware list table. This change will help to
add newer hardware revision to the driver.
o Use ETHER_ADDR_LEN instead of hardcoded value.
o Prefer memory space register mapping over I/O space as the hardware
requires lots of register access to get various consumer/producer
index. Failing to get memory space mapping, sf(4) falls back to I/O
space mapping. Use of memory space register mapping requires
somewhat large memory space(512K), though.
o Switch to simpler bus_{read,write}_{1,2,4}.
o Use PCIR_BAR macro to get BARs.
o Program PCI cache line size if the cache line size was set to 0
and enable PCI MWI.
o Add a new sysctl node 'dev.sf.N.stats' that shows various MAC
counters for Rx/Tx statistics.
o Add a sysctl node to configure interrupt moderation timer. The
timer defers interrupts generation until time specified in timer
control register is expired. The value in the timer register is in
units of 102.4us. The allowable range for the timer is 0 - 31
(0 ~ 3.276ms).
The default value is 1(102.4us). Users can change the timer value
with dev.sf.N.int_mod sysctl(8) variable/loader(8) tunable.
o bus_dma(9) conversion
- Enable 64bit DMA addressing.
- Enable 64bit descriptor format support.
- Apply descriptor ring alignment requirements(256 bytes alignment).
- Apply Rx buffer address alignment requirements(4 bytes alignment).
- Apply 4GB boundary restrictions(Tx/Rx ring and its completion ring
should live in the same 4GB address space.)
- Set number of allowable number of DMA segments to 16. In fact,
AIC-6915 doesn't have a limit for number of DMA segments but it
would be waste of Tx descriptor resource if we allow more than 16.
- Rx/Tx side bus_dmamap_load_mbuf_sg(9) support.
- Added alignment fixup code for strict-alignment architectures.
- Added endianness support code in Tx/Rx descriptor access.
With these changes sf(4) should work on all platforms.
o Don't set if_mtu in device attach, it's handled in ether_ifattach.
o Use our own callout to drive watchdog timer.
o Enable VLAN oversized frames and announce sf(4)'s VLAN capability
to upper layer.
o In sf_detach(), remove mtx_initialized KASSERT as it's not possible
to get there without initialzing the mutex. Also mark that we're
about to detaching so active bpf listeners do not panic the system.
o To reduce PCI register access cycles, Rx completion ring is
directly scanned instead of reading consumer/producer index
registers. In theory, Tx completion ring also can be directly
scanned. However the completion ring is composed of two types
completion(1 for Tx done and 1 and DMA done). So reading producer
index via register access would be more safer way to detect the
ring wrap-around.
o In sf_rxeof(), don't use m_devget(9) to align recevied frames. The
alignment is required only for strict-alignment architectures and
now the alignment is handled by sf_fixup_rx() if required. The
removal of the copy operation in fast path should increase Rx
performance a lot on non-strict-alignemnt architectures such as
i386 and amd64.
o In sf_newbuf(), don't set descriptor valid bit as sf(4) is
programmed to run with normal mode. In normal mode, the valid bit
have no meaning. The valid bit should be used only when the
hardware uses polling(prefetch) mode. The end of descriptor queue
bit could be used if needed, but sf(4) relys on auto-wrapping of
hardware on 256 descriptor queue entries so both valid and
descriptor end bit are not used anymore.
o Don't disable generation of Tx DMA completion as said in datasheet
and use the Tx DMA completion entry instead of relying on Tx done
completion entry. Also added additional Tx completion entry type
check in Tx completion handler.
o Don't blindly reset watchdog timer in sf_txeof(). sf(4) now unarm
the the watchdog only if there are no active Tx descriptors in Tx
queue.
o Don't manually update various counters in driver, instead, use
built-in MAC statistic registers to update them. The statistic
registers are updated in every second.
o Modified Tx underrun handlers to increase the threshold value
in units of 256 bytes. Previously it used to increase 16 bytes
at a time which seems to take too long to stabalize whenever Tx
underrun occurrs.
o In interrupt handler, additional check for the interrupt is
performed such that interrupts only for this device is allowed to
process descriptor rings. Because reading SF_ISR register clears
all interrtups, nuke writing to a SF_ISR register.
o Tx underrun is abonormal condition and SF_ISR_ABNORMALINTR includes
the interrupt. So there is no need to inspect the Tx underrun again
in main interrupt loop.
o Don't blindly reinitialize hardware for abnormal interrupt
condition. sf(4) reintializes the hardware only when it encounters
DMA error which requires an explicit hardware reinitialization.
o Fix a long standing bug that incorrectly clears MAC statistic
registers in sf_init_locked.
o Added strict-alignment safe way of ethernet address reprogramming
as IF_LLADDR may return unaligned address.
o Move sf_reset() to sf_init_locked in order to always reset the
hardware to a known state prior to configuring hardware.
o Set default Rx DMA, Tx DMA paramters as shown in datasheet.
o Enable PCI busmaster logic and autopadding for VLAN frames.
o Rework sf_encap.
- Previously sf(4) used to type 0 of Tx descriptor with padding
enabled to store driver private data. Emebedding private data
structures into descriptors is bad idea as the structure size
would be different between 64bit and 32bit architectures. The
type 0 descriptor allows fixed number of DMA segments in
a descriptor format and provides relatively simple interface to
manage multi-fragmented frames.
However, it wastes lots of Tx descriptors as not all frames are
fragmented as the number of allowable segments in a descriptor.
- To overcome the limitation of type 0 descriptor, switch to type
2 descriptor which allows 64bit DMA addressing and can handle
unliumited number of fragmented DMA segments. The drawback of
type 2 descriptor is in its complexity in managing descriptors
as driver should handle the end of Tx ring manually.
- Manually set Tx desciptor queue end mark and record number of
used descriptors to reclaim used descriptors in sf_txeof().
o Rework sf_start.
- Honor link up/down state before attempting transmission.
- Because sf(4) uses only one of two Tx queues, use low priority
queue instead of high one. This will remove one shift operation
in each Tx kick command.
- Cache last produder index into softc such that subsequenet Tx
operation doesn't need to access producer index register.
o Rewrote sf_stats_update to include all available MAC statistic
counters.
o Employ AIC-6915 firmware from Adaptec and implement firmware
download routine and TCP/UDP checksum offload.
Partial checksum offload support was commented out due to the
possibility of firmware bug in RxGFP.
The firmware can strip VLAN tag in Rx path but the lack of firmware
assistance of VLAN tag insertion in transmit side made it useless
on FreeBSD. Unlike checksum offload, FreeBSD requires both Tx/Rx
hardware VLAN assistance capability. The firmware may also detect
wakeup frame and can wake system up from states other than D0.
However, the lack of wakeup support form D3cold state keep me from
adding WOL capability. Also detecting WOL frame requires firmware
support but it's not yet known to me whether the firmware can
process the WOL frame.
o Changed *_ADDR_HIADDR to *_ADDR_HI to match other definitions of
registers.
o Added definitioan to interrupt moderation related constants.
o Redefined SF_INTRS to include Tx DMA done and DMA errors. Removed
Tx done as it's not needed anymore.
o Added definition for Rx/Tx DMA high priority threshold.
o Nuked unused marco SF_IDX_LO, SF_IDX_HI.
o Added complete MAC statistic register definition.
o Modified sf_stats structure to hold all MAC statistic regiters.
o Nuke various driver private padding data in Tx/Rx descriptor
definition. sf(4) no longer requires private padding. Also remove
unused padding related definitions. This greatly simplifies
descriptor manipulation on 64bit architectures.
o Becase we no longer pad driver private data into descriptor,
remove deprecated/not-applicable comments for padding.
o Redefine Rx/Tx desciptor status. sf(4) doesn't use bit fileds
anymore to support endianness.
Tested by: bruffer (initial version)
be wrong but I couldn't find a way to make it work. In addition, the
number of TxGFP instruction does not match the firmware image size,
so I guess something was wrong when Adaptec generated the TxGFP
firmware from their DDK.
According to datasheet, normally, the first GFP instruction would be
opcode C, WaitForStartOfFrame, to synchronize checksumming with
incoming frame. But the first instruction in TxGFP firmware was
opcode 1, BrToImmIfTrue, so it could not process checksum correctly,
I guess. Checking for RxGFP firmware also indicates the first
instruction should be opcode C. Since the number of instructions in
TxGFP firmware lacks exactly one instruction, I prepended the opcode
C to TxGFP firmware image. With this change, the resulting image size
perfectly matches with the nummber of instructions and Tx checksum
offload seems to work without problems.
- Allow IP in firewall_nat_interface, just like natd_interface
- Allow additional configuration parameters passed to ipfw via
firewall_nat_flags
- Document firewall_nat_* in defaults/rc.conf
Tested by: Albert B. Wang <abwang at gmail.com>
MFC after: 1 month
forget to translate "float" to "double".
ucbtest didn't detect the bug, but exhaustive testing of the float
case relative to the double case eventually did. The bug only affects
args x with |x| ~> 2**19*(pi/2) on non-i386 (i386 is broken in a
different way for large args).
it should never have existed and it has not been used for many years
(floats are reduced faster using doubles). All relevant changes (just
the workaround for broken assignment) have been merged to the double
version.
there is a problem with non-floats (when i386 defaults to extra
precision). This essentially restores yesterday's behaviour for doubles
on i386 (since generic rint() isn't used and everywhere else assumed
working assignment), but for arches that use the generic rint() it
finishes restoring some of 1995's behaviour (don't waste time doing
unnecessary store/load).
variable hack for exp2f() only.
The volatile variable had a surprisingly large cost for exp2f() -- 19
cycles or 15% on i386 in the worst case observed. This is only partly
explained by there being several references to the variable, only one
of which benefited from it being volatile. Arches that have working
assignment are likely to benefit even more from not having any volatile
variable.
exp2() now has a chance of working with extra precision on i386.
exp2() has even more references to the variable, so it would have been
pessimized more by simply declaring the variable as volatile. Even
the temporary volatile variable for STRICT_ASSIGN costs 5-10% on i386,
(A64) so I will change STRICT_ASSIGN() to do an ordinary assignment
until i386 defaults to extra precision.
instead of a volatile cast hack for the float version only. The cast
hack broke with gcc-4, but this was harmless since the float version
hasn't been used for a few years. Merge from the float version so
that the double version has a chance of working on i386 with extra
precision.
See k_rem_pio2f.c rev.1.8 for the original hack.
Convert to _FBSDID().
hack for log1pf() only. The cast hack broke with gcc-4, resulting in
~1 million errors of more than 1 ulp, with a maximum error of ~1.5 ulps.
Now the maximum error for log1pf() on i386 is 0.5034 ulps again (this
depends on extra precision), and log1p() has a chance of working with
extra precision.
See s_log1pf.c 1.8 for the original hack. (It claims only 62343 large
errors).
Convert to _FBSDID(). Another thing broken with gcc-4 is the static
const hack used for rcsids.
lockmgr lkp, when held in exclusive mode, is recursed
- Introduce the function BUF_RECURSED() which does the same for bufobj
locks based on the top of lockmgr_recursed()
- Introduce the function BUF_ISLOCKED() which works like the counterpart
VOP_ISLOCKED(9), showing the state of lockmgr linked with the bufobj
BUF_RECURSED() and BUF_ISLOCKED() entirely replace the usage of bogus
BUF_REFCNT() in a more explicative and SMP-compliant way.
This allows us to axe out BUF_REFCNT() and leaving the function
lockcount() totally unused in our stock kernel. Further commits will
axe lockcount() as well as part of lockmgr() cleanup.
KPI results, obviously, broken so further commits will update manpages
and freebsd version.
Tested by: kris (on UFS and NFS)
- Don't specify vnode operations for mknod, lease, and advlock--let them
fall through to vop_default.
- Implement vop_default with &default_vnodeops, rather than with VOP_PANIC,
so that unimplemented vnode operations are handled in more sensible ways
than panicking, such as EOPNOTSUPP on ACL queries generated by bsdtar,
or mknod.
MFC after: 3 days
fill out all fields, just fill out the ones the file system knows
about. Among other things, this causes the outpuf of "mount" and
"df" to make quite a bit more sense as /dev/cfs0 is specified as the
mountfrom name.
MFC after: 3 days
around assignments not working for gcc on i386. Now volatile hacks
for rint() and rintf() don't needlessly pessimize so many arches
and the remaining pessimizations (for arm and powerpc) can be avoided
centrally.
This cleans up after s_rint.c 1.3 and 1.13 and s_rintf.c 1.3 and 1.9:
- s_rint.c 1.13 broke 1.3 by only using a volatile cast hack in 1 place
when it was needed in 2 places, and the volatile cast hack stopped
working with gcc-4. These bugs only affected correctness tests on
i386 since i386 normally uses asm rint() and doesn't support the
extra precision mode that would break assignments of doubles.
- s_rintf.c 1.9 improved(?) on 1.3 by using a volatile variable hack
instead of an extra-precision variable hack, but it declared 2
variables as volatile when only 1 variable needed to be volatile.
This only affected speed tests on i386 since i386 uses asm rintf().
vnodes during coda_unmount() in order to detect errant use of them
after the vnode references may no longer be valid.
No need to clear the VV_ROOT flag on mi_rootvp flag (especially after
the vnode reference is no longer valid) as this isn't done on other
file systems.
MFC after: 3 days
and then release it when it is closed: we rely on the caller to keep the
vnode around with a valid reference. This avoids vrele() destroying the
vnode vop_close() is being called from during a call to vop_close(), and
a crash due to lockmgr recursing the vnode lock when a Coda unmount
occurs.
MFC after: 3 days
Move all extern variable definitions to associated .h files, move some
extern variable definitions between include files to place them more
appropriately.
MFC after: 3 days
Coda vnode derived from it, in the style of nullfs. This allows files
in the Coda file system to be memory-mapped, such as with execve(2) or
mmap(2).
MFC after: 3 days
Reported by: Rune <u+openafsdev-sr55 at chalmers dot se>
mostly just test corner cases rather than accuracy. Some of the
tests don't pass right now if you compile libm at -O2 due to gcc
constant-folding some things that it shouldn't. I'll fix that
shortly.
long doubles (i386, amd64, ia64) and one for machines with 128-bit
long doubles (sparc64). Other platforms use the double version.
I've only done runtime testing on i386.
Thanks to bde@ for helpful discussions and bugfixes.
