of the newer drivers. The basic problem was
that the driver was pulling the mbuf off the
drbr ring and then when sending with xmit(), encounting
a full transmit ring. Thus the lower layer
xmit() function would return an error, and the
drivers would then append the data back on to the ring.
For TCP this is a horrible scenario sure to bring
on a fast-retransmit.
The fix is to use drbr_peek() to pull the data pointer
but not remove it from the ring. If it fails then
we either call the new drbr_putback or drbr_advance
method. Advance moves it forward (we do this sometimes
when the xmit() function frees the mbuf). When
we succeed we always call advance. The
putback will always copy the mbuf back to the top
of the ring. Note that the putback *cannot* be used
with a drbr_dequeue() only with drbr_peek(). We most
of the time, in putback, would not need to copy it
back since most likey the mbuf is still the same, but
sometimes xmit() functions will change the mbuf via
a pullup or other call. So the optimial case for
the single consumer is to always copy it back. If
we ever do a multiple_consumer (for lagg?) we
will need a test and atomic in the put back possibly
a seperate putback_mc() in the ring buf.
Reviewed by: jhb@freebsd.org, jlv@freebsd.org
This prevents quad igb card on high core machines, without any nmbcluster or
igb queue tuning wedging the boot process if all nics are configured.
Reviewed by: jfv
Approved by: pjd (mentor)
MFC after: 1 week
fail or not. The mbuf pointer is no longer valid, so
can't be reused after.
Fix igb_mq_start() where mbuf pointer was used after
drbr_enqueue().
This eventually leads us to all invocations of
igb_mq_start_locked() called with third argument as NULL.
This allows us to simplify this function.
Submitted by: Karim Fodil-Lemelin <fodillemlinkarim gmail.com>
Reviewed by: jfv
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
adapter->dropped_pkts instead of if_ierrors because if_ierrors is
overwritten by hw stats collection.
Submitted by: Andrew Boyer <aboyer@averesystems.com>
Reviewed by: Jack F Vogel <jfv@freebsd.org>
MFC after: 2 weeks
packet delivery, always enqueue when possible. Also
correct the DEPLETED test as multiple bits might be
set. Thanks to Randall Stewart for the changes!
system with sparse CPU IDs, you can have a valid CPU ID > mp_ncpus (e.g. if
you have two CPUs 0 and 4, with mp_maxid == 4 and mp_ncpus == 2).
Introduced at svn r235210
Submitted by: jhb@
Reviewed by: jfv@
setting in the igb and em driver. This was necessitated by
a shared code change that I was given late in the game, a data
type changed from bool to int, in the last update I dealt with
it by a cast, but it was pointed out (thanks jhb) that there
was a potential problem with this. John suggested this safer
approach, and it is fine with me...
MFC after:2 days (to catch the 9.1 update)
shared code update and small changes in core required
Add support for new i210/i211 devices
Improve queue calculation based on mac type
MFC after:5 days
as possible when using more than one igb(4) adapter. This
means that queues will not be bound to the same CPUs if
there are more CPUs availble.
This is only applicable to a system that has multiple interfaces.
Obtained from: Yahoo! Inc.
MFC after: 3 days
portions were already reapplied in r233708:
- Use a dedicated task to handle deferred transmits from the if_transmit
method instead of reusing the existing per-queue interrupt task.
Reusing the per-queue interrupt task could result in both an interrupt
thread and the taskqueue thread trying to handle received packets on a
single queue resulting in out-of-order packet processing.
- Call ether_ifdetach() earlier in igb_detach().
- Drain tasks and free taskqueues during igb_detach().
MFC after: 1 week
- Do not define the foo_start() methods or set if_start in the ifnet if
multiq transmit is enabled. Also, set if_transmit and if_qflush before
ether_ifattach rather than after when multiq transmit is enabled. This
helps to ensure that the drivers never try to mix different transmit
methods.
- Properly restart transmit during resume. igb(4) was not restarting it
at all, and em(4) was restarting even if the link was down and was
calling the wrong method if multiq transmit was enabled.
- Remove all the 'more' handling for transmit completions. Transmit
completion processing does not have a processing limit, so it always
runs to completion and never has more work to do when it returns.
Instead, the previous code was returning 'true' anytime there were
packets in the queue that weren't still in the process of being
transmitted. The effect was that the driver would continuously
reschedule a task to process TX completions in effect running at 100%
CPU polling the hardware until it finished transmitting all of the
packets in the ring. Now it will just wait for the next TX completion
interrupt.
- Restart packet transmission when the link becomes active.
- Fix the MSI-X queue interrupt handlers to restart packet transmission if
there are pending packets in the relevant software queue (IFQ or buf_ring)
after processing TX completions. This is the root cause for the OACTIVE
hangs as if the MSI-X queue handler drained all the pending packets from
the TX ring, nothing would ever restart it. As such, remove some
previously-added workarounds to reschedule a task to poll the TX ring
anytime OACTIVE was set.
Tested by: sbruno
Reviewed by: jfv
MFC after: 1 week
USERSPACE:
1. add support for devices with different number of rx and tx queues;
2. add better support for zero-copy operation, adding an extra field
to the netmap ring to indicate how many buffers we have already processed
but not yet released (with help from Eddie Kohler);
3. The two changes above unfortunately require an API change, so while
at it add a version field and some spares to the ioctl() argument
to help detect mismatches.
4. update the manual page for the two changes above;
5. update sample applications in tools/tools/netmap
KERNEL:
1. simplify the internal structures moving the global wait queues
to the 'struct netmap_adapter';
2. simplify the functions that map kring<->nic ring indexes
3. normalize device-specific code, helps mainteinance;
4. start exploring the impact of micro-optimizations (prefetch etc.)
in the ixgbe driver.
Use 'legacy' descriptors on the tx ring and prefetch slots gives
about 20% speedup at 900 MHz. Another 7-10% would come from removing
the explict calls to bus_dmamap* in the core (they are effectively
NOPs in this case, but it takes expensive load of the per-buffer
dma maps to figure out that they are all NULL.
Rx performance not investigated.
I am postponing the MFC so i can import a few more improvements
before merging.
Introduce some functions to map NIC ring indexes into netmap ring
indexes and vice versa. This way we can implement the bound
checks only in one place (and hopefully in a correct way).
On passing, make the code and comments more uniform across the
various drivers.
the memory allocator used by netmap. No functional change,
two small bug fixes:
- in if_re.c add a missing bus_dmamap_sync()
- in netmap.c comment out a spurious free() in an error handling block
import:
- Add read-only sysctls for all of the tunables supported by the igb and
em drivers.
- Make the per-instance 'enable_aim' sysctl truly per-instance by having it
change a per-instance variable (which is used to control AIM) rather
than having all of the per-instance sysctls operate on a single global
variable.
While here, restore the previously existing hw.igb.rx_processing_limit
tunable as it is very useful to be able to set a default tunable that
applies to all adapters in the system.
- Two struct members eee_disable are used in a function that expects
an int *, so declare them int, not bool.
- igb_tx_ctx_setup() returns a boolean value, so declare it bool, not int.
- igb_header_split is passed to TUNABLE_INT, so delcare it int, not bool.
- igb_tso_setup() returns a bool, so declare it bool, not boolean_t.
- Do not re-define bool/true/false if the symbols already exist.
MFC after: 2 weeks
Sponsored by: Isilon Systems, LLC
There have still been intermittent problems with apparent TX
hangs for some customers. These have been problematic to reproduce
but I believe these changes will address them. Testing on a number
of fronts have been positive.
EM: there is an important 'chicken bit' fix for 82574 in the shared
code this is supported in the core here.
- The TX path has been tightened up to improve performance. In
particular UDP with jumbo frames was having problems, and the
changes here have improved that.
- OACTIVE has been used more carefully on the theory that some
hangs may be due to a problem in this interaction
- Problems with the RX init code, the "lazy" allocation and
ring initialization has been found to cause problems in some
newer client systems, and as it really is not that big a win
(its not in a hot path) it seems best to remove it.
- HWTSO was broken when VLAN HWTAGGING or HWFILTER is used, I
found this was due to an error in setting up the descriptors
in em_xmit.
IGB:
- TX is also improved here. With multiqueue I realized its very
important to handle OACTIVE only under the CORE lock so there
are no races between the queues.
- Flow Control handling was broken in a couple ways, I have changed
and I hope improved that in this delta.
- UDP also had a problem in the TX path here, it was change to
improve that.
- On some hardware, with the driver static, a weird stray interrupt
seems to sometimes fire and cause a panic in the RX mbuf refresh
code. This is addressed by setting interrupts late in the init
path, and also to set all interrupts bits off at the start of that.
On my hardware, "em" in netmap mode does about 1.388 Mpps
on one card (on an Asus motherboard), and 1.1 Mpps on another
card (PCIe bus). Both seem to be NIC-limited, because
i have the same rate even with the CPU running at 150 MHz.
On the "re" driver the tx throughput is around 420-450 Kpps
on various (8111C and the like) chipsets. On the Rx side
performance seems much better, and i can receive the full
load generated by the "em" cards.
"igb" is untested as i don't have the hardware.
The SYSCTL_NODE macro defines a list that stores all child-elements of
that node. If there's no SYSCTL_DECL macro anywhere else, there's no
reason why it shouldn't be static.
em drivers.
- Make the per-instance 'enable_aim' sysctl truly per-instance by having it
change a per-instance variable (which is used to control AIM) rather
than having all of the per-instance sysctls operate on a single global
variable.
Reviewed by: jfv (earlier version)
MFC after: 1 week
particular flow control and dma coalesce. Also improve the
sysctl operation on those too.
Add IPv6 detection in the ioctl code, this was done for
ixgbe first, carrying that over.
Add resource ability to disable particular adapter.
Add HW TSO capability so vlans can make use of TSO
method instead of reusing the existing per-queue interrupt task.
Reusing the per-queue interrupt task could result in both an interrupt
thread and the taskqueue thread trying to handle received packets on a
single queue resulting in out-of-order packet processing.
- Don't define igb_start() at all on 8.0 and where if_transmit is used.
Replace last remaining call to igb_start() with a loop to kick off
transmit on each queue instead.
- Call ether_ifdetach() earlier in igb_detach().
- Drain tasks and free taskqueues during igb_detach().
Reviewed by: jfv
MFC after: 1 week
- Add the change made in em to the actual unrefreshed number
of descriptors is used as a basis in rxeof on the way out
to determine if more refresh is needed. NOTE: there is a
difference in the ring setup in igb, this is not accidental,
it is necessitated by hardware behavior, when you reset the
newer adapters it will not let you write RDH, it ALWAYS sets
it to 0. Thus the way em does it is not possible.
- Change the sysctl handling of flow control, it will now make
the change dynamically when the variable setting changes rather
than requiring a reset.
- Change the eee sysctl naming, validation found the old unintuitive :)
- Last but not least, some important performance tweaks in the TX
path, I found that UDP behavior could be drastically hindered or
improved with just small changes in the start loop. What I have
here is what testing has shown to be the best overall. Its interesting
to note that changing the clean threshold to start at a full half of
the ring, made a BIG difference in performance. I hope that this
will prove to be advantageous for most workloads.
MFC in a week.
been undergoing test for some weeks. This improves the RX
mbuf handling to avoid system hang due to depletion. Thanks
to all those who have been testing the code, and to Beezar
Liu for the design changes.
Next the igb driver is updated for similar RX changes, but
also to add new features support for our upcoming i350 family
of adapters.
MFC after a week
- this also includes virtualization support on these devices
Correct some vlan issues we were seeing in test, jumbo frames on vlans
did not work correctly, this was all due to confused logic around HW
filters, the new code should now work for all uses.
Important fix: when mbuf resources are depeleted, it was possible to
completely empty the RX ring, and then the RX engine would stall
forever. This is fixed by a flag being set whenever the refresh code
fails due to an mbuf shortage, also the local timer now makes sure
that all queues get an interrupt when it runs, the interrupt code
will then always call rxeof, and in that routine the first thing done
is now to check the refresh flag and call refresh_mbufs. This has been
verified to fix this type 'hang'. Similar code will follow in the other
drivers.
Finally, sync up shared code for the I350 support.
Thanks to everyone that has been reporting issues, and helping in the
debug/test process!!
the size can be smaller than the constant when you are
doing HW TAGGING, and you still need to process this
packet in a normal way. I'm not sure where the notion
to just return came from, but its wrong.
MFC after: 3 days
- Fixes from John Baldwin: vlan shadow tables made per/interface,
make vlan hw setup only happen when capability enabled, and
finally, make a tuneable interrupt rate. Thanks John!
- Tweaked watchdog handling to avoid any false positives, now
detection is in the TX clean path, with only the final check
and init happening in the local timer.
- limit queues to 8 for all devices, with 82576 or 82580 on
larger machines it can get greater than this, and it seems
mostly a resource waste to do so. Even 8 might be high but
it can be manually reduced.
- use 2k, 4k and now 9k clusters based on the MTU size.
- rework the igb_refresh_mbuf() code, its important to
make sure the descriptor is rewritten even when reusing
mbufs since writeback clobbers things.
MFC: in a few days, this delta needs to get to 8.2
- Sync shared code with Intel internal
- New client chipset support added
- em driver - fixes to 82574, limit queues to 1 but use MSIX
- em driver - large changes in TX checksum offload and tso
code, thanks to yongari.
- some small changes for watchdog issues.
- igb driver - local timer watchdog code was missing locking
this and a couple other watchdog related fixes.
- bug in rx discard found by Andrew Boyer, check for null pointer
MFC: a week
- Add a single sysctl procedure to all three drivers to read an arbitrary
register (the register is passed as arg2). Use it to replace existing
routines in igb(4) that used a separate routine for each register, and
to add support for missing stats in em(4) and lem(4).
- Move the 'rx_overruns' and 'watchdog_timeouts' stats out of the MAC stats
section as they are driver stats, not MAC counters.
- Simplify the code that creates per-queue stats in igb(4) to use a single
loop and remove duplicated code.
- Properly read all 64 bits of the 'good octets received/transmitted' in
em(4) and lem(4).
- Actually read the interrupt count registers in em(4), and drop the
'host to card' sysctl stats from em(4) as they are not implemented in
any of the hardware this driver supports.
- Restore several stats to em(4) that were lost in the earlier stats
conversion including per-queue stats.
- Export several MAC stats in em(4) that were exported in igb(4) but not
in em(4).
- Export stats in lem(4) using individual sysctls as in em(4) and igb(4).
Reviewed by: jfv
MFC after: 1 week
configuration function. For failed memory allocations, em(4)/lem(4)
called panic(9) which is not acceptable on production box.
igb(4)/ixgb(4)/ix(4) allocated the required memory in stack which
consumed 768 bytes of stack memory which looks too big.
To address these issues, allocate multicast array memory in device
attach time and make multicast configuration success under any
conditions. This change also removes the excessive use of memory in
stack.
Reviewed by: jfv
