- Add support for IPv6 rx csum offload
- Finally switch mxge from using its own driver lro, to
using tcp_lro
MFC after: 7 days
Sponsored by: Myricom Inc.
tree used it incorrectly, which lead to inaccurate overrated
if_obytes accounting. The drbr(9) used to update ifnet stats on
drbr_enqueue(), which is not accurate since enqueuing doesn't
imply successful processing by driver. Dequeuing neither mean
that. Most drivers also called drbr_stats_update() which did
accounting again, leading to doubled if_obytes statistics. And
in case of severe transmitting, when a packet could be several
times enqueued and dequeued it could have been accounted several
times.
o Thus, make drbr(9) API thinner. Now drbr(9) merely chooses between
ALTQ queueing or buf_ring(9) queueing.
- It doesn't touch the buf_ring stats any more.
- It doesn't touch ifnet stats anymore.
- drbr_stats_update() no longer exists.
o buf_ring(9) handles its stats itself:
- It handles br_drops itself.
- br_prod_bytes stats are dropped. Rationale: no one ever
reads them but update of a common counter on every packet
negatively affects performance due to excessive cache
invalidation.
- buf_ring_enqueue_bytes() reduced to buf_ring_enqueue(), since
we no longer account bytes.
o Drivers handle their stats theirselves: if_obytes, if_omcasts.
o mlx4(4), igb(4), em(4), vxge(4), oce(4) and ixv(4) no longer
use drbr_stats_update(), and update ifnet stats theirselves.
o bxe(4) was the most correct driver, it didn't call
drbr_stats_update(), thus it was the only driver accurate under
moderate load. Now it also maintains stats itself.
o ixgbe(4) had already taken stats from hardware, so just
- drop software stats updating.
- take multicast packet count from hardware as well.
o mxge(4) just no longer needs NO_SLOW_STATS define.
o cxgb(4), cxgbe(4) need no change, since they obtain stats
from hardware.
Reviewed by: jfv, gnn
It seems strchr() and strrchr() are used more often than index() and
rindex(). Therefore, simply migrate all kernel code to use it.
For the XFS code, remove an empty line to make the code identical to
the code in the Linux kernel.
This fixes a long standing bug in mxge(4) where "ifconfig mxge0 $IP"
did not bring the interface into a RUNNING state, like it does on
most (all?) other FreeBSD NIC drivers.
Thanks to gnn for mentioning the bug, and yongari for pointing out that
ether_ioctl() invokes ifp->if_init() in SIOCSIFADDR.
MFC after: 7 days
The Myri10GE NIC will assume all TSO frames contain partial checksum,
and will emit TSO segments with bad TCP checksums if a TSO frame
contains a full checksum. The mxge driver takes care to make sure
that TSO is disabled when checksum offload is disabled for this
reason. However, modules that modify packet contents (like pf) may
end up completing a checksum on a TSO frame, leading to the NIC emitting
TSO segments with bad checksums.
To workaround this, restore the partial checksum in the mxge driver
when we're fed a TSO frame with a full checksum.
Reported by: Bob Healey
MFC after: 3 days
- Re-probe xfp / sfp+ socket on link events, in case user
has changed transceiver
- correctly report current media to avoid confusing lagg (reported by Panasas)
- Report link speed (submitted by yongari)
Reviewed by: yongari (earlier version)
MFC after: 7 days
- Don't leak slice resources when mxge_alloc_rings() fails
- Start taskq threads only after we know attach will succeed. At
boot time, taskqueue_terminate() will loop infinately, waiting
for the threads to exit, and hang the system.
Submitted by: Panasas
MFC After: 3 days
- introduce drbr_needs_enqueue that returns whether the interface/br needs
an enqueue operation: returns true if altq is enabled or there are
already packets in the ring (as we need to maintain packet order)
- update all drbr consumers
- fix drbr_flush
- avoid using the driver queue (IFQ_DRV_*) in the altq case as the
multiqueue consumer does not provide enough protection, serialize altq
interaction with the main queue lock
- make drbr_dequeue_cond work with altq
Discussed with: kmacy, yongari, jfv
MFC after: 4 weeks
Pertinant highlights from Myricom CHANGES file include:
- Make sure invalid external smbus activity cannot affect performance
- Fix to avoid a bug where the link could sometimes stay reported as
up on after unplugging the cable.
- For 8B NIC, make smbus connection passive at init to avoid
possible address conflicts
- Increase number of slices to 17 for multi-slice fw
- Fix a bug where packets dropped because of link_overflow could
be occasionally reported as bad_crc32
- Add selectable failover strategy for dual-port chip: symmetric or primary/backup
- On failover, send RARP broadcast to make the change immediately
known to the network
- Change endianess for PCI Device Serial Number
- For dual-port NICs, time to failover is now a few microsecs
instead of a few millisecs.
MFC after: 3 days
by checking PCI config space when the NIC is not
transmitting. Previously, a h/w fault would not have been
detected if the NIC was down, or handling an RX only
workload.
1) Restore the PCI Express control register after a watchdog
reset. This is required because the device will come out
of watchdog reset with the pectl reg at its default state,
and important BIOS configuration (like max payload size)
could be lost.
2) Call mxge_start_locked() for every tx queue before dropping
the lock in the watchdog handler. This is required, as
the queue's buf ring may have filled during the reset.
- Mark the link as down, so if watchdog reset fails, link watching
failover software can notice it
- Don't send MXGEFW_CMD_ETHERNET_DOWN if the NIC has been reset, it is
not needed, and will fail on a freshly reset NIC.
- Ensure the transmit routines aren't attempting to PIO write to doorbells
while the NIC is being reset.
- Download the correct f/w, rather than using the EEPROM f/w after reset.
- Export a count of the number of watchdog resets via sysctl
- Zero all f/w stats at reset. This will lead to less confusing
diagnostic output when investigating NIC failures.
MFC after: 3 days
loader, because it uses a reserved suffix (_type). Fix
this by removing the "_" and renaming the tunable to
hw.mxge.rss_hashtype. The old (rss_hash_type) tunable is
still fetched, in case people load the driver via scripts.
When both are present in the kernel environment,
the new value (hw.mxge.rss_hashtype) overrides the old
value.
Approved by: re (kib)
IF_ADDR_UNLOCK() across network device drivers when accessing the
per-interface multicast address list, if_multiaddrs. This will
allow us to change the locking strategy without affecting our driver
programming interface or binary interface.
For two wireless drivers, remove unnecessary locking, since they
don't actually access the multicast address list.
Approved by: re (kib)
MFC after: 6 weeks
I tried re-ordering ether_ifdetach(), but this created a new race
where sometimes, when under heavy receive load (>1Mpps) and running
tcpdump, the machine would panic. At panic, the ithread was still in
the original (not dead) if_input() path, and was accessing stale BPF
data structs. By using a dying flag, I can close the interface prior
to if_detach() to be certain the interface cannot send packets up in
the middle of ether_ifdetach.
- Support for 10G-PCIE*-8B*-C (dual-port CX4) NICs
- For dual-port NICs, f/w failover is now a few microsecs
instead of a few millisecs.
- On failover, f/w sends RARP broadcast to make the change
immediately known to the network
- Fixed a bug observed on IBM X3 architecture where
some spurious ecrc errors would be reported when OS enabled
ecrc support.
Sponsored by: Myricom Inc.
- always maintain byte/mcast/drop stats via drbr
- move #define of IFNET_BUF_RING so that its picked
up by all files in the driver
- conditionalize IFNET_BUF_RING on the FreeBSD_version
bump just after it appeared in the tree.
Sponsored by: Myricom Inc.
using bus_dmamap_load_mbuf_sg() on it. This
prevents data corruption when the mxge MTU is
between 4076 and 8172 on machines with 4KB
pages and MXGE_VIRT_JUMBOS is in use (which it
isn't, in -current or -stable)
- Update mxge to use if_transmit(), and the new buf_ring
interfaces, so as to enable multiple transmit queues.
Use of if_transmit() is conditional on IFNET_BUF_RING,
and is enabled by default (as in if_em).
- Record a flow id on receive if receive hashing is active.
I currently only record the rx ring id (0..8) rather than
the 32-bit topelitz hash result, as doing the latter would
require shifting the driver to use a larger rx return ring.
Sponsored by: Myricom, Inc.
- Update to firmware 1.4.39 for dual-chip NIC (10G-PCIE2-xxx)
support, and SFP+ i2c support
- Identify newer "B" NICs (10G-PCIEx-8B-x) correctly, rather than
mis-identifying them as "A" NICs (cosmetic only)
- Identify the IFM_10G_LRM ifmedia type, where applicable.
- Identify ifmedia types for SFP+ based NICs
- Update copyright
Sponsored by: Myricom
MFC after: 1 week
and ifnet functions
- add memory barriers to <machine/atomic.h>
- update drivers to only conditionally define their own
- add lockless producer / consumer ring buffer
- remove ring buffer implementation from cxgb and update its callers
- add if_transmit(struct ifnet *ifp, struct mbuf *m) to ifnet to
allow drivers to efficiently manage multiple hardware queues
(i.e. not serialize all packets through one ifq)
- expose if_qflush to allow drivers to flush any driver managed queues
This work was supported by Bitgravity Inc. and Chelsio Inc.
This update fixes a transmit bug in the multi-queue (MSI-X) firmware
which happens when RDMAs complete out of order, and provides
improved support for the new Myri10GE NIC models (10G-PCIE-8Bx)
Sponsored by: Myricom Inc.
MFC after:3 days
- Support for Myricom 10G-PCIE-8B NICs
- multi-slice firmware: fix a bug when the presence of 32-bit or
64-bit DMA addresses for interrupt queues and data is not uniform across
slices.
- Improves automatic selection between ethp_z8e/eth_z8e
Sponsored by: Myricom Inc.
and handle NIC hardware watchdog resets.
- remove buggy code at the top of mxge_tick() which tried
to detect a race which is already detected in the kernel's
callout code.
- move callout_stop() and callout_reset() into mxge_close()
mxge_open() rather than doing the callout manipulation
all over the place.
- use callout_drain(), rather than callout_stop() to prevent
a potential race between mxge_tick() and mxge_detach()
which could lead to softclock using a destroyed mutex
- restructure the mxge_tick() and mxge_watchdog_reset()
routines to avoid resetting a callout, and then
immediately stopping it if the watchdog reset routine
is called, and fails.
- enable the driver to handle NIC hardware watchdog
resets by restoring the NIC's PCI config space, which is
lost when the NIC hardware watchdog triggers.
Reviewed by: jhb (previus version)
legacy interrupts rather than MSI as a special case. Prior to this
commit, the interrupt handler was doing the slow handshaking with
the device to ensure the legacy interrupt was lowered in both
the legacy and MSI-X case. This handshaking was not
required for MSI-X.
queues (which we call slices). The NIC will steer traffic into up to
hw.mxge.max_slices different receive rings based on a configurable
hash type (hw.mxge.rss_hash_type).
Currently the driver defaults to using a single slice, so the default
behavior is unchanged. Also, transmit from non-zero slices is
disabled currently.
- Fix a bug introduced in 1.4.20 where speculative read by the processor in the
write-only doorbell region would cause a target-abort (as opposed to simply
returning random data). This could manifest itself as NMI or machine freeze
depending on how the BIOS/OS/chipset configuration handles target-abort.
- Add support for new revisions of -R cards (with AEL1002/AEL1010 xaui->xfi)
- Increase an internal timing (dispatch engine): fix possible spurious reset
(seen on very few cards).
on an down mxge interface
- Fix a bug where mxge reported the link state as
active when it wasn't (after ifconfig down).
- Prevent spurious watchdog resets when link partner is not consuming
- Add support for CX4 and popular XFP media detection
- Update the firmware and associated header files to 1.4.25
Approved by: re (kensmith)
- Add custom .c wrappers for the firmware, rather than the standard
firmware(9) generated firmware objects to work around toolchain
problems on ia64 involving linking objects produced by
ld -b -binary into the kernel.
- Move from using Myricom's ".dat" firmware blobs to using Myricom's
zlib compressed ".h" firmware header files. This is done to
facilitate the custom wrappers, and saves a fair amount of wired
memory in the case where the firmware is built in, or preloaded.
- Fix two compile issues in mxge which only appear on non-i386/amd64.
Reviewed by: mlaier, mav (earlier version with just zlib support)
Glanced at by: sam
Approved by: re (kensmith)
- Allow LRO to be enabled / disabled at runtime
- Fix a double-free at module unload time.
- Only update timestamp in lro merge when it is present in the frame
Sponsored by: Myricom
clusters. This helps quite a bit on my low end machines (improves
performance by about 300Kpps when being blasted by a hardware
packet generator).
- Include one extended f/w counter forgotten in earlier commit
Sponsored by: Myricom Inc.
- Remove code to use the special wc_fifo. It has been disabled by default
in our other drivers as it actually slows down transmit by a small amount
- Dynamically determine the amount of space required for the rx_done
ring rather than hardcoding it.
- Compute the number of tx descriptors we are willing to transmit per
frame as the minimum of 128 or 1/4 the tx ring size.
- Fix a typo in the tx dma tag setup which could lead to unnecessary
defragging of TSO packets (and potentially even dropping TSO packets
due to EFBIG being returned).
- Add a counter to keep track of how many times we've needed to
defragment a frame. It should always be zero.
- Export new extended f/w counters via sysctl
Sponsored by: Myricom, Inc.
- Update to the latest (1.4.18) f/w. This f/w introduces a new
receive mode which allows us to use FreeBSD's physically discontinuous
MJUM9BYTES clusters.
- Switch the driver from chaining MJUMPAGESIZE clusters to using
MJUM9BYTES clusters to avoid mbuf chaining overheads. Due to this
change, people running obsolete f/w images will be limited to an MTU of
PAGE_SIZE - 16.
- Add (disabled by default) support for Large Receive Offload.
Sponsored by: Myricom, Inc.
- Update to latest (1.4.17) firmware.
- Use the new MXGEFW_CMD_UNALIGNED_TEST (added in firmare 1.4.16) to
have the firmware tell us if the PCIe chipset supports aligned PCIe
completions.
- Hard to maintain, and frequently out of date whitelist of PCIe
chipsets known to produce aligned completions removed, as it has been
replaced in its role of selecting the correct firmware to run by the
use of MXGEFW_CMD_UNALIGNED_TEST.
- Break the dma test out of mxge_reset() and into its own function
(mxge_dma_test()) so it can be used by both the normal DMA test, and
to run the unaligned test.
- Improved support for enabling ECRCs
Sponsored by: Myricom Inc.
were never freed, but the big ring was freed twice.
-Don't supply rx hw csums for frames which are padded beyond the
length specified in the ip header. If the padding is non-zero,
the hw csum will be incorrect for such frames.
Sponsored by: Myricom
been defragged and had their headers in the same cluster as their
payload would be fed to the NIC in header-sized chunks, and would
likely exceed the number of available transmit descriptors.
- If a TSO frame exceeds the number of available transmit descriptors,
don't leak busdmma resources when freeing it.
Sponsored by: Myricom Inc.
obtaining and releasing shared and exclusive locks. The algorithms for
manipulating the lock cookie are very similar to that rwlocks. This patch
also adds support for exclusive locks using the same algorithm as mutexes.
A new sx_init_flags() function has been added so that optional flags can be
specified to alter a given locks behavior. The flags include SX_DUPOK,
SX_NOWITNESS, SX_NOPROFILE, and SX_QUITE which are all identical in nature
to the similar flags for mutexes.
Adaptive spinning on select locks may be enabled by enabling the
ADAPTIVE_SX kernel option. Only locks initialized with the SX_ADAPTIVESPIN
flag via sx_init_flags() will adaptively spin.
The common cases for sx_slock(), sx_sunlock(), sx_xlock(), and sx_xunlock()
are now performed inline in non-debug kernels. As a result, <sys/sx.h> now
requires <sys/lock.h> to be included prior to <sys/sx.h>.
The new kernel option SX_NOINLINE can be used to disable the aforementioned
inlining in non-debug kernels.
The size of struct sx has changed, so the kernel ABI is probably greatly
disturbed.
MFC after: 1 month
Submitted by: attilio
Tested by: kris, pjd
When submitting rx buffers and not using WC fifo, always replace the
invalid DMA address with the real one, otherwise allocation failures
could lead to the invalid DMA address being given to the NIC, and
that would cause the receive side to lockup.
treated as multicast frames and filtered, but when only when "adopting"
running firmware. By "adopting", I mean using pre-existing firmware
loaded from eeprom at PCI reset, rather than firmware loaded by the
driver.
a version that i posted earlier on the -current mailing list,
and subsequent feedback received.
The core of the change is just in sys/firmware.h and kern/subr_firmware.c,
while other files are just adaptation of the clients to the ABI change
(const-ification of some parameters and hiding of internal info,
so this is fully compatible at the binary level).
In detail:
- reduce the amount of information exported to clients in struct firmware,
and constify the pointer;
- internally, document and simplify the implementation of the various
functions, and make sure error conditions are dealt with properly.
The diffs are large, but the code is really straightforward now (i hope).
Note also that there is a subtle issue with the implementation of
firmware_register(): currently, as in the previous version, we just
store a reference to the 'imagename' argument, but we should rather
copy it because there is no guarantee that this is a static string.
I realised this while testing this code, but i prefer to fix it in
a later commit -- there is no regression with respect to the past.
Note, too, that the version in RELENG_6 has various bugs including
missing locks around the module release calls, mishandling of modules
loaded by /boot/loader, and so on, so an MFC is absolutely necessary
there. I was just postponing it until this cleanup to avoid doing
things twice.
MFC after: 1 week
for pci_cfg_restore() to be exported. It was tested using a
hackily accessed pci_cfg_restore().
- Add ifmedia_removeall() to mxge_detach() in order to stop leaking
an ifaddr
- Fix a small acounting bug introduced by the locking code shuffle
which could cause spurious watchdog resets now that we have a
watchdog.
Sponsored by: Myricom
locking in preparation for adding a watchdog handler (callouts must
not use sleepable locks). This required shuffling memory and
interrupt allocation to the attach routine rather than if_ioctl so as
to avoid potential sleeps while bringing up the interface.
- initialize ifq_drv_maxlen correctly
- mark the interface as jumbo capable
- keep stats on the number of times the hw transmit queue filled and
was restarted.