Age | Commit message (Collapse) | Author | Files | Lines |
|
Move from synchronous flushing of io and ctrl events from transports to
applications to an async model via a new session_input input node that
runs in interrupt mode. Events are coalesced per application worker.
On the one hand, this helps by minimizing message queue locking churn.
And on the other, it opens the possibility for further optimizations of
event message generation, obviates need for rx rescheduling rpcs and is
a first step towards a fully async data/io rx path.
Type: improvement
Signed-off-by: Florin Coras <fcoras@cisco.com>
Change-Id: Id6bebcb65fc9feef8aa02ddf1af6d9ba6f6745ce
|
|
App transports like TLS can close sessions on tx and consequently
generate new events. That can realloc the event pool.
Type: fix
Signed-off-by: Florin Coras <fcoras@cisco.com>
Change-Id: I58a97502acc1182b3d051ba1aa9e0e98c16f4593
|
|
Type: fix
Signed-off-by: Florin Coras <fcoras@cisco.com>
Change-Id: I35391fb8c39defbe0e57a241a357c3c98e8cef54
|
|
Type: improvement
Signed-off-by: Florin Coras <fcoras@cisco.com>
Signed-off-by: Filip Tehlar <ftehlar@cisco.com>
Change-Id: I7177ada23e5a69ec8e362ec98b98010c3b44b3d7
|
|
wrk->event_elts has 5 elements if no user events
Type: fix
Signed-off-by: Xiaoming Jiang <jiangxiaoming@outlook.com>
Change-Id: Ib38fab422304efc470e20ccb7121442f05bf8bf3
|
|
Type: feature
Signed-off-by: Filip Tehlar <ftehlar@cisco.com>
Change-Id: I02d9bb5292b32ffb1b2f05daccd8a7d5dba05125
|
|
Rename unused SESSION_IO_EVT_BUILTIN_TX to SESSION_IO_EVT_TX_MAIN and
leverage it for non-connected udp tx.
Non-connected udp sessions are listeners and are therefore allocated on
main thread. Consequently, whenever session queue node is not polling
main, tx events generated by external applications might be missed or
processed with some delay. To solve this, request that apps use
SESSION_IO_EVT_TX_MAIN tx events as opposed to SESSION_IO_EVT_TX and
send that to first worker as opposed to main.
Type: fix
Signed-off-by: Florin Coras <fcoras@cisco.com>
Change-Id: I5df5ac3dc80c0f192b2eefb1d465e9deefe8786b
|
|
Specify the number of max_batches when applying for dma config.
Skip this round when no batch available from vlib_dma_batch_new.
Type: improvement
Signed-off-by: Marvin Liu <yong.liu@intel.com>
Change-Id: Ic6e0acf81ba4fc3ed33aea6ac6990ef841021c59
|
|
Type: fix
Signed-off-by: Florin Coras <fcoras@cisco.com>
Change-Id: I70596ffcf90fa4cd57092584cb7a454f44208943
|
|
Allow apps/vcl to provide updated local ips for dgrams. In particular,
allow sessions bound to 0/0 to send data with valid local ips.
Type: improvement
Signed-off-by: Florin Coras <fcoras@cisco.com>
Change-Id: I50a086b1c252731a32a15b6a181ad3dba0c687e0
|
|
Type: fix
Signed-off-by: Florin Coras <fcoras@cisco.com>
Signed-off-by: Steven Luong <sluong@cisco.com>
Change-Id: Ibebe9e4ab7331c3ae66c9502e910368acaba51ec
|
|
Type: fix
Signed-off-by: Florin Coras <fcoras@cisco.com>
Change-Id: I1a3393b579caeb5dc25b60bc1b4a71706fb07051
|
|
Some upon apps(e.g. Nginx-quic) package it's several protocol buffers into a struct msg
which is a combination of gso_buffer and gso_size.
but if HostStack regardless the gso_size to the buffer and split the buffer with default mss,
that cause peer client failed on parsing the package.
Type: improvement
Signed-off-by: Dou Chao <chao.dou@intel.com>
Change-Id: I805eb642be826038ba96d1b85dad8ec0c0f6c459
Signed-off-by: Dou Chao <chao.dou@intel.com>
|
|
Type: improvement
Signed-off-by: Florin Coras <fcoras@cisco.com>
Change-Id: I035e3fdbb52eca010ad7b2c20ca2930cb1645978
|
|
To aid sesipon debug, add session event log in SM debug to track
the session state.
Type: improvement
Signed-off-by: Steven Luong <sluong@cisco.com>
Change-Id: I6909cf969cd5b6a3ea5a06d08ae32c2f1d48f686
|
|
- add session debug cli to enable fine control of which event logs
are enable/disable with below syntax
session debug {show | group <list> level <n>}
list may be entered with a dash, "0-4"
or it may be entered with a comma, "0,1,4"
- fix compilation errors when SESSION_EVT is enable
- change SESSION_EVT_FREE_HANDLER to use DEC_SESSION_ED instead of
DEC_SESSION_ETD because the transport may already be free when the
handler is called
Type: improvement
Signed-off-by: Steven Luong <sluong@cisco.com>
Change-Id: Iab2989e0a847bb59002ef16494eebcc1d112b2ae
|
|
add dma support to session, acclerate host-stack with dma
Type: feature
Signed-off-by: Marvin Liu <yong.liu@intel.com>
Signed-off-by: Junfeng Wang <drenfong.wang@intel.com>
Change-Id: I3d492921d69d9e3e0b34d33adc33fba3bde9e1cc
|
|
Type: fix
Signed-off-by: Florin Coras <fcoras@cisco.com>
Change-Id: I79bfe91e81983b619c61a32285d8e038c2654147
|
|
Fix race with connects for iperf3 udp test.
Type: fix
Signed-off-by: Florin Coras <fcoras@cisco.com>
Change-Id: Ief725b80047911e87ba24736dc0a60aa8bcdac50
|
|
Minimize amount of rpcs from first worker to main
Type: improvement
Signed-off-by: Florin Coras <fcoras@cisco.com>
Change-Id: I3794ff028a17d18b7bff69ede2b62e1e2d45ae77
|
|
Type: improvement
Signed-off-by: Florin Coras <fcoras@cisco.com>
Change-Id: I4fd7ae435514eb986543302c7e3e69e04acba8cf
|
|
Also reset send params flags before calling transports to avoid explicit
resets in all transports.
Type: improvement
Signed-off-by: Florin Coras <fcoras@cisco.com>
Change-Id: I1da7b3fab009728e7fee4199425ced933fa8a122
|
|
Instead of constantly scanning all transport vfts for update time
functions, build list at transport enable time.
Type: improvement
Signed-off-by: Florin Coras <fcoras@cisco.com>
Change-Id: Id5c07cc03ee1fdd072ebbbd40119d1a440a5e3b1
|
|
Type: improvement
Signed-off-by: Florin Coras <fcoras@cisco.com>
Change-Id: I1025cccd784f80b557847f69c3ea1ada5c9de60d
|
|
This ensures the scheduler always tracks sessions that are descheduled,
i.e., do not have events in the old io events list. When app retries to
send, clear descheduled flag and potentially the pacer.
Consequently, transports no longer need to reset the pacer when
sessions are rescheduled after a long app tx pause.
This also fixes a tcp bug whereby the pacer was reset too often when
snd_una was equal to snd_nxt as there was no way to distinguish betwen
app tx breaks and congestion.
Type: improvement
Signed-off-by: Florin Coras <fcoras@cisco.com>
Change-Id: Id3cc6c98cd76299e15030e504380dcf3c04c5189
|
|
- Always check session ownership
- Improve test for main thread rpc
Type: improvement
Signed-off-by: Florin Coras <fcoras@cisco.com>
Change-Id: I5fa60f7f5de199af0966987f9ce9a4cc8180cf98
|
|
Type: feature
Signed-off-by: Filip Tehlar <ftehlar@cisco.com>
Change-Id: I835675267c997b5dc92a0aaccdb58648bc786bb9
|
|
Type: fix
Signed-off-by: Filip Tehlar <ftehlar@cisco.com>
Change-Id: I6548274f8c0ae2a183b1d221cb195de445c2819f
|
|
vlib_buffer_enqueue_to_next() requires to allow overflow of up to 63
elements of 'buffer' and 'nexts' array.
- add helper to compute the minimum size
- fix occurences in session and async crypto
Type: fix
Change-Id: If8d7eebc5bf9beba71ba194aec0f79b8eb6d5843
Signed-off-by: Benoît Ganne <bganne@cisco.com>
|
|
The app may send packet to different remote via sendto() with
same socket (without calling connect()). In such scenario, remote
address need to be filled in.
Type: fix
Signed-off-by: liuyacan <liuyacan@corp.netease.com>
Change-Id: I49deb308668c05442f237e97acc9bfa239782482
|
|
CLIB_PREFETCH (s->tx_fifo, 2 * CLIB_CACHE_LINE_BYTES, LOAD);
sizeof(svm_fifo_t) is 128 bytes
Note on 64B cacheline size Arm machine,
CLIB_CACHE_LINE_BYTES 128
CLIB_CACHE_PREFETCH_BYTES 6
above CLIB_PREFETCH () macro will be expand to
__builtin_prefetch(s->tx_fifo)
__builtin_prefetch(s->tx_fifo + 64)
__builtin_prefetch(s->tx_fifo + 128) << prefetch out of range
__builtin_prefetch(s->tx_fifo + 192) << the same here
Solution:
Change to CLIB_PREFETCH (s->tx_fifo, sizeof (*(s->tx_fifo)), LOAD);
Type: fix
Signed-off-by: Tianyu Li <tianyu.li@arm.com>
Reviewed-by: Lijian Zhang <lijian.zhang@arm.com>
Change-Id: I745cbce3dbe5afcab53c39189d18392f569df5aa
|
|
Tcp may want to send acks in established, but the app
has not called accept() yet.
Type: improvement
Signed-off-by: liuyacan <liuyacan@corp.netease.com>
Change-Id: I43b8cd386e533ca95c8ec260a0a1f695ea140358
|
|
Type: fix
Signed-off-by: Florin Coras <fcoras@cisco.com>
Change-Id: I9301fbbcd611033b4b6ad5313edbc66840f5bb3a
|
|
Type: fix
Signed-off-by: Florin Coras <fcoras@cisco.com>
Change-Id: I56c4682aef59ed0e69073f9001341c425e65bd48
|
|
Type: fix
Signed-off-by: Florin Coras <fcoras@cisco.com>
Change-Id: Ibe32f7965f8cf457c39845713b029c8a4647ee55
|
|
Type: improvement
Signed-off-by: Florin Coras <fcoras@cisco.com>
Change-Id: Id91c3be57d49745cb3db6c768a8d5d14133f899e
|
|
Avoid grabbing the worker barrier if there's no work to be done.
Type: improvement
Signed-off-by: Florin Coras <fcoras@cisco.com>
Change-Id: Ice3de5df41cd1752aba3419ad2e2dd82f30e9bfb
|
|
Type: refactor
Signed-off-by: Florin Coras <fcoras@cisco.com>
Change-Id: I750c856ac81d951e8c0e62c710e0f35a0c80d6f9
|
|
Type: improvement
Signed-off-by: Florin Coras <fcoras@cisco.com>
Change-Id: Ie8a15c50531f3ccd5f91dbc0779e4d9c0d146844
|
|
Avoids dispatching ctrl events generated while handling the
current pending list.
Type: improvement
Signed-off-by: Florin Coras <fcoras@cisco.com>
Change-Id: Ibeaf901ba4cf58a68fbd88e5ec3c23f6c2f6f145
|
|
Type: refactor
Signed-off-by: Florin Coras <fcoras@cisco.com>
Change-Id: I10ec410fb7f3acb47128dda23510162dc13b20d0
|
|
Type: improvement
Signed-off-by: Florin Coras <fcoras@cisco.com>
Change-Id: Ifa47e1500e5cfb3c717f87b1d21131b9531c9005
|
|
Some app(e.g. Envoy) may call shutdown() instead of close() when
draining connection.
Type: improvement
Signed-off-by: liuyacan <liuyacan@corp.netease.com>
Change-Id: I9543b9ca3caa87b10b134fd1fc4019124e41e4d2
|
|
trajectory trace has been broken for a while because we used to save the
buffer trajectory in a vector pointed to in opaque2. This does not work
well when opaque2 is copied (eg. because of a clone) as 2 buffers end up
sharing the same vector.
This dedicates a full cacheline in the buffer metadata instead when
trajectory is compiled in. No dynamic allocation, no sharing, no tears.
Type: refactor
Change-Id: I6a028ca1b48d38f393a36979e5e452c2dd48ad3f
Signed-off-by: Benoît Ganne <bganne@cisco.com>
|
|
Type: fix
Signed-off-by: Florin Coras <fcoras@cisco.com>
Change-Id: I18eb1a71ac22017f75f835a23a57b023b46524ce
|
|
Pass tls/quic crypto context using extended config instead of bloating
conect/listen messages.
Type: refactor
Signed-off-by: Florin Coras <fcoras@cisco.com>
Change-Id: I0bc637ae310e6c31ef1e16847501dcb81453ee94
|
|
Type: feature
Signed-off-by: Florin Coras <fcoras@cisco.com>
Change-Id: Ic8d9386fef37ffd3446aaeb93a96ee6d60633831
|
|
Type: feature
Signed-off-by: Florin Coras <fcoras@cisco.com>
Change-Id: Ifdd6024daf044751895bb8d2deabad41d3a80c92
|
|
Type: fix
Signed-off-by: Florin Coras <fcoras@cisco.com>
Change-Id: Id2c06f49bcae1a26a2d16d58f9da7a4d3175bf3c
|
|
Experimental support for session layer interrupt mode. When enabled
(use-private-rx-mqs must be set) session queue node switches to
interrupt state when lightly loaded, i.e., no events and less than 1
vector/dispatch.
Because transport protocols require a periodic time update, when in
interrupt state the session queue node workers register a timerfd with
the unix-epoll-input node that when triggered signals, i.e., wakes up,
the queue node. Under light load, the timer is set to trigger every 1ms
whereas if no session is allocated, the worker moves to idle state and
the timeout is set to 100ms.
Type: feature
Signed-off-by: Florin Coras <fcoras@cisco.com>
Change-Id: I905b00777fbc025faf9c4074fce4c516cd139387
|