- Ask the firmware for the number of frames that can be stuffed in one
work request.
- Modify mp_ring to increase the likelihood of tx coalescing when there
are just one or two threads that are doing most of the tx. Add teeth
to the abdication mechanism by pushing the consumer lock into mp_ring.
This reduces the likelihood that a consumer will get stuck with all
the work even though it is above its budget.
- Add support for coalesced tx WR to the VF driver. This, with the
changes above, results in a 7x improvement in the tx pps of the VF
driver for some common cases. The firmware vets the L2 headers
submitted by the VF driver and it's a big win if the checks are
performed for a batch of packets and not each one individually.
Reviewed by: jhb@
MFC after: 2 weeks
Sponsored by: Chelsio Communications
Differential Revision: https://reviews.freebsd.org/D25454
There were quite a few places where port_info was being accessed only to
get to the adapter.
Reviewed by: jhb@
MFC after: 1 week
Sponsored by: Chelsio Communications
Differential Revision: https://reviews.freebsd.org/D25432
o Shrink sglist(9) functions to work with multipage mbufs down from
four functions to two.
o Don't use 'struct mbuf_ext_pgs *' as argument, use struct mbuf.
o Rename to something matching _epg.
Reviewed by: gallatin
Differential Revision: https://reviews.freebsd.org/D24598
The following series of patches addresses three things:
Now that array of pages is embedded into mbuf, we no longer need
separate structure to pass around, so struct mbuf_ext_pgs is an
artifact of the first implementation. And struct mbuf_ext_pgs_data
is a crutch to accomodate the main idea r359919 with minimal churn.
Also, M_EXT of type EXT_PGS are just a synonym of M_NOMAP.
The namespace for the newfeature is somewhat inconsistent and
sometimes has a lengthy prefixes. In these patches we will
gradually bring the namespace to "m_epg" prefix for all mbuf
fields and most functions.
Step 1 of 4:
o Anonymize mbuf_ext_pgs_data, embed in m_ext
o Embed mbuf_ext_pgs
o Start documenting all this entanglement
Reviewed by: gallatin
Differential Revision: https://reviews.freebsd.org/D24598
While the original implementation of unmapped mbufs was a large
step forward in terms of reducing cache misses by enabling mbufs
to carry more than a single page for sendfile, they are rather
cache unfriendly when accessing the ext_pgs metadata and
data. This is because the ext_pgs part of the mbuf is allocated
separately, and almost guaranteed to be cold in cache.
This change takes advantage of the fact that unmapped mbufs
are never used at the same time as pkthdr mbufs. Given this
fact, we can overlap the ext_pgs metadata with the mbuf
pkthdr, and carry the ext_pgs meta directly in the mbuf itself.
Similarly, we can carry the ext_pgs data (TLS hdr/trailer/array
of pages) directly after the existing m_ext.
In order to be able to carry 5 pages (which is the minimum
required for a 16K TLS record which is not perfectly aligned) on
LP64, I've had to steal ext_arg2. The only user of this in the
xmit path is sendfile, and I've adjusted it to use arg1 when
using unmapped mbufs.
This change is almost entirely mechanical, except that we
change mb_alloc_ext_pgs() to no longer allow allocating
pkthdrs, the change to avoid ext_arg2 as mentioned above,
and the removal of the ext_pgs zone,
This change saves roughly 2% "raw" CPU (~59% -> 57%), or over
3% "scaled" CPU on a Netflix 100% software kTLS workload at
90+ Gb/s on Broadwell Xeons.
In a follow-on commit, I plan to remove some hacks to avoid
access ext_pgs fields of mbufs, since they will now be in
cache.
Many thanks to glebius for helping to make this better in
the Netflix tree.
Reviewed by: hselasky, jhb, rrs, glebius (early version)
Sponsored by: Netflix
Differential Revision: https://reviews.freebsd.org/D24213
This reduces the lines bouncing around between the driver rx ithread and
the netmap rxsync thread. There is no net change in the size of the
struct (it continues to waste a lot of space).
This kind of split was originally proposed in D17869 by Marc De La
Gueronniere @ Verisign, Inc.
MFC after: 1 week
Sponsored by: Chelsio Communications
already allocating from the safe zone and the allocation fails.
This bug was introduced in r357481.
MFC after: 3 days
Sponsored by: Chelsio Communications
r357614 added CTLFLAG_NEEDGIANT to make it easier to find nodes that are
still not MPSAFE (or already are but aren’t properly marked).
Use it in preparation for a general review of all nodes.
This is non-functional change that adds annotations to SYSCTL_NODE and
SYSCTL_PROC nodes using one of the soon-to-be-required flags.
Mark all obvious cases as MPSAFE. All entries that haven't been marked
as MPSAFE before are by default marked as NEEDGIANT
Approved by: kib (mentor, blanket)
Commented by: kib, gallatin, melifaro
Differential Revision: https://reviews.freebsd.org/D23718
It duplicated the kern_tls_records stat and was not conditional on NIC
TLS being enabled.
Reviewed by: np
Sponsored by: Chelsio Communications
Differential Revision: https://reviews.freebsd.org/D23670
This simplifies the driver's rx fast path as well as the bookkeeping
code that tracks various rx buffer sizes and layouts.
MFC after: 1 week
Sponsored by: Chelsio Communications
ext_arg2 is the only item in the third cacheline in an mbuf and could be
cold by the time rxb_free runs. Put the information needed by rxb_free
in the same line as the refcount, which is very likely to be hot given
that rxb_free runs when the refcount is decremented and reaches 0.
MFC after: 1 week
Sponsored by: Chelsio Communications
requested.
This is a tradeoff between PCIe efficiency during large packet rx and
packing efficiency during small packet rx.
MFC after: 1 week
Sponsored by: Chelsio Communications
CPL_TX_PKT_XT disables the internal parser on the chip and instead
relies on the driver to provide the exact length of the L2 and L3
headers. This allows hw checksumming and TSO to be used with L2 and
L3 encapsulations that the chip doesn't understand directly.
Note that netmap tx still uses the old CPL as it never uses the hw
to generate the checksum on tx.
Reviewed by: jhb@
MFC after: 1 month
Sponsored by: Chelsio Communications
Differential Revision: https://reviews.freebsd.org/D22788
TX_PKTS2 is more efficient within the firmware and this improves netmap
Tx by a few Mpps in some common scenarios.
MFC after: 1 week
Sponsored by: Chelsio Communications
This adds support for ifnet (NIC) KTLS using Chelsio T6 adapters.
Unlike the TOE-based KTLS in r353328, NIC TLS works with non-TOE
connections.
NIC KTLS on T6 is not able to use the normal TSO (LSO) path to segment
the encrypted TLS frames output by the crypto engine. Instead, the
TOE is placed into a special setup to permit "dummy" connections to be
associated with regular sockets using KTLS. This permits using the
TOE to segment the encrypted TLS records. However, this approach does
have some limitations:
1) Regular TOE sockets cannot be used when the TOE is in this special
mode. One can use either TOE and TOE-based KTLS or NIC KTLS, but
not both at the same time.
2) In NIC KTLS mode, the TOE is only able to accept a per-connection
timestamp offset that varies in the upper 4 bits. Put another way,
only connections whose timestamp offset has the 28 lower bits
cleared can use NIC KTLS and generate correct timestamps. The
driver will refuse to enable NIC KTLS on connections with a
timestamp offset with any of the lower 28 bits set. To use NIC
KTLS, users can either disable TCP timestamps by setting the
net.inet.tcp.rfc1323 sysctl to 0, or apply a local patch to the
tcp_new_ts_offset() function to clear the lower 28 bits of the
generated offset.
3) Because the TCP segmentation relies on fields mirrored in a TCB in
the TOE, not all fields in a TCP packet can be sent in the TCP
segments generated from a TLS record. Specifically, for packets
containing TCP options other than timestamps, the driver will
inject an "empty" TCP packet holding the requested options (e.g. a
SACK scoreboard) along with the segments from the TLS record.
These empty TCP packets are counted by the
dev.cc.N.txq.M.kern_tls_options sysctls.
Unlike TOE TLS which is able to buffer encrypted TLS records in
on-card memory to handle retransmits, NIC KTLS must re-encrypt TLS
records for retransmit requests as well as non-retransmit requests
that do not include the start of a TLS record but do include the
trailer. The T6 NIC KTLS code tries to optimize some of the cases for
requests to transmit partial TLS records. In particular it attempts
to minimize sending "waste" bytes that have to be given as input to
the crypto engine but are not needed on the wire to satisfy mbufs sent
from the TCP stack down to the driver.
TCP packets for TLS requests are broken down into the following
classes (with associated counters):
- Mbufs that send an entire TLS record in full do not have any waste
bytes (dev.cc.N.txq.M.kern_tls_full).
- Mbufs that send a short TLS record that ends before the end of the
trailer (dev.cc.N.txq.M.kern_tls_short). For sockets using AES-CBC,
the encryption must always start at the beginning, so if the mbuf
starts at an offset into the TLS record, the offset bytes will be
"waste" bytes. For sockets using AES-GCM, the encryption can start
at the 16 byte block before the starting offset capping the waste at
15 bytes.
- Mbufs that send a partial TLS record that has a non-zero starting
offset but ends at the end of the trailer
(dev.cc.N.txq.M.kern_tls_partial). In order to compute the
authentication hash stored in the trailer, the entire TLS record
must be sent as input to the crypto engine, so the bytes before the
offset are always "waste" bytes.
In addition, other per-txq sysctls are provided:
- dev.cc.N.txq.M.kern_tls_cbc: Count of sockets sent via this txq
using AES-CBC.
- dev.cc.N.txq.M.kern_tls_gcm: Count of sockets sent via this txq
using AES-GCM.
- dev.cc.N.txq.M.kern_tls_fin: Count of empty FIN-only packets sent to
compensate for the TOE engine not being able to set FIN on the last
segment of a TLS record if the TLS record mbuf had FIN set.
- dev.cc.N.txq.M.kern_tls_records: Count of TLS records sent via this
txq including full, short, and partial records.
- dev.cc.N.txq.M.kern_tls_octets: Count of non-waste bytes (TLS header
and payload) sent for TLS record requests.
- dev.cc.N.txq.M.kern_tls_waste: Count of waste bytes sent for TLS
record requests.
To enable NIC KTLS with T6, set the following tunables prior to
loading the cxgbe(4) driver:
hw.cxgbe.config_file=kern_tls
hw.cxgbe.kern_tls=1
Reviewed by: np
Sponsored by: Chelsio Communications
Differential Revision: https://reviews.freebsd.org/D21962
NIC KTLS will add a new TLS send tag type in cxgbe(4) that is a
distinct tag from a ratelimit tag. To support this, refactor
cxgbe_snd_tag to be a simple send tag with a type and convert the
existing ratelimit tag to a new cxgbe_rate_tag structure.
Reviewed by: np
Sponsored by: Chelsio Communications
Differential Revision: https://reviews.freebsd.org/D22072
Since cxgbe(4) uses sglist instead of bus_dma, this required updates
to the code that generates scatter/gather lists for packets. Also,
unmapped mbufs are always sent via DMA and never as immediate data in
the payload of a work request.
Submitted by: gallatin (earlier version)
Reviewed by: gallatin, hselasky, rrs
Discussed with: np
Sponsored by: Netflix
Differential Revision: https://reviews.freebsd.org/D20616
- Perform ifp mismatch checks (to determine if a send tag is allocated
for a different ifp than the one the packet is being output on), in
ip_output() and ip6_output(). This avoids sending packets with send
tags to ifnet drivers that don't support send tags.
Since we are now checking for ifp mismatches before invoking
if_output, we can now try to allocate a new tag before invoking
if_output sending the original packet on the new tag if allocation
succeeds.
To avoid code duplication for the fragment and unfragmented cases,
add ip_output_send() and ip6_output_send() as wrappers around
if_output and nd6_output_ifp, respectively. All of the logic for
setting send tags and dealing with send tag-related errors is done
in these wrapper functions.
For pseudo interfaces that wrap other network interfaces (vlan and
lagg), wrapper send tags are now allocated so that ip*_output see
the wrapper ifp as the ifp in the send tag. The if_transmit
routines rewrite the send tags after performing an ifp mismatch
check. If an ifp mismatch is detected, the transmit routines fail
with EAGAIN.
- To provide clearer life cycle management of send tags, especially
in the presence of vlan and lagg wrapper tags, add a reference count
to send tags managed via m_snd_tag_ref() and m_snd_tag_rele().
Provide a helper function (m_snd_tag_init()) for use by drivers
supporting send tags. m_snd_tag_init() takes care of the if_ref
on the ifp meaning that code alloating send tags via if_snd_tag_alloc
no longer has to manage that manually. Similarly, m_snd_tag_rele
drops the refcount on the ifp after invoking if_snd_tag_free when
the last reference to a send tag is dropped.
This also closes use after free races if there are pending packets in
driver tx rings after the socket is closed (e.g. from tcpdrop).
In order for m_free to work reliably, add a new CSUM_SND_TAG flag in
csum_flags to indicate 'snd_tag' is set (rather than 'rcvif').
Drivers now also check this flag instead of checking snd_tag against
NULL. This avoids false positive matches when a forwarded packet
has a non-NULL rcvif that was treated as a send tag.
- cxgbe was relying on snd_tag_free being called when the inp was
detached so that it could kick the firmware to flush any pending
work on the flow. This is because the driver doesn't require ACK
messages from the firmware for every request, but instead does a
kind of manual interrupt coalescing by only setting a flag to
request a completion on a subset of requests. If all of the
in-flight requests don't have the flag when the tag is detached from
the inp, the flow might never return the credits. The current
snd_tag_free command issues a flush command to force the credits to
return. However, the credit return is what also frees the mbufs,
and since those mbufs now hold references on the tag, this meant
that snd_tag_free would never be called.
To fix, explicitly drop the mbuf's reference on the snd tag when the
mbuf is queued in the firmware work queue. This means that once the
inp's reference on the tag goes away and all in-flight mbufs have
been queued to the firmware, tag's refcount will drop to zero and
snd_tag_free will kick in and send the flush request. Note that we
need to avoid doing this in the middle of ethofld_tx(), so the
driver grabs a temporary reference on the tag around that loop to
defer the free to the end of the function in case it sends the last
mbuf to the queue after the inp has dropped its reference on the
tag.
- mlx5 preallocates send tags and was using the ifp pointer even when
the send tag wasn't in use. Explicitly use the ifp from other data
structures instead.
- Sprinkle some assertions in various places to assert that received
packets don't have a send tag, and that other places that overwrite
rcvif (e.g. 802.11 transmit) don't clobber a send tag pointer.
Reviewed by: gallatin, hselasky, rgrimes, ae
Sponsored by: Netflix
Differential Revision: https://reviews.freebsd.org/D20117
Drivers can now pass up numa domain information via the
mbuf numa domain field. This information is then used
by TCP syncache_socket() to associate that information
with the inpcb. The domain information is then fed back
into transmitted mbufs in ip{6}_output(). This mechanism
is nearly identical to what is done to track RSS hash values
in the inp_flowid.
Follow on changes will use this information for lacp egress
port selection, binding TCP pacers to the appropriate NUMA
domain, etc.
Reviewed by: markj, kib, slavash, bz, scottl, jtl, tuexen
Sponsored by: Netflix
Differential Revision: https://reviews.freebsd.org/D20028
Recent firmwares prefer to use a different format for viid internally
and this change allows them to do so.
MFC after: 1 week
Sponsored by: Chelsio Communications
"slow" interrupt handler:
- Expand the list of INT_CAUSE registers known to the driver.
- Add decode information for many more bits but decouple it from the
rest of intr_info so that it is entirely optional.
- Call t4_fatal_err exactly once, and from the top level PL intr handler.
t4_fatal_err:
- Use t4_shutdown_adapter from the common code to stop the adapter.
- Stop servicing slow interrupts after the first fatal one.
Driver/firmware interaction:
- CH_DUMP_MBOX: note whether the mailbox being dumped is a command or a
reply or something else.
- Log the raw value of pcie_fw for some errors.
- Use correct log levels (debug vs. error).
Sponsored by: Chelsio Communications
through.
cxgb4vf doesn't own the buffer size list but still expects the first two
entries to be 4K and some power of 2 respectively. The BSD cxgbe
doesn't care where its preferred buffer sizes are as long as they're in
the list somewhere, so just move its entries towards the end as a
workaround.
MFC after: 1 month
Sponsored by: Chelsio Communicatons
- Use PH_loc.eight[1] as a general 'cflags' (Chelsio flags) field to
describe properties of a queued packet. The MC_RAW_WR flag
indicates an mbuf holding a raw work request. mbuf_cflags() returns
the current flags.
- Raw work request mbufs are allocated via alloc_wr_mbuf() which will
allocate a single contiguous range to hold the mbuf data. The
consumer can use mtod() to obtain the start of the work request and
write the required work request in the buffer. The mbuf can then be
enqueued directly to the txq via mp_ring_enqueue().
- Since raw work requests might potentially send arbitrary work
requests, only set the EQUIQ and EQUEQ bits on work requests that
support them such as the normal tunneled Ethernet packet work
requests.
Reviewed by: np
Sponsored by: Chelsio Communications
Differential Revision: https://reviews.freebsd.org/D17811
Currently this is a no-op, but will matter in the future when
cannot_use_txpkts() starts checking other conditions than just
needs_tso().
Sponsored by: Chelsio Communications
The bits that explicitly request cidx updates do not work reliably with
all possible WRs that can be sent over the queue. The F_FW_WR_EQUIQ
requests that still remain may also have to be replaced with explicit
credit flush WRs in the future.
MFC after: 2 days
Sponsored by: Chelsio Communications
There used to be one control queue per adapter (the mgmtq) that was
initialized during adapter init and one per port that was initialized
later during port init. This change moves all the control queues (one
per port/channel) to the adapter so that they are initialized during
adapter init and are available before any port is up. This allows the
driver to issue ctrlq work requests over any channel without having to
bring up any port.
MFH: 2 weeks
Sponsored by: Chelsio Communications
