In case that VMBUS_CHAN_ISPRIMARY is needed in the early place of
channel setup.
MFC after: 1 week
Sponsored by: Microsoft OSTC
Differential Revision: https://reviews.freebsd.org/D7108
This paves way for the further cleanup/disentangle.
MFC after: 1 week
Sponsored by: Microsoft OSTC
Differential Revision: https://reviews.freebsd.org/D7092
This paves way to nuke the hv_device, which is actually an unncessary
indirection.
MFC after: 1 week
Sponsored by: Microsoft OSTC
Differential Revision: https://reviews.freebsd.org/D7033
This prepares to remove the unnecessary offer message embedding in
hv_vmbus_channel.
MFC after: 1 week
Sponsored by: Microsoft OSTC
Differential Revision: https://reviews.freebsd.org/D7015
This prepares to remove the unnecessary offer message embedding in
hv_vmbus_channel.
MFC after: 1 week
Sponsored by: Microsoft OSTC
Differential Revision: https://reviews.freebsd.org/D7014
For multi-channel devices, once the primary channel is closed,
a set of 'rescind' messages for sub-channels will be delivered
by Hypervisor. Sub-channel MUST be freed according to these
'rescind' messages; directly re-openning sub-channels in the
same fashion as the primary channel's re-opening does NOT work
at all.
After the primary channel is re-opened, requested # of sub-
channels will be delivered though 'channel offer' messages, and
this set of newly offered channels can be opened along side with
the primary channel.
This unbreaks the MTU setting for hn(4), which requires re-
openning all existsing channels upon MTU change.
MFC after: 1 week
Sponsored by: Microsoft OSTC
Differential Revision: https://reviews.freebsd.org/D6978
Instead of global variable, vmbus version is accessed through
a vmbus DEVMETHOD now.
MFC after: 1 week
Sponsored by: Microsoft OSTC
Differential Revision: https://reviews.freebsd.org/D6953
This greatly reduces the oqdrops under heavy workload.
For TCP send/recv test (10K concurrent connections):
oqdrops is reduced by 17% on sending side, and 57% on receiving side.
For nginx-1.8/wrk-4 1KB object test (10K concurrent connections,
4 requests/connection):
oqdrops is reduced by 44% on nginx side, and 10% on wrk side.
MFC after: 1 week
Sponsored by: Microsoft OSTC
Since the sub-channel offers are synchronized, we can do our own
channel setup without using the sub-channel creation callback.
This paves the way to whack the sub-channel creation callback.
MFC after: 1 week
Sponsored by: Microsoft OSTC
8 gives the best performance in both Azure and local Hyper-V on both
10Ge and 40Ge. More rings are still allowed by manual configuration.
Reviewed by: Dexuan Cui <decui microsoft com>
MFC after: 1 week
Sponsored by: Microsoft OSTC
Differential Revision: https://reviews.freebsd.org/D5879
And factor out tcp_lro_rx_done, which deduplicates the same logic with
netinet/tcp_lro.c
Reviewed by: gallatin (1st version), hps, zbb, np, Dexuan Cui <decui microsoft com>
Sponsored by: Microsoft OSTC
Differential Revision: https://reviews.freebsd.org/D5725
This mainly used to improve ACK timeliness when multiple RX rings
are enabled.
This value gives the best performance in both Azure and Hyper-V
environment, w/ both 10Ge and 40Ge using non-{INVARIANTS,WITNESS}
kernel.
MFC after: 1 week
Sponsored by: Microsoft OSTC
Differential Revision: https://reviews.freebsd.org/D5691
So that functions shared w/ attach path could use if_printf().
While I'm here, remove unnecessary if_dunit and if_dname assignment.
MFC after: 1 week
Sponsored by: Microsoft OSTC
Differential Revision: https://reviews.freebsd.org/D5576
Each channel contains one RX ring and one TX ring. And we
try to distribute the channels to different evenly.
Note: Currently we don't have enough information to extract
the RSS type and RSS hash value from the received packets.
This greatly improves the TX/RX performance for 8 virtual CPU
Hyper-V over 10Ge: it can max out 10Ge for TCP when multiple
RX/TX rings are enabled.
This almost doubles the TX/RX performance for locally connected
Hyper-Vs: was 6Gbps w/ 128 TCP streams, now 11Gbps w/ multiple
RX/TX rings enabled.
It is not enabled by default; it will be switched on after more
tests.
Collaborated with: Hongjiang Zhang <honzhan microsoft com>
MFC after: 2 week
Sponsored by: Microsoft OSTC
And since the host may not being able to allocate the # of rings
requested by us, save the # of rings allocated by the host in the
ring_inuse counters; use ring_inuse counters for run time operation.
This paves the way for the upcoming vRSS support.
MFC after: 1 week
Sponsored by: Microsoft OSTC
So that the host could dispatch the TX done back to this TX ring's
owner channel
MFC after: 1 week
Sponsored by: Microsoft OSTC
Differential Revision: https://reviews.freebsd.org/D5498
Chimney sending buffer still needs conversion, which will be done
along with the upcoming vRSS support.
MFC after: 1 week
Sponsored by: Microsoft OSTC
Differential Revision: https://reviews.freebsd.org/D5457
This fixes the TX/RX ring selection for TX/RX done.
MFC after: 1 week
Sponsored by: Microsoft OSTC
Differential Revision: https://reviews.freebsd.org/D5454
This is preamble to associate the TX/RX rings to their channel.
While I'm here, revoke unused netvsc_recv_rollup.
MFC after: 1 week
Sponsored by: Microsoft OSTC
Differential Revision: https://reviews.freebsd.org/D5453
This is the preamble to pass channel back to hn(4) upon TX/RX done.
Reviewed by: Hongjiang Zhang <honzhan microsoft com>
MFC after: 1 week
Sponsored by: Microsoft OSTC
Differential Revision: https://reviews.freebsd.org/D5452
IF_PREPEND promises out-of-order packet sending when the TX desc list
is depleted. It was overlooked and copied blindly when the transmission
path was partially rewritten.
MFC after: 1 week
Sponsored by: Microsoft OSTC
Differential Revision: https://reviews.freebsd.org/D5386
It will be shared w/ the upcoming ifnet.if_transmit method
implementation.
No functional change.
MFC after: 1 week
Sponsored by: Microsoft OSTC
Differential Revision: https://reviews.freebsd.org/D5385
need to include it explicitly when <vm/vm_param.h> is already included.
Suggested by: alc
Reviewed by: alc
Differential Revision: https://reviews.freebsd.org/D5379
It is only used in hv_netvsc_drv_freebsd.c; and rename it to hn_tx_done()
mainly to reserve "xmit" for ifnet.if_transmit implement.
While I'm here, remove unapplied comment.
Reviewed by: adrian
Approved by: adrian (mentor)
MFC after: 1 week
Sponsored by: Microsoft OSTC
Differential Revision: https://reviews.freebsd.org/D5345
Preamble to implement the ifnet.if_transmit method.
Reviewed by: adrian
Approved by: adrian (mentor)
MFC after: 1 week
Sponsored by: Microsoft OSTC
Differential Revision: https://reviews.freebsd.org/D5344
Tested on Windows Server 2012.
Reviewed by: adrian
Approved by: adrian (mentor)
MFC after: 1 week
Sponsored by: Microsoft OSTC
Differential Revision: https://reviews.freebsd.org/D5317
It will be used to help tracking host side transmission ring selection
issue; and it will be turned on by default, once we have concrete result.
Reviewed by: adrian, Jun Su <junsu microsoft com>
Approved by: adrian (mento)
MFC after: 1 week
Sponsored by: Microsoft OSTC
Differential Revision: https://reviews.freebsd.org/D5316
So one spinlock is avoided, which would be potentially dangerous for
virtual machine, if the spinlock holder was scheduled out by the host,
as noted by royger.
Old spinlock based txdesc list is still kept around, so we could have
a safe fallback.
No performance regression nor improvement is observed.
Reviewed by: adrian
Approved by: adrian (mentor)
MFC after: 1 week
Sponsored by: Microsoft OSTC
Differential Revision: https://reviews.freebsd.org/D5290
This paves the way for upcoming vRSS stuffs and eases more code cleanup.
Reviewed by: adrian
Approved by: adrian (mentor)
MFC after: 1 week
Sponsored by: Microsoft OSTC
Differential Revision: https://reviews.freebsd.org/D5283
Performance stays same; so no need to use fast taskqueue here.
Suggested by: royger
Reviewed by: adrian
Approved by: adrian (mentor)
MFC after: 1 week
Sponsored by: Microsoft OSTC
Differential Revision: https://reviews.freebsd.org/D5282
This also eases experiment on the non-fast taskqueue.
Reviewed by: adrian, Jun Su <junsu microsoft com>
Approved by: adrian (mentor)
MFC after: 1 week
Sponsored by: Microsoft OSTC
Differential Revision: https://reviews.freebsd.org/D5276
This paves the way for upcoming vRSS stuffs and eases more code cleanup.
Reviewed by: adrian
Approved by: adrian (mentor)
MFC after: 1 week
Sponsored by: Microsoft OSTC
Differential Revision: https://reviews.freebsd.org/D5275
And use SYSCTL+CTLFLAG_RDTUN for them.
Suggested by: adrian
Reviewed by: adrian, Hongjiang Zhang <honzhan microsoft com>
Approved by: adrian (mentor)
MFC after: 1 week
Sponsored by: Microsoft OSTC
Differential Revision: https://reviews.freebsd.org/D5274
This one gives the best performance so far.
Reviewed by: adrian
Approved by: adrian (mentor)
MFC after: 1 week
Sponsored by: Microsoft OSTC
Differential Revision: https://reviews.freebsd.org/D5273
It is off by default. This eases further experimenting on this driver.
Reviewed by: adrian
Approved by: adrian (mentor)
MFC after: 1 week
Sponsored by: Microsoft OSTC
Differential Revision: https://reviews.freebsd.org/D5272
Set TCP ACK append limit to 1, i.e. aggregate 2 ACKs at most. Aggregating
anything more than 2 hurts TCP sending performance in hyperv. This
significantly improves the TCP sending performance when the number of
concurrent connetion is low (2~8). And it greatly stabilizes the TCP
sending performance in other cases.
Set TCP data segments aggregation length limit to 37500. Without this
limitation, hn(4) could aggregate ~45 TCP data segments for each
connection (even at 64 or more connections) before dispatching them to
socket code; large aggregation slows down ACK sending and eventually
hurts/destabilizes TCP reception performance. This setting stabilizes
and improves TCP reception performance for >4 concurrent connections
significantly.
Make them sysctls so they could be adjusted.
Reviewed by: adrian, gallatin (previous version), hselasky (previous version)
Approved by: adrian (mentor)
MFC after: 1 week
Sponsored by: Microsoft OSTC
Differential Revision: https://reviews.freebsd.org/D5185
And convert rndis non-hot path spinlock to mutex.
Submitted by: Jun Su <junsu microsoft com>
Reviewed by: adrian, sephe
Approved by: adrian (mentor)
MFC after: 1 week
Sponsored by: Microsoft OSTC
Differential Revision: https://reviews.freebsd.org/D5081
It is off by default. This eases more experiment on hn(4).
Reviewed by: adrian, Hongjiang Zhang <honzhan microsoft com>
Approved by: adrian (mentor)
MFC after: 1 week
Sponsored by: Microsoft OSTC
Differential Revision: https://reviews.freebsd.org/D5175
This significantly increases LRO aggregation ratio when there are
large amount of connections (improves reception performance a lot).
Reviewed by: adrian
Approved by: adrian (mentor)
MFC after: 1 week
Sponsored by: Microsoft OSTC
Differential Revision: https://reviews.freebsd.org/D5167
hn(4) only has one RX ring currently, so default 8 LRO entries
are too small.
Reviewed by: adrian
Approved by: adrian (mentor)
MFC after: 1 week
Sponsored by: Microsoft OSTC
Differential Revision: https://reviews.freebsd.org/D5166
We lost half of the chimney sending space, because we mis-used
ffs() on a 64 bits mask, where ffsl() should be used.
While I'm here:
- Use system atomic operation instead.
- Stringent chimney sending index assertion.
Reviewed by: adrian
Approved by: adrian (mentor)
MFC after: 1 week
Sponsored by: Microsoft OSTC
Differential Revision: https://reviews.freebsd.org/D5159
It will be shared w/ upcoming ifnet.if_transmit implementaion.
No functional changes.
Reviewed by: adrian
Approved by: adrian (mentor)
MFC after: 1 week
Sponsored by: Microsoft OSTC
Differential Revision: https://reviews.freebsd.org/D5158
So that:
- TCP/IP stack will not do unnecessary IP header checksum for TSO
packets.
- Reduce guest load for non-TSO IP packets.
Reviewed by: adrian
Approved by: adrian (mentor)
MFC after: 1 week
Sponsored by: Microsoft OSTC
Differential Revision: https://reviews.freebsd.org/D5099
- For non-TSO offloading, we don't need to access mbuf to know
which csum offloading is requested, we can just use the
CSUM_{IP,TCP,UDP} in the csum_flags.
- For TSO offloading, we still can depend on CSUM_{TSO4,TSO6}
in the csum_flags to tell whether the TSO packet is an IPv4
TSO packet or an IPv6 TSO packet.
This streamlines csum offloading handling (remove the two goto)
and allows us the nuke the unnecessary get_transport_proto_type().
Reviewed by: adrian, Hongjiang Zhang <honzhan microsoft com>
Approved by: adrian (mentor)
MFC after: 1 week
Sponsored by: Microsoft OSTC
Differential Revision: https://reviews.freebsd.org/D5098
- Record csum features in softc, so we don't need to duplicate the
logic from attach path to ioctl path.
- Protect if_capenable and if_hwassist changes by main lock.
- Prefer turn on/off bits in if_hwassist explicitly instead of using
XOR.
Reviewed by: adrian, Hongjiang Zhang <honzhan microsoft com>
Approved by: adrian (mentor)
MFC after: 1 week
Sponsored by: Microsoft OSTC
Differential Revision: https://reviews.freebsd.org/D5085
- Avoid main lock contention by trylock for if_start, if that fails,
schedule TX taskqueue for if_start
- Don't do direct sending if the packet to be sent is large, e.g.
TSO packet.
This change gives me stable 9.1Gbps TCP sending performance w/ TSO
over a 10Gbe directly connected network (the performance fluctuated
between 4Gbps and 9Gbps before this commit). It also improves non-
TSO TCP sending performance a lot.
Reviewed by: adrian, royger
Approved by: adrian (mentor)
Sponsored by: Microsoft OSTC
Differential Revision: https://reviews.freebsd.org/D5074
According to all available information, VMSWITCH always does the
TCP segment checksum verification before sending the segment to
guest.
Reviewed by: adrian, delphij, Hongjiang Zhang <honzhan microsoft com>
Approved by: adrian (mentor)
Sponsored by: Microsoft OSTC
Differential Revision: https://reviews.freebsd.org/D4991
All used fields are setup one by one, so there is no need to zero
out this large struct.
While I'm here, move the stack variable near its usage.
Reviewed by: adrian, delphij, Jun Su <junsu microsoft com>
Approved by: adrian (mentor)
Sponsored by: Microsoft OSTC
Differential Revision: https://reviews.freebsd.org/D4978
While I'm here, move stack variables near their usage.
Reviewed by: adrian, delphij, Jun Su <junsu microsoft com>
Approved by: adrian (mentor)
Sponsored by: Microsoft OSTC
Differential Revision: https://reviews.freebsd.org/D4977
- Avoid unnecessary malloc/free on transmission path.
- busdma(9)-fy transmission path.
- Properly handle IFF_DRV_OACTIVE. This should fix the network
stalls reported by many.
- Properly setup TSO parameters.
- Properly handle bpf(4) tapping. This 5 times the performance
during TCP sending test, when there is one bpf(4) attached.
- Allow size of chimney sending be tuned on a running system.
Default value still needs more test to determine.
Reviewed by: adrian, delphij
Approved by: adrian (mentor)
Sponsored by: Microsoft OSTC
Differential Revision: https://reviews.freebsd.org/D4972
- Add optimizing LRO wrapper which pre-sorts all incoming packets
according to the hash type and flowid. This prevents exhaustion of
the LRO entries due to too many connections at the same time.
Testing using a larger number of higher bandwidth TCP connections
showed that the incoming ACK packet aggregation rate increased from
~1.3:1 to almost 3:1. Another test showed that for a number of TCP
connections greater than 16 per hardware receive ring, where 8 TCP
connections was the LRO active entry limit, there was a significant
improvement in throughput due to being able to fully aggregate more
than 8 TCP stream. For very few very high bandwidth TCP streams, the
optimizing LRO wrapper will add CPU usage instead of reducing CPU
usage. This is expected. Network drivers which want to use the
optimizing LRO wrapper needs to call "tcp_lro_queue_mbuf()" instead
of "tcp_lro_rx()" and "tcp_lro_flush_all()" instead of
"tcp_lro_flush()". Further the LRO control structure must be
initialized using "tcp_lro_init_args()" passing a non-zero number
into the "lro_mbufs" argument.
- Make LRO statistics 64-bit. Previously 32-bit integers were used for
statistics which can be prone to wrap-around. Fix this while at it
and update all SYSCTL's which expose LRO statistics.
- Ensure all data is freed when destroying a LRO control structures,
especially leftover LRO entries.
- Reduce number of memory allocations needed when setting up a LRO
control structure by precomputing the total amount of memory needed.
- Add own memory allocation counter for LRO.
- Bump the FreeBSD version to force recompilation of all KLDs due to
change of the LRO control structure size.
Sponsored by: Mellanox Technologies
Reviewed by: gallatin, sbruno, rrs, gnn, transport
Tested by: Netflix
Differential Revision: https://reviews.freebsd.org/D4914
If the NVSP protocol version is not greater than NVSP_PROTOCOL_VERSION_2,
then the recv buffer size is 15MB, otherwise the buffer size is 16MB.
Submitted by: Hongjiang Zhang <honzhan microsoft com>
Reviewed by: royger, Dexuan Cui <decui microsoft com>, adrian
Approved by: adrian (mentor)
Sponsored by: Microsoft OSTC
Differential Revision: https://reviews.freebsd.org/D4814
This one mainly avoids mbuf cluster allocation for TCP ACKs during
TCP sending tests. And it gives me ~200Mbps improvement (4.7Gbps
-> 4.9Gbps), when running iperf3 TCP sending test w/ 16 connections.
While I'm here, nuke the unnecessary zeroing out pkthdr.csum_flags.
Reviewed by: adrain
Approved by: adrian (mentor)
Sponsored by: Microsoft OSTC
Differential Revision: https://reviews.freebsd.org/D4853
Many applications and kernel modules (e.g. bridge) rely on the ifmedia
status report; give them what they want.
Submitted by: Dexuan Cui <decui microsoft com>
Reviewed by: Jun Su <junsu microsoftc com>, me, adrian
Modified by: me (minor)
Original differential: https://reviews.freebsd.org/D4611
Differential Revision: https://reviews.freebsd.org/D4852
Approved by: adrian (mentor)
Sponsored by: Microsoft OSTC
- Implement the LRO using tcp_lro APIs, and LRO is enabled by default.
- Add several stats sysctl nodes.
- Check IP/TCP length before sending the packet to tcp_lro_rx(), if host
does not provide RX csum information (*); and add an option through
sysctl to always trust host TCP segment csum checks (default is off).
- Add sysctl to control the LRO entry depth; it is disabled by default.
It is used to avoid holding too much TCP segments in driver. Limiting
the LRO entry depth helps a lot in a one/two streams RX test.
This one 3x the RX performance on my local test (3Gbps -> 10Gbps), and
~2x the RX performance over a directly connected 40Ge network (5Gbps ->
9Gbps).
(*) It seems the host stops supplying csum information, once the network
load is high. This still needs investigation...
Reviewed by: Hongjiang Zhang <honzhan microsoft com>,
Dexuan Cui <decui microsoft com>,
Jun Su <junsu microsoft com>,
delphij
Tested by: me (local),
Hongjiang Zhang <honzhan microsoft com>
(directly connected 40Ge)
Approved by: delphij (mentor), adrian (mentor, no objection)
With feedback from: delphij, Hongjiang Zhang <honzhan microsoft com>
Sponsored by: Microsoft OSTC
Differential Revision: https://reviews.freebsd.org/D4824
Windows Server 2012 and earlier hosts.
Submitted by: whu
Reviewed by: royger
Approved by: royger
MFC after: 3 days
Relnotes: No
Sponsored by: Microsoft OSTC
Differential Revision: https://reviews.freebsd.org/D3086
offload support was introduced in r284746.
While here also fix the ioctl() handler for IPv4 added in r279819,
which was never compiled in given opt_inet.h was not included.
- Bump link state when stopping or starting the interface;
- Don't handle SIOCGIFADDR specially, similar to r277103.
This change is based on a previous revision from Andy Zhang
(Microsoft) who did the diagnostic work and many thanks to
them for their help in supporting the HyperV work.
PR: kern/187203
MFC after: 2 weeks
it, except Ethernet, where it carried ng_ether(4) pointer.
For now carry the pointer in if_l2com directly.
Sponsored by: Netflix
Sponsored by: Nginx, Inc.
and holding a reference prior to calling further into the hyperv
stack.
Added missing FreeBSD idents.
Submitted by: Microsoft hyperv dev team
Approved by: re@ (gjb)
- change the SI_SUB_RUN_SCHEDULER sysinits in hv_utilc and
hv_netvsc_drv_freebsd.c to SI_SUB_KTHREAD_IDLE, since the
former is no longer in FreeBSD.
The use of these SYSINITs can probably be removed.
be pulled into FreeBSD. From now, FreeBSD will be considered the
upstream repo.
First step: move the drivers away from the contrib area and into
the base system.
A follow-on commit will include the drivers in the amd64 GENERIC kernel.
