On Tue, 2024-03-12 at 22:27 +0800, Wen Gu wrote:
> 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-#2: some prepare work for loopback-ism.
> - Patch #3-#7: implement loopback-ism device. Noted that loopback-ism now
> serves only SMC and no userspace interface exposed.
> - Patch #10-#15: 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) 81433 143938(+76.75%)
>
> 2. sockperf
>
> - serv: <smc_run> taskset -c <cpu> sockperf sr --tcp
> - clnt: <smc_run> taskset -c <cpu> sockperf { tp | pp } --tcp --msg-size={ 64000 for tp | 14 for pp } -i 127.0.0.1 -t 30
>
> TCP SMC-lo
> Bandwidth(MBps) 4903.07 7978.69(+62.73%)
> Latency(us) 6.095 3.539(-41.94%)
>
> 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 161665.67 244272.41(+51.10%)
>
> 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) 88790.23 117474.30(+32.31%)
> SET(Requests/s) 87508.20 118623.96(+35.57%)
>
>
> Change log:
>
> v3->v2:
> - 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@xxxxxxxxxxxxxxxxx/
> - 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@xxxxxxxxxxxxxxxxx/
> - 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.
>
## Form letter - net-next-closed
The merge window for v6.9 has begun and we have already posted our pull
request. Therefore net-next is closed for new drivers, features, code
refactoring and optimizations. We are currently accepting bug fixes
only.
Please repost when net-next reopens after March 25th.
RFC patches sent for review only are obviously welcome at any time.
See:
https://www.kernel.org/doc/html/next/process/maintainer-netdev.html#development-cycle
--
pw-bot: defer
