It seems RTL8169/RTL8168/RTL810xE has a kind of interrupt
moderation mechanism but it is not documented at all. The magic
value dramatically reduced number of interrupts without noticeable
performance drops so apply it to all RTL8169/RTL8169 controllers.
Vendor's FreeBSD driver also applies it to RTL810xE controllers but
their Linux driver explicitly cleared the register, so do not
enable interrupt moderation for RTL810xE controllers.
While I'm here sort 8169 specific registers.
Obtained from: RealTek FreeBSD driver
useful counters like rl_missed_pkts is 16 bits quantity which is
too small to hold meaningful information happened in a second. This
means driver should frequently read these counters in order not to
lose accuracy and that approach is too inefficient in driver's
view. Moreover it seems there is no way to trigger an interrupt to
detect counter near-full or wraparound event as well as lacking
clearing the MAC counters. Another limitation of reading the
counters from RealTek controllers is lack of interrupt firing at
the end of DMA cycle of MAC counter read request such that driver
have to poll the end of the DMA which is a time consuming process
as well as inefficient. The more severe issue of the MAC counter
read request is it takes too long to complete the DMA. All these
limitation made maintaining MAC counters in driver impractical. For
now, just provide simple sysctl interface to trigger reading the
MAC counters. These counters could be used to track down driver
issues. Users can read MAC counters maintained in controller with
the following command.
#sysctl dev.re.0.stats=1
While I'm here add check for validity of dma map and allocated
memory before unloading/freeing them.
Tested by: rmacklem
checksum offload frames. Software workaround used for broken
controllers(RTL8169, RTL8168, RTL8168B) seem to cause watchdog
timeouts on RTL8139C+.
Introduce a new flag RL_FLAG_AUTOPAD to mark automatic padding
feature of controller and set it for RTL8139C+ and controllers that
use new descriptor format. This fixes watchdog timeouts seen on
RTL8139C+.
Reported by: Dimitri Rodis < DimitriR <> integritasystems dot com >
Tested by: Dimitri Rodis < DimitriR <> integritasystems dot com >
command whenever Tx completion interrupt is raised. The Tx poll
bit is cleared when all packets waiting to be transferred have been
processed. This means the second Tx poll command can be silently
ignored as the Tx poll bit could be still active while processing
of previous Tx poll command is in progress.
To address the issue re(4) used to invoke the Tx poll command in Tx
completion handler whenever it detects there are pending packets in
TxQ. However that still does not seem to completely eliminate
watchdog timeouts seen on RealTek PCIe controllers. To fix the
issue kick Tx poll command only after Tx completion interrupt is
raised as this would indicate Tx is now idle state such that it can
accept new Tx poll command again. While here apply this workaround
for PCIe based controllers as other controllers does not seem to
have this limitation.
Tested by: Victor Balada Diaz < victor <> bsdes DOT net >
out of sleep mode prior to accessing to PHY. This should fix device
attach failure seen on these controllers. Also enable the sleep
mode when device is put into sleep state.
PR: kern/123123, kern/123053
established a valid link or not. In miibus_statchg handler add a
check for established link is valid one for the controller(e.g.
1000baseT is not a valid link for fastethernet controllers.)
o Added a flag RE_FLAG_FASTETHER to mark fastethernet controllers.
o Added additional check to know whether we've really encountered
watchdog timeouts or missed Tx completion interrupts. This change
may help to track down the cause of watchdog timeouts.
o In interrupt handler, removed a check for link state change
interrupt. Not all controllers have the bit and re(4) did not
rely on the event for a long time. In addition, re(4) didn't
request the interrupt in RL_IMR register.
Tested by: rpaulo
cable tuning. This has helped in some installations for hardware
deployed by a former employer. Made optional because the lists aren't
full of complaints about these cards... even when they were wildly
popular.
Reviewed by: attilio@, jhb@, trhodes@ (all an older version of the patch)
- The hardware does not support DAC so limit DMA address space to
4GB.
- Removed BUS_DMA_ALLOC_NOW flag.
- Created separated Tx buffer and Rx buffer DMA tags. Previously
it used to single DMA tag and it was not possible to specify
different DMA restrictions.
- Apply 4 bytes alignment limitation of Tx buffer.
- Apply 8 bytes alignment limitation of Rx buffer.
- Tx side bus_dmamap_load_mbuf_sg(9) support.
- Preallocate Tx DMA maps as creating DMA maps take very long time
on architectures that require real DMA maps.
- Adjust guard buffer size to 1522 + 8 as it should include VLAN
and additional reserved bytes in Rx buffer.
- Plug memory leak in device detach. Previously wrong buffer
address was used to free allocated memory.
- Added rl_list_rx_init() to clear Rx buffer and cleared the
buffer.
- Don't destroy DMA maps in rl_txeof() as the DMA map should be
reused. There is no reason to destroy/recreate the DMA maps in
this driver.
- Removed rl_dma_map_rxbuf()/rl_dma_map_txbuf() callbacks.
- The hardware does not support descriptor based DMA on Tx side
and the Tx buffer address should be aligned on 4 bytes boundary
as well as manual padding for short frames. Because of this
hardware limitation rl(4) always used to invoke m_defrag(9) to
get a 4 bytes aligned single buffer. However m_defrag(9) takes
a lot of CPU cycles on slow machines and not all packets need
the help of m_defrag(9). Armed with the information, don't
invoke m_defrag(9) if the following conditions are true.
1. Buffer is not fragmented.
2. Buffer is aligned on 4 bytes boundary.
3. Manual padding is not necessary.
4. Or padding is necessary but upper stack passed a writable
buffer and the space needed for padding is satisfied.
This change combined with preallocated DMA maps greatly
increased Tx performance of driver on sparc64.
- Moved bus_dmamap_sync(9) in rl_start_locked() to rl_encap() and
corrected memory synchronization operation specifier of
bus_dmamap_sync(9).
- Removed bus_dmamap_unload(9) in rl_stop(). There is no need to
reload/unload Rx buffer as rl(4) always have to copy from the
buffer. It just needs proper bus_dmamap_sync(9) calls before
copying the received frame.
With this change rl(4) should work on systems with more than 4GB
memory.
PR: kern/128143
11bits. This limits the maximum interface MTU size in TSO case
as upper stack should not generate TCP segments with MSS greater
than the limit. Armed with this information, disable TSO if
interface MTU is greater than the limit.
generation of RTL810x PCIe fast ethernet controller. Note, Tx/Rx
descriptor format is different from that of first generation of
RTL8101E series. Jumbo frame is not supported for RTL810x
family.
Tested by: NAGATA Shinya ( maya AT negeta DOT com )
ATM Tx/Rx checksum offload is supported but TSO and jumbo frame is
not yet supported. Because these newer controllers use different
descriptor formats, a flag RL_FLAG_DESCV2 flag was introduced to
handle that case in Tx/Rx handler. Also newer controllers seems to
require to not touch 'enable Tx/Rx bit' in RL_CPLUS_CMD register
so don't blindly try to set that bits.
Note, it seems that there is still power-saving related issue where
driver fails to attach PHY. Rebooting seems to fix that issue but
number of required reboots varys.
Many thanks to users that helped during developement. I really
appreciate their patient and test/feedbacks.
a dedicated flag that represents controller capabilities/events.
This will simplify many part of code that requires different
workaround for each controller revisions and will enhance
readability.
While I'm here move PHY wakeup code up before mii_phy_probe() which
seems to help to wake PHY in some cases.
o Increased number of Rx/Tx descriptors to 256 for 8169 GigEs
because it's hard to push the hardware to the limit with default
64 descriptors.
TSO requires large number of Tx descriptors to pass a full sized
TCP segment(65535 bytes IP packet) to hardware. Previously it
consumed 32 Tx descriptors, assuming MCLBYTES DMA segment size,
to send the TCP segment which means re(4) couldn't queue more
than two full sized IP packets.
For 8139C+ it still uses 64 Rx/Tx descriptors due to its hardware
limitations. With this changes there are (very) small waste of
memory for 8139C+ users but I don't think it would affect 8139C+
users for most cases.
o Various bus_dma(9) fixes.
- The hardware supports DAC so allow 64bit DMA operations.
- Removed BUS_DMA_ALLOC_NOW flag.
- Increased DMA segment size to 4096 from MCLBYTES because TSO
consumes too many descriptors with MCLBYTES DMA segment size.
- Tx/Rx side bus_dmamap_load_mbuf_sg(9) support. With these
changes the code is more readable than previous one and got a
(slightly) better performance as it doesn't need to pass/
decode arguments to/from callback function.
- Removed unnecessary callback function re_dmamap_desc() and
nuked rl_dmaload_arg structure which was used in the callback.
- Additional protection for DMA map load failure. In case of
failure reuse current map instead of returning a bogus DMA
map.
- Deferred DMA map unloading/sync operation for maximum
performance until we really need to load new DMA map. If we
happen to reuse current map(e.g. input error) there is no need
to sync/unload/load again.
- The number of allowable Tx DMA segments for a mbuf chains are
now 32 instead of magic nseg value. If the number of available
Tx descriptors are short enough to send highly fragmented mbuf
chains an optimized re_defrag() is called to collapse mbuf
chains which is supposed to be much faster than m_defrag(9).
re_defrag() was borrowed from ath(4).
- Separated Rx/Tx DMA tag from a common DMA tag such that Rx DMA
tag correctly uses DMA maps that were created with DMA alignment
restriction(8bytes alignments). Tx DMA tag does not have such
alignment limitation.
- Added additional sanity checks for DMA ring map load failure.
- Added additional spare Rx DMA map for graceful handling of Rx
DMA map load failure.
- Fixed misused bus_dmamap_sync(9) and added missing
bus_dmamap_sync(9) in re_encap()/re_txeof()/re_rxeof().
o Enabled TSO again as re(4) have reasonable number of Tx
descriptors.
o Don't touch DMA address of a Tx descriptor in re_txeof(). It's
not needed.
o Fix incorrect update of if_ierrors counter. For Rx buffer
shortage it should update if_qdrops as the buffer is reused.
o Added checks for unsupported H/W revisions and return ENXIO for
these hardwares. This is required to remove resource allocation
code in re_probe as other drivers do in device probe routine.
o Modified descriptor index manipulation macros as it's now possible
to have different number of descriptors for Rx/Tx.
o In re_start, to save a lock operation, use IFQ_DRV_IS_EMPTY before
trying to invoke IFQ_DRV_DEQUEUE. Also don't blindly call re_encap
since we already know the number of available Tx descriptors in
advance.
o Removed RL_TX_DESC_THLD which was used to reserve RL_TX_DESC_THLD
descriptors in Tx path. There is no such a limitation mentioned in
8139C+/8169/8110/8168/8101/8111 datasheet and it seems to work ok
without reserving RL_TX_DESC_THLD descriptors.
o Fix a comment for RL_GTXSTART. The register is 8bits register.
o Added comments for 8169/8139C+ hardware restrictions on descriptors.
o Removed forward declaration for "struct rl_softc", it's not needed.
o Added a new structure rl_txdesc for Tx descriptor managements and
a structure rl_rxdesc for Rx descriptor managements.
o Removed unused member variable rl_intlock in driver softc. There are
still several unused member variables which are supposed to be used
to access hardware statistics counters. But it seems that accessing
hardware counters were not implemented yet.
Ethernet Controller. Multicast filtering wasn't tested and needs more
expore. While I'm here change complex if statements with switch
statement which would improve readability.
Reported by: Abdullah Ibn Hamad Al-Marri < wearabnet AT yahoo DOT ca >
Tested by: Abdullah Ibn Hamad Al-Marri < wearabnet AT yahoo DOT ca >
Ever since switching to adaptive polling re(4) occasionally spews
watchdog timeouts on systems with MSI capability. This change is
minimal one for supporting MSI and re(4) also needs MSIX support
for RTL8111C in future. Because softc structure of re(4) is shared
with rl(4), rl(4) was touched to use the modified softc.
Reported by: cnst
Tested by: cnst
Approved by: re (kensmith)
would be 93C46(1Kbit) or 93C56(2Kbit). One of differences between them
is number of address lines required to access the EEPROM. For example,
93C56 EEPROM needs 8 address lines to read/write data. If 93C56
recevied premature end of required number of serial clock(CLK) to set
OP code/address of EEPROM, the result would be unexpected behavior.
Previously it tried to detect 93C46, which requires 6 address lines,
and then assumed it would be 93C56 if read data was not expected
value. However, this approach didn't work in some models/situations
as 93C56 requries 8 address lines to access its data. In order to fix
it, change EEPROM probing order such that 93C56 is detected reliably.
While I'm here change hard-coded address line numbers with defined
constant to enhance readability.
PR: 112710
Approved by: re (mux)
Previously whenever PROMISC mode turned on/off link renegotiation
occurs and it could resulted in network unavailability for serveral
seconds.(Depending on switch STP settings it could last several tens
seconds.)
Reported by: Prokofiev S.P. < proks AT logos DOT uptel DOT net >
Tested by: Prokofiev S.P. < proks AT logos DOT uptel DOT net >
operation as it ran out of free descriptors or if there are too many
segments in the first place, call bus_dmamap_unload() in order to
unload the already loaded segments.
For trying to map the defragmented mbuf (chain) in re_encap() this
introduces re_dma_map_desc() setting arg.rl_maxsegs to 0 as a new
failure mode. Previously we just ignored this case, corrupting our
view of the TX ring.
o In re_txeof():
- Don't clear IFF_DRV_OACTIVE unless there are at least 4 free TX
descriptors. Further down the road re_encap() will bail if there
aren't at least 4 free TX descriptors, causing re_start() to
abort and prepend the dequeued mbuf again so it makes no sense
to pretend we could process mbufs again when in fact we won't.
While at it replace this magic 4 with a macro RL_TX_DESC_THLD
throughout this driver.
- Don't cancel the watchdog timeout as soon as there's at least one
free TX descriptor but instead only if all descriptors have been
handled. It's perfectly normal, especially in the DEVICE_POLLING
case, that re_txeof() is called when only a part of the enqueued
TX descriptors have been handled, causing the watchdog to be
disarmed prematurely.
o In re_encap():
- If m_defrag() fails just drop the packet like other NIC drivers
do. This should only happen when there's a mbuf shortage, in which
case it was possible to end up with an IFQ full of packets which
couldn't be processed as they couldn't be defragmented as they
were taking up all the mbufs themselves. This includes adjusting
re_start() to not trying to prepend the mbuf (chain) if re_encap()
has freed it.
- Remove dupe initialization of members of struct rl_dmaload_arg to
values that didn't change since trying to process the fragmented
mbuf chain.
While at it remove an unused member from struct rl_dmaload_arg.
o In re_start() remove a abandoned, banal comment. The corresponding
code was moved to re_attach() some time ago.
With these changes re(4) now survives one day (until stopped) of
hammering out packets here.
Reviewed by: yongari
MFC after: 2 weeks
- Change the workaround for the autopad/checksum offload bug so that
instead of lying about the map size, we actually create a properly
padded mbuf and map it as usual. The other trick works, but is ugly.
This approach also gives us a chance to zero the pad space to avoid
possibly leaking data.
- With the PCIe devices, it looks issuing a TX command while there's
already a transmission in progress doesn't have any effect. In other
words, if you send two packets in rapid succession, the second one may
end up sitting in the TX DMA ring until another transmit command is
issued later in the future. Basically, if re_txeof() sees that there
are still descriptors outstanding, it needs to manually resume the
TX DMA channel by issuing another TX command to make sure all
transmissions are flushed out. (The PCI devices seem to keep the
TX channel moving until all descriptors have been consumed. I'm not
sure why the PCIe devices behave differently.)
(You can see this issue if you do the following test: plug an re(4)
interface into another host via crossover cable, and from the other
host do 'ping -c 2 <host with re(4) NIC>' to prime the ARP cache,
then do 'ping -c 1 -s 1473 <host with re(4) NIC>'. You're supposed
to see two packets sent in response, but you may only see one. If
you do 'ping -c 1 -s 1473 <host with re(4) NIC>' again, you'll
see two packets, but one will be the missing fragment from the last
ping, followed by one of the fragments from this ping.)
- Add the PCI ID for the US Robotics 997902 NIC, which is based on
the RTL8169S.
- Add a tsleep() of 1 second in re_detach() after the interrupt handler
is disconnected. This should allow any tasks queued up by the ISR
to drain. Now, I know you're supposed to use taskqueue_drain() for
this, but something about the way taskqueue_drain() works with
taskqueue_fast queues doesn't seem quite right, and I refuse to be
tricked into fixing it.
- Correct the PCI ID for the 8169SC/8110SC in the device list (I added
the macro for it to if_rlreg.h before, but forgot to use it.)
- Remove the extra interrupt spinlock I added previously. After giving it
some more thought, it's not really needed.
- Work around a hardware bug in some versions of the 8169. When sending
very small IP datagrams with checksum offload enabled, a conflict can
occur between the TX autopadding feature and the hardware checksumming
that can corrupt the outbound packet. This is the reason that checksum
offload sometimes breaks NFS: if you're using NFS over UDP, and you're
very unlucky, you might find yourself doing a fragmented NFS write where
the last fragment is smaller than the minimum ethernet frame size (60
bytes). (It's rare, but if you keep NFS running long enough it'll
happen.) If checksum offload is enabled, the chip will have to both
autopad the fragment and calculate its checksum header. This confuses
some revs of the 8169, causing the packet that appears on the wire
to be corrupted. (The IP addresses and the checksum field are mangled.)
This will cause the NFS write to fail. Unfortunately, when NFS retries,
it sends the same write request over and over again, and it keeps
failing, so NFS stays wedged.
(A simple way to provoke the failure is to connect the failing system
to a network with a known good machine and do "ping -s 1473 <badhost>"
from the good system. The ping will fail.)
Someone had previously worked around this using the heavy-handed
approahch of just disabling checksum offload. The correct fix is to
manually pad short frames where the TCP/IP stack has requested
checksum offloading. This allows us to have checksum offload turned
on by default but still let NFS work right.
- Not a bug, but change the ID strings for devices with hardware rev
0x30000000 and 0x38000000 to both be 8168B/8111B. According to RealTek,
they're both the same device, but 0x30000000 is an earlier silicon spin.
cards: the chips are all marked "RTL8111B", but they put stickers on the
back that say "RTL8168B/8111B". The manual says there's only one HWREV code
for both the 8111B and 8168B devices, which is 0x30000000, but the cards
they sent me actually report HWREV of 0x38000000. Deciding to trust the
hardware in front of me rather than a possibly incorrect manual (it wouldn't
be the first time the HWREVs were incorrectly documented), I changed the
8168 revision code. It turns out this was a mistake though: 0x30000000
really is a valid for the 8168.
There are two possible reasons for there to be two different HWREVs:
1) 0x30000000 is used only for the 8168B and 0x38000000 is only for
the 8111B.
2) There were 8111/8168 rev A devices which both used code 0x30000000,
and the 8111B/8168B both use 0x38000000.
The product list on the RealTek website doesn't mention the existence of
any 8168/8111 rev A chips being in production though, and I've never seen
one, so until I get clarification from RealTek, I'm going to assume that
0x30000000 is just for the 8168B and 0x38000000 is for the 8111B only.
So, the HWREV code for the 8168 has been put back to 0x30000000,
a new 8111 HWREV code has been added, and there are now separate
entries for recognizing both devices in the device list. This will
allow all devices to work, though if it turns out I'm wrong I may
need to change the ID strings
latter is a PCIe 10/100 chip.
Finally fix the EEPROM reading code so that we can access the EEPROMs on all
devices. In order to access the EEPROM, we must select 'EEPROM programming'
mode, and then set the EEPROM chip select bit. Previously, we were setting
both bits simultaneously, which doesn't work: they must be set in the
right sequence.
Always obtain the station address from the EEPROM, now that EEPROM
reading works correctly.
Make the TX interrupt moderation code based on the internal timer
optional and turned off by default.
Make the re_diag() routine conditional and off by default. When it is
on, only use it for the original 8169, which was the only device that
that really needed it.
Modify interrupt handling to use a fast interrupt handler and fast
taskqeueue.
Correct the rgephy driver so that it only applies the DSP fixup for
PHY revs 0 and 1. Later chips are fixed and don't need the fixup.
Make the rgephy driver advertise both 1000_FD and 1000_HD bits in
autoneg mode. A couple of the devices don't autoneg correctly unless
configured this way.
