Merge branch 'net-smc-smc-intra-os-shortcut-with-loopback-ism'
authorPaolo Abeni <pabeni@redhat.com>
Tue, 30 Apr 2024 11:24:50 +0000 (13:24 +0200)
committerPaolo Abeni <pabeni@redhat.com>
Tue, 30 Apr 2024 11:24:51 +0000 (13:24 +0200)
commite458a9addfb296bf086c5e15ecbb2a8cf51e1b5e
tree2fe804316b776ea1fd62dd849951b5148013054f
parente5c5f3596de224422561d48eba6ece5210d967b3
parentc3a910f2380fe294d14e42af66af3d3eed8fecbf
Merge branch 'net-smc-smc-intra-os-shortcut-with-loopback-ism'

Wen Gu says:

====================
net/smc: SMC intra-OS shortcut with loopback-ism

This patch set acts as the second part of the new version of [1] (The first
part can be referred from [2]), the updated things of this version are listed
at the end.

- Background

SMC-D is now used in IBM z with ISM function to optimize network interconnect
for intra-CPC communications. Inspired by this, we try to make SMC-D available
on the non-s390 architecture through a software-implemented Emulated-ISM device,
that is the loopback-ism device here, to accelerate inter-process or
inter-containers communication within the same OS instance.

- Design

This patch set includes 3 parts:

 - Patch #1: some prepare work for loopback-ism.
 - Patch #2-#7: implement loopback-ism device and adapt SMC-D for it.
   loopback-ism now serves only SMC and no userspace interfaces exposed.
 - Patch #8-#11: memory copy optimization for intra-OS scenario.

The loopback-ism device is designed as an ISMv2 device and not be limited to
a specific net namespace, ends of both inter-process connection (1/1' in diagram
below) or inter-container connection (2/2' in diagram below) can find the same
available loopback-ism and choose it during the CLC handshake.

 Container 1 (ns1)                              Container 2 (ns2)
 +-----------------------------------------+    +-------------------------+
 | +-------+      +-------+      +-------+ |    |        +-------+        |
 | | App A |      | App B |      | App C | |    |        | App D |<-+     |
 | +-------+      +---^---+      +-------+ |    |        +-------+  |(2') |
 |     |127.0.0.1 (1')|             |192.168.0.11       192.168.0.12|     |
 |  (1)|   +--------+ | +--------+  |(2)   |    | +--------+   +--------+ |
 |     `-->|   lo   |-` |  eth0  |<-`      |    | |   lo   |   |  eth0  | |
 +---------+--|---^-+---+-----|--+---------+    +-+--------+---+-^------+-+
              |   |           |                                  |
 Kernel       |   |           |                                  |
 +----+-------v---+-----------v----------------------------------+---+----+
 |    |                            TCP                               |    |
 |    |                                                              |    |
 |    +--------------------------------------------------------------+    |
 |                                                                        |
 |                           +--------------+                             |
 |                           | smc loopback |                             |
 +---------------------------+--------------+-----------------------------+

loopback-ism device creates DMBs (shared memory) for each connection peer.
Since data transfer occurs within the same kernel, the sndbuf of each peer
is only a descriptor and point to the same memory region as peer DMB, so that
the data copy from sndbuf to peer DMB can be avoided in loopback-ism case.

 Container 1 (ns1)                              Container 2 (ns2)
 +-----------------------------------------+    +-------------------------+
 | +-------+                               |    |        +-------+        |
 | | App C |-----+                         |    |        | App D |        |
 | +-------+     |                         |    |        +-^-----+        |
 |               |                         |    |          |              |
 |           (2) |                         |    |     (2') |              |
 |               |                         |    |          |              |
 +---------------|-------------------------+    +----------|--------------+
                 |                                         |
 Kernel          |                                         |
 +---------------|-----------------------------------------|--------------+
 | +--------+ +--v-----+                           +--------+ +--------+  |
 | |dmb_desc| |snd_desc|                           |dmb_desc| |snd_desc|  |
 | +-----|--+ +--|-----+                           +-----|--+ +--------+  |
 | +-----|--+    |                                 +-----|--+             |
 | | DMB C  |    +---------------------------------| DMB D  |             |
 | +--------+                                      +--------+             |
 |                                                                        |
 |                           +--------------+                             |
 |                           | smc loopback |                             |
 +---------------------------+--------------+-----------------------------+

- Benchmark Test

 * Test environments:
      - VM with Intel Xeon Platinum 8 core 2.50GHz, 16 GiB mem.
      - SMC sndbuf/DMB size 1MB.

 * Test object:
      - TCP: run on TCP loopback.
      - SMC lo: run on SMC loopback-ism.

1. ipc-benchmark (see [3])

 - ./<foo> -c 1000000 -s 100

                            TCP                  SMC-lo
Message
rate (msg/s)              84991                  151293(+78.01%)

2. sockperf

 - serv: <smc_run> sockperf sr --tcp
 - clnt: <smc_run> sockperf { tp | pp } --tcp --msg-size={ 64000 for tp | 14 for pp } -i 127.0.0.1 -t 30

                            TCP                  SMC-lo
Bandwidth(MBps)        5033.569                7987.732(+58.69%)
Latency(us)               5.986                   3.398(-43.23%)

3. nginx/wrk

 - serv: <smc_run> nginx
 - clnt: <smc_run> wrk -t 8 -c 1000 -d 30 http://127.0.0.1:80

                           TCP                   SMC-lo
Requests/s           187951.76                267107.90(+42.12%)

4. redis-benchmark

 - serv: <smc_run> redis-server
 - clnt: <smc_run> redis-benchmark -h 127.0.0.1 -q -t set,get -n 400000 -c 200 -d 1024

                           TCP                   SMC-lo
GET(Requests/s)       86132.64                118133.49(+37.15%)
SET(Requests/s)       87374.40                122887.86(+40.65%)

Change log:
v7->v6
- Patch #2: minor: remove unnecessary 'return' of inline smc_loopback_exit().
- Patch #10: minor: directly return 0 instead of 'rc' in smcd_cdc_msg_send().
- all: collect the Reviewed-by tags.

v6->RFC v5
Link: https://lore.kernel.org/netdev/20240414040304.54255-1-guwen@linux.alibaba.com/
- Patch #2: make the use of CONFIG_SMC_LO cleaner.
- Patch #5: mark some smcd_ops that loopback-ism doesn't support as
  optional and check for the support when they are called.
- Patch #7: keep loopback-ism at the beginning of the SMC-D device list.
- Some expression changes in commit logs and comments.

RFC v5->RFC v4:
Link: https://lore.kernel.org/netdev/20240324135522.108564-1-guwen@linux.alibaba.com/
- Patch #2: minor changes in description of config SMC_LO and comments.
- Patch #10: minor changes in comments and if(smc_ism_support_dmb_nocopy())
  check in smcd_cdc_msg_send().
- Patch #3: change smc_lo_generate_id() to smc_lo_generate_ids() and SMC_LO_CHID
  to SMC_LO_RESERVED_CHID.
- Patch #5: memcpy while holding the ldev->dmb_ht_lock.
- Some expression changes in commit logs.

RFC v4->v3:
Link: https://lore.kernel.org/netdev/20240317100545.96663-1-guwen@linux.alibaba.com/
- The merge window of v6.9 is open, so post this series as an RFC.
- Patch #6: since some information fed back by smc_nl_handle_smcd_dev() dose
  not apply to Emulated-ISM (including loopback-ism here), loopback-ism is
  not exposed through smc netlink for the time being. we may refactor this
  part when smc netlink interface is updated.

v3->v2:
Link: https://lore.kernel.org/netdev/20240312142743.41406-1-guwen@linux.alibaba.com/
- Patch #11: use tasklet_schedule(&conn->rx_tsklet) instead of smcd_cdc_rx_handler()
  to avoid possible recursive locking of conn->send_lock and use {read|write}_lock_bh()
  to acquire dmb_ht_lock.

v2->v1:
Link: https://lore.kernel.org/netdev/20240307095536.29648-1-guwen@linux.alibaba.com/
- All the patches: changed the term virtual-ISM to Emulated-ISM as defined by SMCv2.1.
- Patch #3: optimized the description of SMC_LO config. Avoid exposing loopback-ism
  to sysfs and remove all the knobs until future definition clear.
- Patch #3: try to make lockdep happy by using read_lock_bh() in smc_lo_move_data().
- Patch #6: defaultly use physical contiguous DMB buffers.
- Patch #11: defaultly enable DMB no-copy for loopback-ism and free the DMB in
  unregister_dmb or detach_dmb when dmb_node->refcnt reaches 0, instead of using
  wait_event to keep waiting in unregister_dmb.

v1->RFC:
Link: https://lore.kernel.org/netdev/20240111120036.109903-1-guwen@linux.alibaba.com/
- Patch #9: merge rx_bytes and tx_bytes as xfer_bytes statistics:
  /sys/devices/virtual/smc/loopback-ism/xfer_bytes
- Patch #10: add support_dmb_nocopy operation to check if SMC-D device supports
  merging sndbuf with peer DMB.
- Patch #13 & #14: introduce loopback-ism device control of DMB memory type and
  control of whether to merge sndbuf and DMB. They can be respectively set by:
  /sys/devices/virtual/smc/loopback-ism/dmb_type
  /sys/devices/virtual/smc/loopback-ism/dmb_copy
  The motivation for these two control is that a performance bottleneck was
  found when using vzalloced DMB and sndbuf is merged with DMB, and there are
  many CPUs and CONFIG_HARDENED_USERCOPY is set [4]. The bottleneck is caused
  by the lock contention in vmap_area_lock [5] which is involved in memcpy_from_msg()
  or memcpy_to_msg(). Currently, Uladzislau Rezki is working on mitigating the
  vmap lock contention [6]. It has significant effects, but using virtual memory
  still has additional overhead compared to using physical memory.
  So this new version provides controls of dmb_type and dmb_copy to suit
  different scenarios.
- Some minor changes and comments improvements.

RFC->old version([1]):
Link: https://lore.kernel.org/netdev/1702214654-32069-1-git-send-email-guwen@linux.alibaba.com/
- Patch #1: improve the loopback-ism dump, it shows as follows now:
  # smcd d
  FID  Type  PCI-ID        PCHID  InUse  #LGs  PNET-ID
  0000 0     loopback-ism  ffff   No        0
- Patch #3: introduce the smc_ism_set_v2_capable() helper and set
  smc_ism_v2_capable when ISMv2 or virtual ISM is registered,
  regardless of whether there is already a device in smcd device list.
- Patch #3: loopback-ism will be added into /sys/devices/virtual/smc/loopback-ism/.
- Patch #8: introduce the runtime switch /sys/devices/virtual/smc/loopback-ism/active
  to activate or deactivate the loopback-ism.
- Patch #9: introduce the statistics of loopback-ism by
  /sys/devices/virtual/smc/loopback-ism/{{tx|rx}_tytes|dmbs_cnt}.
- Some minor changes and comments improvements.

[1] https://lore.kernel.org/netdev/1695568613-125057-1-git-send-email-guwen@linux.alibaba.com/
[2] https://lore.kernel.org/netdev/20231219142616.80697-1-guwen@linux.alibaba.com/
[3] https://github.com/goldsborough/ipc-bench
[4] https://lore.kernel.org/all/3189e342-c38f-6076-b730-19a6efd732a5@linux.alibaba.com/
[5] https://lore.kernel.org/all/238e63cd-e0e8-4fbf-852f-bc4d5bc35d5a@linux.alibaba.com/
[6] https://lore.kernel.org/all/20240102184633.748113-1-urezki@gmail.com/
====================

Link: https://lore.kernel.org/r/20240428060738.60843-1-guwen@linux.alibaba.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>