- 13 Sep, 2024 8 commits
-
-
Jakub Kicinski authored
Willem de Bruijn says: ==================== selftests/net: packetdrill: netns and two imports From: Willem de Bruijn <willemb@google.com> 1/3: run in nets, as discussed, and add missing CONFIGs 2/3: import tcp/zerocopy 3/3: import tcp/slow_start ==================== Link: https://patch.msgid.link/20240912005317.1253001-1-willemdebruijn.kernel@gmail.comSigned-off-by:
Jakub Kicinski <kuba@kernel.org>
-
Willem de Bruijn authored
Same import process as previous tests. Also add CONFIG_NET_SCH_FQ to config, as one test uses that. Same test process as previous tests. Both with and without debug mode. Recording the steps once: make mrproper vng --build \ --config tools/testing/selftests/net/packetdrill/config \ --config kernel/configs/debug.config vng -v --run . --user root --cpus 4 -- \ make -C tools/testing/selftests TARGETS=net/packetdrill run_tests Link: https://github.com/linux-netdev/nipa/wiki/How-to-run-netdev-selftests-CI-style#how-to-buildSigned-off-by:Willem de Bruijn <willemb@google.com> Acked-by:
Matthieu Baerts (NGI0) <matttbe@kernel.org> Link: https://patch.msgid.link/20240912005317.1253001-4-willemdebruijn.kernel@gmail.comSigned-off-by:
-
Willem de Bruijn authored
Same as initial tests, import verbatim from github.com/google/packetdrill, aside from: - update `source ./defaults.sh` path to adjust for flat dir - add SPDX headers - remove author statements if any - drop blank lines at EOF (new) Also import set_sysctls.py, which many scripts depend on to set sysctls and then restore them later. This is no longer strictly needed for namespacified sysctl. But not all sysctls are namespacified, and doesn't hurt if they are. Signed-off-by:
-
Willem de Bruijn authored
Run packetdrill tests inside netns. They may change system settings, such as sysctl. Also expand config with a few more needed CONFIGs. Link: https://lore.kernel.org/netdev/20240910152640.429920be@kernel.org/Signed-off-by:
-
Jakub Kicinski authored
David Arinzon says: ==================== ENA driver metrics changes This patchset contains an introduction of new metrics available to ENA users. ==================== Link: https://patch.msgid.link/20240909084704.13856-1-darinzon@amazon.comSigned-off-by:
-
David Arinzon authored
ENA currently supports the following customer metrics: - `bw_in_allowance_exceeded` - `bw_out_allowance_exceeded` - `conntrack_allowance_exceeded` - `linklocal_allowance_exceeded` - `pps_allowance_exceeded` This patch adds a new metric named: `conntrack_allowance_available`. Information about these metrics is available in [1]. In addition, the interface between the driver and the device has been upgraded to allow more flexibility and expendability to additional metrics in the future. [1]: https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/monitoring-network-performance-ena.html#network-performance-metricsSigned-off-by:
Ron Beider <rbeider@amazon.com> Signed-off-by:
Shahar Itzko <itzko@amazon.com> Signed-off-by:
David Arinzon <darinzon@amazon.com> Link: https://patch.msgid.link/20240909084704.13856-3-darinzon@amazon.comSigned-off-by:
-
David Arinzon authored
ENA Express metrics, called `ena_srd` are exposed to customers via `ethtool`. The metrics allow customers to check the configuration (mode), tx/rx counters as well as resource utilization. The documentation is also updated to provide a general explanation about ENA Express as well as links for further information about metrics and configurations. Signed-off-by:
Igor Chauskin <igorch@amazon.com> Signed-off-by:
-
git://git.kernel.org/pub/scm/linux/kernel/git/netdev/netJakub Kicinski authored
Cross-merge networking fixes after downstream PR. No conflicts (sort of) and no adjacent changes. This merge reverts commit b3c9e65e ("net: hsr: remove seqnr_lock") from net, as it was superseded by commit 430d67bd ("net: hsr: Use the seqnr lock for frames received via interlink port.") in net-next. Signed-off-by:
-
- 12 Sep, 2024 32 commits
-
-
git://git.kernel.org/pub/scm/linux/kernel/git/netdev/netLinus Torvalds authored
Pull networking fixes from Paolo Abeni: "Including fixes from netfilter. There is a recently notified BT regression with no fix yet. I do not think a fix will land in the next week. Current release - regressions: - core: tighten bad gso csum offset check in virtio_net_hdr - netfilter: move nf flowtable bpf initialization in nf_flow_table_module_init() - eth: ice: stop calling pci_disable_device() as we use pcim - eth: fou: fix null-ptr-deref in GRO. Current release - new code bugs: - hsr: prevent NULL pointer dereference in hsr_proxy_announce() Previous releases - regressions: - hsr: remove seqnr_lock - netfilter: nft_socket: fix sk refcount leaks - mptcp: pm: fix uaf in __timer_delete_sync - phy: dp83822: fix NULL pointer dereference on DP83825 devices - eth: revert "virtio_net: rx enable premapped mode by default" - eth: octeontx2-af: Modify SMQ flush sequence to drop packets Previous releases - always broken: - eth: mlx5: fix bridge mode operations when there are no VFs - eth: igb: Always call igb_xdp_ring_update_tail() under Tx lock" * tag 'net-6.11-rc8' of git://git.kernel.org/pub/scm/linux/kernel/git/netdev/net: (36 commits) net: netfilter: move nf flowtable bpf initialization in nf_flow_table_module_init() net: tighten bad gso csum offset check in virtio_net_hdr netlink: specs: mptcp: fix port endianness net: dpaa: Pad packets to ETH_ZLEN mptcp: pm: Fix uaf in __timer_delete_sync net: libwx: fix number of Rx and Tx descriptors net: dsa: felix: ignore pending status of TAS module when it's disabled net: hsr: prevent NULL pointer dereference in hsr_proxy_announce() selftests: mptcp: include net_helper.sh file selftests: mptcp: include lib.sh file selftests: mptcp: join: restrict fullmesh endp on 1st sf netfilter: nft_socket: make cgroupsv2 matching work with namespaces netfilter: nft_socket: fix sk refcount leaks MAINTAINERS: Add ethtool pse-pd to PSE NETWORK DRIVER dt-bindings: net: tja11xx: fix the broken binding selftests: net: csum: Fix checksums for packets with non-zero padding net: phy: dp83822: Fix NULL pointer dereference on DP83825 devices virtio_net: disable premapped mode by default Revert "virtio_net: big mode skip the unmap check" Revert "virtio_net: rx remove premapped failover code" ... -
Linus Torvalds authored
Merge tag 'platform-drivers-x86-v6.11-7' of git://git.kernel.org/pub/scm/linux/kernel/git/pdx86/platform-drivers-x86 Pull x86 platform driver fixes from Ilpo Järvinen: - asus-wmi: Disable OOBE that interferes with backlight control - panasonic-laptop: Two fixes to SINF array handling * tag 'platform-drivers-x86-v6.11-7' of git://git.kernel.org/pub/scm/linux/kernel/git/pdx86/platform-drivers-x86: platform/x86: asus-wmi: Disable OOBE experience on Zenbook S 16 platform/x86: panasonic-laptop: Allocate 1 entry extra in the sinf array platform/x86: panasonic-laptop: Fix SINF array out of bounds accesses
-
Linus Torvalds authored
As Jann points out, PFN mappings are special, because unlike normal memory mappings, there is no lifetime information associated with the mapping - it is just a raw mapping of PFNs with no reference counting of a 'struct page'. That's all very much intentional, but it does mean that it's easy to mess up the cleanup in case of errors. Yes, a failed mmap() will always eventually clean up any partial mappings, but without any explicit lifetime in the page table mapping itself, it's very easy to do the error handling in the wrong order. In particular, it's easy to mistakenly free the physical backing store before the page tables are actually cleaned up and (temporarily) have stale dangling PTE entries. To make this situation less error-prone, just make sure that any partial pfn mapping is torn down early, before any other error handling. Reported-and-tested-by:
Jann Horn <jannh@google.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Jason Gunthorpe <jgg@ziepe.ca> Cc: Simona Vetter <simona.vetter@ffwll.ch> Signed-off-by:
Linus Torvalds <torvalds@linux-foundation.org>
-
Lorenzo Bianconi authored
Move nf flowtable bpf initialization in nf_flow_table module load routine since nf_flow_table_bpf is part of nf_flow_table module and not nf_flow_table_inet one. This patch allows to avoid the following kernel warning running the reproducer below: $modprobe nf_flow_table_inet $rmmod nf_flow_table_inet $modprobe nf_flow_table_inet modprobe: ERROR: could not insert 'nf_flow_table_inet': Invalid argument [ 184.081501] ------------[ cut here ]------------ [ 184.081527] WARNING: CPU: 0 PID: 1362 at kernel/bpf/btf.c:8206 btf_populate_kfunc_set+0x23c/0x330 [ 184.081550] CPU: 0 UID: 0 PID: 1362 Comm: modprobe Kdump: loaded Not tainted 6.11.0-0.rc5.22.el10.x86_64 #1 [ 184.081553] Hardware name: Red Hat OpenStack Compute, BIOS 1.14.0-1.module+el8.4.0+8855+a9e237a9 04/01/2014 [ 184.081554] RIP: 0010:btf_populate_kfunc_set+0x23c/0x330 [ 184.081558] RSP: 0018:ff22cfb38071fc90 EFLAGS: 00010202 [ 184.081559] RAX: 0000000000000001 RBX: 0000000000000001 RCX: 0000000000000000 [ 184.081560] RDX: 000000000000006e RSI: ffffffff95c00000 RDI: ff13805543436350 [ 184.081561] RBP: ffffffffc0e22180 R08: ff13805543410808 R09: 000000000001ec00 [ 184.081562] R10: ff13805541c8113c R11: 0000000000000010 R12: ff13805541b83c00 [ 184.081563] R13: ff13805543410800 R14: 0000000000000001 R15: ffffffffc0e2259a [ 184.081564] FS: 00007fa436c46740(0000) GS:ff1380557ba00000(0000) knlGS:0000000000000000 [ 184.081569] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 184.081570] CR2: 000055e7b3187000 CR3: 0000000100c48003 CR4: 0000000000771ef0 [ 184.081571] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 [ 184.081572] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 [ 184.081572] PKRU: 55555554 [ 184.081574] Call Trace: [ 184.081575] <TASK> [ 184.081578] ? show_trace_log_lvl+0x1b0/0x2f0 [ 184.081580] ? show_trace_log_lvl+0x1b0/0x2f0 [ 184.081582] ? __register_btf_kfunc_id_set+0x199/0x200 [ 184.081585] ? btf_populate_kfunc_set+0x23c/0x330 [ 184.081586] ? __warn.cold+0x93/0xed [ 184.081590] ? btf_populate_kfunc_set+0x23c/0x330 [ 184.081592] ? report_bug+0xff/0x140 [ 184.081594] ? handle_bug+0x3a/0x70 [ 184.081596] ? exc_invalid_op+0x17/0x70 [ 184.081597] ? asm_exc_invalid_op+0x1a/0x20 [ 184.081601] ? btf_populate_kfunc_set+0x23c/0x330 [ 184.081602] __register_btf_kfunc_id_set+0x199/0x200 [ 184.081605] ? __pfx_nf_flow_inet_module_init+0x10/0x10 [nf_flow_table_inet] [ 184.081607] do_one_initcall+0x58/0x300 [ 184.081611] do_init_module+0x60/0x230 [ 184.081614] __do_sys_init_module+0x17a/0x1b0 [ 184.081617] do_syscall_64+0x7d/0x160 [ 184.081620] ? __count_memcg_events+0x58/0xf0 [ 184.081623] ? handle_mm_fault+0x234/0x350 [ 184.081626] ? do_user_addr_fault+0x347/0x640 [ 184.081630] ? clear_bhb_loop+0x25/0x80 [ 184.081633] ? clear_bhb_loop+0x25/0x80 [ 184.081634] ? clear_bhb_loop+0x25/0x80 [ 184.081637] entry_SYSCALL_64_after_hwframe+0x76/0x7e [ 184.081639] RIP: 0033:0x7fa43652e4ce [ 184.081647] RSP: 002b:00007ffe8213be18 EFLAGS: 00000246 ORIG_RAX: 00000000000000af [ 184.081649] RAX: ffffffffffffffda RBX: 000055e7b3176c20 RCX: 00007fa43652e4ce [ 184.081650] RDX: 000055e7737fde79 RSI: 0000000000003990 RDI: 000055e7b3185380 [ 184.081651] RBP: 000055e7737fde79 R08: 0000000000000007 R09: 000055e7b3179bd0 [ 184.081651] R10: 0000000000000001 R11: 0000000000000246 R12: 0000000000040000 [ 184.081652] R13: 000055e7b3176fa0 R14: 0000000000000000 R15: 000055e7b3179b80 Fixes: 391bb659 ("netfilter: Add bpf_xdp_flow_lookup kfunc") Signed-off-by:
Lorenzo Bianconi <lorenzo@kernel.org> Acked-by:
Florian Westphal <fw@strlen.de> Acked-by:
Pablo Neira Ayuso <pablo@netfilter.org> Link: https://patch.msgid.link/20240911-nf-flowtable-bpf-modprob-fix-v1-1-f9fc075aafc3@kernel.orgSigned-off-by:
Paolo Abeni <pabeni@redhat.com>
-
git://git.kernel.org/pub/scm/linux/kernel/git/netfilter/nfPaolo Abeni authored
Pablo Neira Ayuso says: ==================== Netfilter fixes for net The following batch contains two fixes from Florian Westphal: Patch #1 fixes a sk refcount leak in nft_socket on mismatch. Patch #2 fixes cgroupsv2 matching from containers due to incorrect level in subtree. netfilter pull request 24-09-12 * tag 'nf-24-09-12' of git://git.kernel.org/pub/scm/linux/kernel/git/netfilter/nf: netfilter: nft_socket: make cgroupsv2 matching work with namespaces netfilter: nft_socket: fix sk refcount leaks ==================== Link: https://patch.msgid.link/20240911222520.3606-1-pablo@netfilter.orgSigned-off-by:
-
Jakub Kicinski authored
Parthiban Veerasooran says: ==================== Add support for OPEN Alliance 10BASE-T1x MACPHY Serial Interface This patch series contain the below updates, - Adds support for OPEN Alliance 10BASE-T1x MACPHY Serial Interface in the net/ethernet/oa_tc6.c. Link to the spec: ----------------- https://opensig.org/download/document/OPEN_Alliance_10BASET1x_MAC-PHY_Serial_Interface_V1.1.pdf - Adds driver support for Microchip LAN8650/1 Rev.B1 10BASE-T1S MACPHY Ethernet driver in the net/ethernet/microchip/lan865x/lan865x.c. Link to the product: -------------------- https://www.microchip.com/en-us/product/lan8650 Testing Details: ---------------- The driver performance was tested using iperf3 in the below two setups separately. Setup 1: -------- Node 0 - Raspberry Pi 4 with LAN8650 MAC-PHY Node 1 - Raspberry Pi 4 with EVB-LAN8670-USB USB Stick Setup 2: -------- Node 0 - SAMA7G54-EK with LAN8650 MAC-PHY Node 1 - Raspberry Pi 4 with EVB-LAN8670-USB USB Stick Achieved maximum of 9.4 Mbps. Some systems like Raspberry Pi 4 need performance mode enabled to get the proper clock speed for SPI. Refer below link for more details. https://github.com/raspberrypi/linux/issues/3381#issuecomment-1144723750 ==================== Link: https://patch.msgid.link/20240909082514.262942-1-Parthiban.Veerasooran@microchip.comSigned-off-by:
-
Parthiban Veerasooran authored
The LAN8650/1 combines a Media Access Controller (MAC) and an Ethernet PHY to enable 10BASE-T1S networks. The Ethernet Media Access Controller (MAC) module implements a 10 Mbps half duplex Ethernet MAC, compatible with the IEEE 802.3 standard and a 10BASE-T1S physical layer transceiver integrated into the LAN8650/1. The communication between the Host and the MAC-PHY is specified in the OPEN Alliance 10BASE-T1x MACPHY Serial Interface (TC6). Reviewed-by:
Conor Dooley <conor.dooley@microchip.com> Reviewed-by:
Andrew Lunn <andrew@lunn.ch> Signed-off-by:
Parthiban Veerasooran <Parthiban.Veerasooran@microchip.com> Link: https://patch.msgid.link/20240909082514.262942-15-Parthiban.Veerasooran@microchip.comSigned-off-by:
-
Parthiban Veerasooran authored
The LAN8650/1 is designed to conform to the OPEN Alliance 10BASE-T1x MAC-PHY Serial Interface specification, Version 1.1. The IEEE Clause 4 MAC integration provides the low pin count standard SPI interface to any microcontroller therefore providing Ethernet functionality without requiring MAC integration within the microcontroller. The LAN8650/1 operates as an SPI client supporting SCLK clock rates up to a maximum of 25 MHz. This SPI interface supports the transfer of both data (Ethernet frames) and control (register access). By default, the chunk data payload is 64 bytes in size. The Ethernet Media Access Controller (MAC) module implements a 10 Mbps half duplex Ethernet MAC, compatible with the IEEE 802.3 standard. 10BASE-T1S physical layer transceiver integrated is into the LAN8650/1. The PHY and MAC are connected via an internal Media Independent Interface (MII). Reviewed-by:
-
Parthiban Veerasooran authored
Zero align receive frame feature can be enabled to align all receive ethernet frames data to start at the beginning of any receive data chunk payload with a start word offset (SWO) of zero. Receive frames may begin anywhere within the receive data chunk payload when this feature is not enabled. Reviewed-by:
-
Parthiban Veerasooran authored
The MAC-PHY interrupt is asserted when the following conditions are met. Receive chunks available - This interrupt is asserted when the previous data footer had no receive data chunks available and once the receive data chunks become available for reading. On reception of the first data header this interrupt will be deasserted. Transmit chunk credits available - This interrupt is asserted when the previous data footer indicated no transmit credits available and once the transmit credits become available for transmitting transmit data chunks. On reception of the first data header this interrupt will be deasserted. Extended status event - This interrupt is asserted when the previous data footer indicated no extended status and once the extended event become available. In this case the host should read status #0 register to know the corresponding error/event. On reception of the first data header this interrupt will be deasserted. Reviewed-by:
-
Parthiban Veerasooran authored
SPI rx data buffer can contain one or more receive data chunks. A receive data chunk consists a 64 bytes receive data chunk payload followed a 4 bytes data footer at the end. The data footer contains the information needed to determine the validity and location of the receive frame data within the receive data chunk payload and the host can use these information to generate ethernet frame. Initially the receive chunks available will be updated from the buffer status register and then it will be updated from the footer received on each spi data transfer. Tx data valid or empty chunks equal to the number receive chunks available will be transmitted in the MOSI to receive all the rx chunks. Additionally the receive data footer contains the below information as well. The received footer will be examined for the receive errors if any. Reviewed-by:
-
Parthiban Veerasooran authored
The transmit ethernet frame will be converted into multiple transmit data chunks. Each transmit data chunk consists of a 4 bytes header followed by a 64 bytes transmit data chunk payload. The 4 bytes data header occurs at the beginning of each transmit data chunk on MOSI. The data header contains the information needed to determine the validity and location of the transmit frame data within the data chunk payload. The number of transmit data chunks transmitted to mac-phy is limited to the number transmit credits available in the mac-phy. Initially the transmit credits will be updated from the buffer status register and then it will be updated from the footer received on each spi data transfer. The received footer will be examined for the transmit errors if any. Reviewed-by:
-
Parthiban Veerasooran authored
Enabling Configuration Synchronization bit (SYNC) in the Configuration Register #0 enables data communication in the MAC-PHY. The state of this bit is reflected in the data footer SYNC bit. Reviewed-by:
-
Parthiban Veerasooran authored
This patch adds c45 registers direct access support in Microchip's LAN865x internal PHY. OPEN Alliance 10BASE-T1x compliance MAC-PHYs will have both C22 and C45 registers space. If the PHY is discovered via C22 bus protocol it assumes it uses C22 protocol and always uses C22 registers indirect access to access C45 registers. This is because, we don't have a clean separation between C22/C45 register space and C22/C45 MDIO bus protocols. Resulting, PHY C45 registers direct access can't be used which can save multiple SPI bus access. To support this feature, set .read_mmd/.write_mmd in the PHY driver to call .read_c45/.write_c45 in the OPEN Alliance framework drivers/net/ethernet/oa_tc6.c Reviewed-by:
-
Parthiban Veerasooran authored
Internal PHY is initialized as per the PHY register capability supported by the MAC-PHY. Direct PHY Register Access Capability indicates if PHY registers are directly accessible within the SPI register memory space. Indirect PHY Register Access Capability indicates if PHY registers are indirectly accessible through the MDIO/MDC registers MDIOACCn defined in OPEN Alliance specification. Currently the direct register access is only supported. Reviewed-by:
-
Parthiban Veerasooran authored
This will unmask the following error interrupts from the MAC-PHY. tx protocol error rx buffer overflow error loss of framing error header error The MAC-PHY will signal an error by setting the EXST bit in the receive data footer which will then allow the host to read the STATUS0 register to find the source of the error. Reviewed-by:
-
Parthiban Veerasooran authored
Reset complete bit is set when the MAC-PHY reset completes and ready for configuration. Additionally reset complete bit in the STS0 register has to be written by one upon reset complete to clear the interrupt. Reviewed-by:
-
Parthiban Veerasooran authored
Implement register read operation according to the control communication specified in the OPEN Alliance 10BASE-T1x MACPHY Serial Interface document. Control read commands are used by the SPI host to read registers within the MAC-PHY. Each control read commands are composed of a 32 bits control command header. The MAC-PHY ignores all data from the SPI host following the control header for the remainder of the control read command. Control read commands can read either a single register or multiple consecutive registers. When multiple consecutive registers are read, the address is automatically post-incremented by the MAC-PHY. Reading any unimplemented or undefined registers shall return zero. Reviewed-by:
-
Parthiban Veerasooran authored
Implement register write operation according to the control communication specified in the OPEN Alliance 10BASE-T1x MACPHY Serial Interface document. Control write commands are used by the SPI host to write registers within the MAC-PHY. Each control write commands are composed of a 32 bits control command header followed by register write data. The MAC-PHY ignores the final 32 bits of data from the SPI host at the end of the control write command. The write command and data is also echoed from the MAC-PHY back to the SPI host to enable the SPI host to identify which register write failed in the case of any bus errors. Control write commands can write either a single register or multiple consecutive registers. When multiple consecutive registers are written, the address is automatically post-incremented by the MAC-PHY. Writing to any unimplemented or undefined registers shall be ignored and yield no effect. Reviewed-by:
-
Parthiban Veerasooran authored
The IEEE 802.3cg project defines two 10 Mbit/s PHYs operating over a single pair of conductors. The 10BASE-T1L (Clause 146) is a long reach PHY supporting full duplex point-to-point operation over 1 km of single balanced pair of conductors. The 10BASE-T1S (Clause 147) is a short reach PHY supporting full / half duplex point-to-point operation over 15 m of single balanced pair of conductors, or half duplex multidrop bus operation over 25 m of single balanced pair of conductors. Furthermore, the IEEE 802.3cg project defines the new Physical Layer Collision Avoidance (PLCA) Reconciliation Sublayer (Clause 148) meant to provide improved determinism to the CSMA/CD media access method. PLCA works in conjunction with the 10BASE-T1S PHY operating in multidrop mode. The aforementioned PHYs are intended to cover the low-speed / low-cost applications in industrial and automotive environment. The large number of pins (16) required by the MII interface, which is specified by the IEEE 802.3 in Clause 22, is one of the major cost factors that need to be addressed to fulfil this objective. The MAC-PHY solution integrates an IEEE Clause 4 MAC and a 10BASE-T1x PHY exposing a low pin count Serial Peripheral Interface (SPI) to the host microcontroller. This also enables the addition of Ethernet functionality to existing low-end microcontrollers which do not integrate a MAC controller. Reviewed-by:
-
Jakub Kicinski authored
Mina Almasry says: ==================== Device Memory TCP Device memory TCP (devmem TCP) is a proposal for transferring data to and/or from device memory efficiently, without bouncing the data to a host memory buffer. * Problem: A large amount of data transfers have device memory as the source and/or destination. Accelerators drastically increased the volume of such transfers. Some examples include: - ML accelerators transferring large amounts of training data from storage into GPU/TPU memory. In some cases ML training setup time can be as long as 50% of TPU compute time, improving data transfer throughput & efficiency can help improving GPU/TPU utilization. - Distributed training, where ML accelerators, such as GPUs on different hosts, exchange data among them. - Distributed raw block storage applications transfer large amounts of data with remote SSDs, much of this data does not require host processing. Today, the majority of the Device-to-Device data transfers the network are implemented as the following low level operations: Device-to-Host copy, Host-to-Host network transfer, and Host-to-Device copy. The implementation is suboptimal, especially for bulk data transfers, and can put significant strains on system resources, such as host memory bandwidth, PCIe bandwidth, etc. One important reason behind the current state is the kernel’s lack of semantics to express device to network transfers. * Proposal: In this patch series we attempt to optimize this use case by implementing socket APIs that enable the user to: 1. send device memory across the network directly, and 2. receive incoming network packets directly into device memory. Packet _payloads_ go directly from the NIC to device memory for receive and from device memory to NIC for transmit. Packet _headers_ go to/from host memory and are processed by the TCP/IP stack normally. The NIC _must_ support header split to achieve this. Advantages: - Alleviate host memory bandwidth pressure, compared to existing network-transfer + device-copy semantics. - Alleviate PCIe BW pressure, by limiting data transfer to the lowest level of the PCIe tree, compared to traditional path which sends data through the root complex. * Patch overview: ** Part 1: netlink API Gives user ability to bind dma-buf to an RX queue. ** Part 2: scatterlist support Currently the standard for device memory sharing is DMABUF, which doesn't generate struct pages. On the other hand, networking stack (skbs, drivers, and page pool) operate on pages. We have 2 options: 1. Generate struct pages for dmabuf device memory, or, 2. Modify the networking stack to process scatterlist. Approach #1 was attempted in RFC v1. RFC v2 implements approach #2. ** part 3: page pool support We piggy back on page pool memory providers proposal: https://github.com/kuba-moo/linux/tree/pp-providers It allows the page pool to define a memory provider that provides the page allocation and freeing. It helps abstract most of the device memory TCP changes from the driver. ** part 4: support for unreadable skb frags Page pool iovs are not accessible by the host; we implement changes throughput the networking stack to correctly handle skbs with unreadable frags. ** Part 5: recvmsg() APIs We define user APIs for the user to send and receive device memory. Not included with this series is the GVE devmem TCP support, just to simplify the review. Code available here if desired: https://github.com/mina/linux/tree/tcpdevmem This series is built on top of net-next with Jakub's pp-providers changes cherry-picked. * NIC dependencies: 1. (strict) Devmem TCP require the NIC to support header split, i.e. the capability to split incoming packets into a header + payload and to put each into a separate buffer. Devmem TCP works by using device memory for the packet payload, and host memory for the packet headers. 2. (optional) Devmem TCP works better with flow steering support & RSS support, i.e. the NIC's ability to steer flows into certain rx queues. This allows the sysadmin to enable devmem TCP on a subset of the rx queues, and steer devmem TCP traffic onto these queues and non devmem TCP elsewhere. The NIC I have access to with these properties is the GVE with DQO support running in Google Cloud, but any NIC that supports these features would suffice. I may be able to help reviewers bring up devmem TCP on their NICs. * Testing: The series includes a udmabuf kselftest that show a simple use case of devmem TCP and validates the entire data path end to end without a dependency on a specific dmabuf provider. ** Test Setup Kernel: net-next with this series and memory provider API cherry-picked locally. Hardware: Google Cloud A3 VMs. NIC: GVE with header split & RSS & flow steering support. ==================== Link: https://patch.msgid.link/20240910171458.219195-1-almasrymina@google.comSigned-off-by:
-
Mina Almasry authored
Add dmabuf information to page_pool stats: $ ./cli.py --spec ../netlink/specs/netdev.yaml --dump page-pool-get ... {'dmabuf': 10, 'id': 456, 'ifindex': 3, 'inflight': 1023, 'inflight-mem': 4190208}, {'dmabuf': 10, 'id': 455, 'ifindex': 3, 'inflight': 1023, 'inflight-mem': 4190208}, {'dmabuf': 10, 'id': 454, 'ifindex': 3, 'inflight': 1023, 'inflight-mem': 4190208}, {'dmabuf': 10, 'id': 453, 'ifindex': 3, 'inflight': 1023, 'inflight-mem': 4190208}, {'dmabuf': 10, 'id': 452, 'ifindex': 3, 'inflight': 1023, 'inflight-mem': 4190208}, {'dmabuf': 10, 'id': 451, 'ifindex': 3, 'inflight': 1023, 'inflight-mem': 4190208}, {'dmabuf': 10, 'id': 450, 'ifindex': 3, 'inflight': 1023, 'inflight-mem': 4190208}, {'dmabuf': 10, 'id': 449, 'ifindex': 3, 'inflight': 1023, 'inflight-mem': 4190208}, And queue stats: $ ./cli.py --spec ../netlink/specs/netdev.yaml --dump queue-get ... {'dmabuf': 10, 'id': 8, 'ifindex': 3, 'type': 'rx'}, {'dmabuf': 10, 'id': 9, 'ifindex': 3, 'type': 'rx'}, {'dmabuf': 10, 'id': 10, 'ifindex': 3, 'type': 'rx'}, {'dmabuf': 10, 'id': 11, 'ifindex': 3, 'type': 'rx'}, {'dmabuf': 10, 'id': 12, 'ifindex': 3, 'type': 'rx'}, {'dmabuf': 10, 'id': 13, 'ifindex': 3, 'type': 'rx'}, {'dmabuf': 10, 'id': 14, 'ifindex': 3, 'type': 'rx'}, {'dmabuf': 10, 'id': 15, 'ifindex': 3, 'type': 'rx'}, Suggested-by:Mina Almasry <almasrymina@google.com> Reviewed-by:
-
Mina Almasry authored
ncdevmem is a devmem TCP netcat. It works similarly to netcat, but it sends and receives data using the devmem TCP APIs. It uses udmabuf as the dmabuf provider. It is compatible with a regular netcat running on a peer, or a ncdevmem running on a peer. In addition to normal netcat support, ncdevmem has a validation mode, where it sends a specific pattern and validates this pattern on the receiver side to ensure data integrity. Suggested-by:
Stanislav Fomichev <sdf@fomichev.me> Signed-off-by:
-
Mina Almasry authored
Add documentation outlining the usage and details of devmem TCP. Signed-off-by:
Bagas Sanjaya <bagasdotme@gmail.com> Reviewed-by:
Donald Hunter <donald.hunter@gmail.com> Link: https://patch.msgid.link/20240910171458.219195-12-almasrymina@google.comSigned-off-by:
-
Mina Almasry authored
Add an interface for the user to notify the kernel that it is done reading the devmem dmabuf frags returned as cmsg. The kernel will drop the reference on the frags to make them available for reuse. Signed-off-by:
Kaiyuan Zhang <kaiyuanz@google.com> Signed-off-by:
Pavel Begunkov <asml.silence@gmail.com> Reviewed-by:
Eric Dumazet <edumazet@google.com> Link: https://patch.msgid.link/20240910171458.219195-11-almasrymina@google.comSigned-off-by:
-
Mina Almasry authored
In tcp_recvmsg_locked(), detect if the skb being received by the user is a devmem skb. In this case - if the user provided the MSG_SOCK_DEVMEM flag - pass it to tcp_recvmsg_devmem() for custom handling. tcp_recvmsg_devmem() copies any data in the skb header to the linear buffer, and returns a cmsg to the user indicating the number of bytes returned in the linear buffer. tcp_recvmsg_devmem() then loops over the unaccessible devmem skb frags, and returns to the user a cmsg_devmem indicating the location of the data in the dmabuf device memory. cmsg_devmem contains this information: 1. the offset into the dmabuf where the payload starts. 'frag_offset'. 2. the size of the frag. 'frag_size'. 3. an opaque token 'frag_token' to return to the kernel when the buffer is to be released. The pages awaiting freeing are stored in the newly added sk->sk_user_frags, and each page passed to userspace is get_page()'d. This reference is dropped once the userspace indicates that it is done reading this page. All pages are released when the socket is destroyed. Signed-off-by:
-
Mina Almasry authored
For device memory TCP, we expect the skb headers to be available in host memory for access, and we expect the skb frags to be in device memory and unaccessible to the host. We expect there to be no mixing and matching of device memory frags (unaccessible) with host memory frags (accessible) in the same skb. Add a skb->devmem flag which indicates whether the frags in this skb are device memory frags or not. __skb_fill_netmem_desc() now checks frags added to skbs for net_iov, and marks the skb as skb->devmem accordingly. Add checks through the network stack to avoid accessing the frags of devmem skbs and avoid coalescing devmem skbs with non devmem skbs. Signed-off-by:
-
Mina Almasry authored
Make skb_frag_page() fail in the case where the frag is not backed by a page, and fix its relevant callers to handle this case. Signed-off-by:
-
Mina Almasry authored
Implement a memory provider that allocates dmabuf devmem in the form of net_iov. The provider receives a reference to the struct netdev_dmabuf_binding via the pool->mp_priv pointer. The driver needs to set this pointer for the provider in the net_iov. The provider obtains a reference on the netdev_dmabuf_binding which guarantees the binding and the underlying mapping remains alive until the provider is destroyed. Usage of PP_FLAG_DMA_MAP is required for this memory provide such that the page_pool can provide the driver with the dma-addrs of the devmem. Support for PP_FLAG_DMA_SYNC_DEV is omitted for simplicity & p.order != 0. Signed-off-by:
-
Mina Almasry authored
Convert netmem to be a union of struct page and struct netmem. Overload the LSB of struct netmem* to indicate that it's a net_iov, otherwise it's a page. Currently these entries in struct page are rented by the page_pool and used exclusively by the net stack: struct { unsigned long pp_magic; struct page_pool *pp; unsigned long _pp_mapping_pad; unsigned long dma_addr; atomic_long_t pp_ref_count; }; Mirror these (and only these) entries into struct net_iov and implement netmem helpers that can access these common fields regardless of whether the underlying type is page or net_iov. Implement checks for net_iov in netmem helpers which delegate to mm APIs, to ensure net_iov are never passed to the mm stack. Signed-off-by: -
Mina Almasry authored
Implement netdev devmem allocator. The allocator takes a given struct netdev_dmabuf_binding as input and allocates net_iov from that binding. The allocation simply delegates to the binding's genpool for the allocation logic and wraps the returned memory region in a net_iov struct. Signed-off-by:
-
Mina Almasry authored
Add a netdev_dmabuf_binding struct which represents the dma-buf-to-netdevice binding. The netlink API will bind the dma-buf to rx queues on the netdevice. On the binding, the dma_buf_attach & dma_buf_map_attachment will occur. The entries in the sg_table from mapping will be inserted into a genpool to make it ready for allocation. The chunks in the genpool are owned by a dmabuf_chunk_owner struct which holds the dma-buf offset of the base of the chunk and the dma_addr of the chunk. Both are needed to use allocations that come from this chunk. We create a new type that represents an allocation from the genpool: net_iov. We setup the net_iov allocation size in the genpool to PAGE_SIZE for simplicity: to match the PAGE_SIZE normally allocated by the page pool and given to the drivers. The user can unbind the dmabuf from the netdevice by closing the netlink socket that established the binding. We do this so that the binding is automatically unbound even if the userspace process crashes. The binding and unbinding leaves an indicator in struct netdev_rx_queue that the given queue is bound, and the binding is actuated by resetting the rx queue using the queue API. The netdev_dmabuf_binding struct is refcounted, and releases its resources only when all the refs are released. Signed-off-by:
Daniel Vetter <daniel.vetter@ffwll.ch> Reviewed-by:
-
