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
previous names, 'ptag' and 'pmap' -- p stands for packet.
This change reduces the difference between the code in stable/9
and head, and also helps using the same ixgbe_netmap.h on both branches.
Approved by: Jack Vogel
thought I've decided its overkill,a simple tuneable for
each RX and TX limit, and then init sets the ring values
based on that, should be sufficient.
More importantly, fix a bug causing a panic, when changing
the define style to IXGBE_LEGACY_TX a taskqueue init was
inadvertently set #ifdef when it should be #ifndef.
the revamped sysctl code did not work, and needed a change. This
makes the limit get set at the time that all sysctl stats are
created and is actually more elegant imho anyway.
TX hot path by getting rid of index calculations and simply
managing pointers. Much of the creative code is due to my
coworker here at Intel, Alex Duyck, thanks Alex!
Also, this whole series of patches was given the critical
eye of Gleb Smirnoff and is all the better for it, thanks
Gleb!
- add a limit for both RX and TX, change the default to 256
- change the sysctl usage to be common, and now to be called
during init for each ring.
- the TX limit is not yet used, but the changes in the last
patch in this series uses the value.
- the motivation behind these changes is to improve data
locality in the final code.
- rxeof interface changes since it now gets limit from the
ring struct
defines (at Gleb's request). Also, change the defines around
the old transmit code to IXGBE_LEGACY_TX, I do this to make
it possible to define this regardless of the OS level (it is
not defined by default). There are also a couple changed
comments for clarity.
these are FCOE stats (fiber channel over ethernet), something that
FreeBSD does not yet have, they were mistaken for flow control by
the implementor I believe. Secondly, the real flow control stats
are oddly named with a 'link' tag on the front, it was requested
by my validation engineer to make these stats have the same name as
the igb driver for clarity and that seemed reasonable to me.
multiqueue code, this functionality has proven to be more
trouble than it was worth. Thanks to Gleb for a second
critical look over my code and help in the patches!
- Testing TSO6 has led me to discover that HW RSC is
a problematic feature, it is ONLY designed to work
with IPv4 in the first place, and if IP forwarding
is done it can't be disabled as LRO in the stack,
also initial testing we've done at Intel shows an
equal performance using TSO[46] on the TX and LRO
on RX, if you ran older code on 82599 or later hardware
you actually could have detrimental performance for
this reason. So I am disabling the feature by default
and all our adapters will now use LRO instead.
- If you have flow control off and multiple queues it
was possible when the buffer of one queue becomes
full that all RX movement is stalled, to eliminate
this problem a feature bit is now set that will allow
packets to be dropped when full rather than stall.
Note, the default is to have flow control on, and this
keeps this from happening.
- Because of the recent fixes in the stack, LRO is now
auto-disabled when problematic, so I have decided to
enable it by default in the capabilities in the driver.
- There are some 1G modules used by some customers, a couple
small tweaks to properly support those in the media code.
- A note: we have now done some testing of TSO6 and using
LRO with IPv6 and it all works great!! Seeing line rate
in both directions in best cases. Thanks bz for your
excellent work!!
this was designed to keep duplicate null vlan tags from
being added. When doing vlans purely via the switch
this problem will occur. Reported by external customer.
device drivers that used to provide this feature.
This is a subset of 241856 (which was reverted)
Reviewed by: des
Approved by: cperciva (implicit)
MFC after: 1 week
around the problem where high speed interfaces (such as ixgbe(4))
are not able to report real ifi_baudrate. bascially, take a spare
byte from struct if_data and use it to store ifi_baudrate power
factor. in other words,
real ifi_baudrate = ifi_baudrate * 10 ^ ifi_baudrate power factor
this should be backwards compatible with old binaries. use ixgbe(4)
as an example on how drivers would set ifi_baudrate power factor
Discussed with: kib, scottl, glebius
MFC after: 1 week
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
- 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 and lock
contention.
- Don't define ixgbe_start() at all where if_transmit is used.
Tested by: Vijay Singh
Reviewed by: jfv
MFC after: 2 weeks
this case, allocate a plain mbuf and copy the frame into it, then send the
copy up the stack, leaving the original mbuf+cluster in place in the
receive ring for immediate re-use. This saves a trip through 2 of the
3 zones of the compound mbuf allocator, a trip through busdma, and a trip
through the 1 of the 3 mbuf destructors. For our load at Netflix, this can
lower CPU consumption by as much as 20%. The copy algorithm is based on
investigative work from Luigi Rizzo earlier in the year.
Reviewed by: jfv
Obtained from: Netflix
- Add a couple of new devices
- Flow control changes in shared and core code
- Bug fix to Flow Director for 82598
- Shared code sync to internal with required core change
Thanks to those helping in the testing and improvements to this driver!
MFC after:5 days
Put a bandaid to prevent ixgbe(4) from completely locking up the system
under high load. Our platform has a few CPU cores and a single active
ixgbe(4) port with 4 queues. Under high enough traffic load, at about
7.5GBs and 700,000 packets/sec (outbound), the entire system would
deadlock. What we found was that each CPU was in an endless loop on a
different ix taskqueue thread. The OACTIVE flag had gotten set on each
queue, and the ixgbe_handle_queue() function was continuously rescheduling
itself via the taskqueue_enqueue. Since all CPUs were busy with their
taskqueue threads, the ixgbe_local_timer() function couldn't run to clear
the OACTIVE flag.
Submitted by: scottl
MFC after: 1 week
Add TSO6 and LRO/IPv6 support.
Fix the module Makefile to at least properly inlcude opt_inet6.h
and allow builds without INET or INET6.
Sponsored by: The FreeBSD Foundation
Sponsored by: iXsystems
Reviewed by: gnn (as part of the whole)
MFC After: 3 days
Contrarily to what i wrote in my previous commit, the 82599
does include the CRC in the length. The operating mode is
reset in ixgbe_init_locked() and so we need to hook into
the places where the two registers (HLREG0 and RDRXCTL) are
modified.
values as in the Intel driver 3.8.21 for linux. The fact that it
is standard in the above driver suggests that it has no bad side
effects.
But of course there must be a reason for enabling features, not
just "it does not harm", so here it is a good one:
Prefetching enables full line rate even using a single queue (14.88
Mpps, compared to ~12 Mpps without prefetch). This in turn is
terribly useful when one wants to schedule traffic.
For obvious reasons the difference is only visible with netmap
or other high speed solutions, but presumably the advantage
should be in the order of a fraction of a microsecond when
starting transmission on an empty queue.
Discussed with Jack Vogel.
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.
- remove experimental code for disabling CRC
- use the correct constant for conversion between interrupt rate
and EITR values (the previous values were off by a factor of 2)
- make dev.ix.N.queueM.interrupt_rate a RW sysctl variable.
Changing individual values affects the queue immediately,
and propagates to all interfaces at the next reinit.
- add dev.ix.N.queueM.irqs rdonly sysctl, to export the actual
interrupt counts
Netmap-related changes for ixgbe:
- use the "new" format for TX descriptors in netmap mode.
- pass interrupt mitigation delays to the user process doing poll()
on a netmap file descriptor.
On the RX side this means we will not check the ring more than once
per interrupt. This gives the process a chance to sleep and process
packets in larger batches, thus reducing CPU usage.
On the TX side we take this even further: completed transmissions are
reclaimed every half ring even if the NIC interrupts more often.
This saves even more CPU without any additional tx delays.
Generic Netmap-related changes:
- align the netmap_kring to cache lines so that there is no false sharing
(possibly useful for multiqueue NICs and MSIX interrupts, which are
handled by different cores). It's a minor improvement but it does not
cost anything.
Reviewed by: Jack Vogel
Approved by: Jack Vogel
1. correct the initialization of RDT when there is an ixgbe_init()
while a netmap client is active. This code was previously
in ixgbe_initialize_receive_units() but RDT is overwritten
shortly afterwards in ixgbe_init_locked()
2. add code (not active yet) to disable CRCSTRIP while in netmap mode.
From all evidence i could gather, it seems that when the 82599 has to
write a data block that is not a full cache line, it first reads
the line (64 bytes) and then writes back the updated version.
This hurts reception of min-sized frames, which are only 60 bytes
if the CRC is stripped: i could never get above 11Mpps
(received from one queue) with CRCSTRIP enabled, whyle 64+4-byte
packets reach 14.2 Mpps (the theoretical maximum).
Leaving the CRC in gets us 14.88Mpps for 60+4 byte frames,
(and penalizes 64+4). The min-size case is important not just because
it looks good in benchmarks, but also because this is the size
of pure acks.
Note we cannot leave CRCSTRIP on by default because it is
incompatible with some other features (LRO etc.)
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
- {ixgbe,ixv}_header_split is passed to TUNABLE_INT, so delcare it
int, not bool.
- {ixgbe,ixv}_tx_ctx_setup() returns a boolean value, so declare it
bool, not int.
- {ixgbe,ixv}_tso_setup() returns a bool, so declare it bool, not boolean_t.
- {ixgbe,ixv}_txeof() returns a bool, so declare it bool, not boolean_t.
- Do not re-define bool if the symbol already exists.
MFC after: 2 weeks
Sponsored by: Isilon Systems, LLC
A link reset now is completely transparent for the netmap client:
even if the NIC resets its own ring (e.g. restarting from 0),
the client will not see any change in the current rx/tx positions,
because the driver will keep track of the offset between the two.
2. make the device-specific code more uniform across different drivers
There were some inconsistencies in the implementation of the netmap
support routines, now drivers have been aligned to a common
code structure.
3. import netmap support for ixgbe . This is implemented as a very
small patch for ixgbe.c (233 lines, 11 chunks, mostly comments:
in total the patch has only 54 lines of new code) , as most of
the code is in an external file sys/dev/netmap/ixgbe_netmap.h ,
following some initial comments from Jack Vogel about making
changes less intrusive.
(Note, i have emailed Jack multiple times asking if he had
comments on this structure of the code; i got no reply so
i assume he is fine with it).
Support for other drivers (em, lem, re, igb) will come later.
"ixgbe" is now the reference driver for netmap support. Both the
external file (sys/dev/netmap/ixgbe_netmap.h) and the device-specific
patches (in sys/dev/ixgbe/ixgbe.c) are heavily commented and should
serve as a reference for other device drivers.
Tested on i386 and amd64 with the pkt-gen program in tools/tools/netmap,
the sender does 14.88 Mpps at 1050 Mhz and 14.2 Mpps at 900 MHz
on an i7-860 with 4 cores and 82599 card. Haven't tried yet more
aggressive optimizations such as adding 'prefetch' instructions
in the time-critical parts of the code.
The current code was rounding down the maximum frame size instead of
routing up, resulting in a read size of 1024 bytes, in the non-jumbo
frame case, and splitting the packets across multiple mbufs.
Consequently the above problem exposed another issue, which is when
packets were splitted across multiple mbufs, and all of the mbufs in the
chain have the M_PKTHDR flag set.
Submitted by: original patch by Ray Ruvinskiy at BlueCoat dot com
Reviewed by: jfv, kmacy, rwatson
Approved by: re (rwatson)
MFC after: 5 days
crusty, and this still isn't perfect, but its at least a bit
more recent.
Secondly, a few improvements to the driver from Andrew Boyer,
support hint to allow devices to not attach, add VLAN_HWTSO
capability so vlans can use TSO, fix in the interrupt handler
to make sure the stack TX queue is processed. Oh, and also
make sure IPv6 does not cause a re-init in the ioctl routine.
Thanks for your efforts Andrew!
Thanks to Claudio Jeker for noticing the ixgbe_xmit() routine
was not correctly swapping the dma map from the first to the
last descriptor in a multi-descriptor transmission, corrected
this.
- Also a couple minor tweaks to the TX code from the same source.
- Add the INET ioctl code which has been missing from this driver,
and which caused IP aliases to reset the interface.
- Last, some minor logic changes that just reflect upcoming
hardware support, but have no other functional effect now.
MFC after a week
