FreeBSD source tree https://cgit-dev.freebsd.org/src/atom?h=main 2026-04-20T17:18:04Z 2026-04-20T17:18:04Z John Baldwin jhb@FreeBSD.org 2026-04-20T17:18:04Z urn:sha1:505f381cb96ee25bcf0df11cfaf15a601d969500 If soreserve() or sosetopt() (to set TCP_NODELAY) fails after claiming the socket from the file descriptor, explicitly close the socket before returning failure. Sponsored by: Chelsio Communications Differential Revision: https://reviews.freebsd.org/D55493 2026-04-12T18:33:07Z Gleb Smirnoff glebius@FreeBSD.org 2026-04-12T18:33:07Z urn:sha1:40dbb06fa73cac37d57563c07e55efd0cabbd488 The inpcb flag INP_DROPPED served two purposes. It was used by TCP and subsystems running on top of TCP as a flag that marks a connection that is now in TCPS_CLOSED, but was in some other state before (not a new-born connection). Create a new TCP flag TF_DISCONNECTED for this purpose. The in_pcbdrop() was a TCP's version of in_pcbdisconnect() that also sets INP_DROPPED. Use in_pcbdisconnect() instead. Second purpose of INP_DROPPED was a negative lookup mask in inp_smr_lock(), as SMR-protected lookup may see inpcbs that had been removed from the hash. We already have had INP_INHASHLIST that marks inpcb that is in hash. Convert it into INP_UNCONNECTED with the opposite meaning. This allows to combine it with INP_FREED for the negative lookup mask. The Chelsio/ToE and kTLS changes are done with some style refactoring, like moving inp/tp assignments up and using macros for that. However, no deep thinking was taken to check if those checks are really needed, it could be that some are not. Reviewed by: rrs Differential Revision: https://reviews.freebsd.org/D56186 2025-11-12T20:04:48Z Navdeep Parhar np@FreeBSD.org 2025-11-12T20:04:48Z urn:sha1:e353cb88d4581303a4d9af090974ec88463802e6 Fixes: ec0cd287f55f nvmf_che: NVMe-TCP offload support for Chelsio T7 adapters Sponsored by: Chelsio Communications 2025-11-10T15:50:48Z John Baldwin jhb@FreeBSD.org 2025-11-10T15:50:48Z urn:sha1:ec0cd287f55f7ea93ff4ccfa4de0f70eca5fef75 This provides an alternative NVMe over TCP transport which uses PDU offload for TOE connections on a T7. Similar to iSCSI offload via cxgbei.ko, nvmf_che uses DDP when possible to enable the NIC to DMA received data directly into I/O data buffers (pages from a struct bio on the host side, pages from a CTL I/O requests on the controller side) to avoid copying data on the host CPU. nvmf_che is also able to receive a stream of C2H or H2C PDUs for a single data transfer when using DDP without processing the header of each PDU. Unlike cxgbei, nvmf_che aims to be mostly transparent to end users. Notably, neither nvmecontrol or ctld have to be explicitly asked to use an offload. Instead, TCP queue pairs are claimed by this driver whenever they are eligible (e.g., using TOE). The main restriction of nvmf_che compared to the software TCP transport is that Chelsio adapters have a restriction on the largest PDU that can be sent and received. When sending data, nvmf_che is able to split large C2H or H2C data requests across multiple PDUs without affecting nvmf(4) or nvmft(4). To avoid overly large PDUs when using nvmf(4), nvmf_che reports a data transfer limit that is honored by nvmf(4). This ensures that the remote controller's PDUs will never be too large (since the command transfer size is limited to one PDU) and also ensures that nvmf(4) will never to try to send a command PDU with ICD that is too large. For nvmft(4), overly large command PDUs due to ICD are avoided by clamping the size of the reported IOCCSZ in the controller data. However, to ensure that H2C PDUs are sufficiently small, nvmf_che will only claim queue pairs which advertised a suitable MAXH2CDATA parameter during queue negotiation. For ctld(8), this can be achieved by setting the MAXH2CDATA option in a transport-group, e.g. for T7: transport-group tg0 { discovery-auth-group no-authentication listen tcp 0.0.0.0 listen tcp [::] listen discovery-tcp 0.0.0.0 listen discovery-tcp [::] option MAXH2CDATA 32488 } Sponsored by: Chelsio Communications