TID RDMA READ PROTOCOL(9)

💅 TID RDMA READ protocol — the hfi1 glow-up, but make it a manpage 🦋

okay so real talk, this page lives in the Kernel API Manual (section 9, iykyk) and it's literally the blueprint for the TID RDMA READ protocol. no cap, it's iconic. 🎀

🌸 DESCRIPTION 🌸

bestie, picture this: it's an end-to-end protocol at the hfi1 level between two nodes, and its whole main-character move is improving performance by avoiding data copy on the requester side. the way it does that?? it takes a qualified RDMA READ request and converts it into a TID RDMA READ request on the requester side, and after that it handles the request and response completely differently. slay. 💅

but here's the vibe check — to be qualified, the RDMA READ request should meet all of these, fr fr:

• ✨ The total data length should be greater than 256K;
• ✨ The total data length should be a multiple of 4K page size;
• ✨ Each local scatter-gather entry should be 4K page aligned;
• ✨ Each local scatter-gather entry should be a multiple of 4K page size.

miss a rule? then it's not getting the TID RDMA READ treatment, periodt. it has to check every box. 🙈

🔥 SEE ALSO 🔥

the whole squad, in order, no gatekeeping 🫶 — and yes, the kernel file is literally ./drivers/infiniband/hw/hfi1/tid_rdma.c:

• qp_to_rcd(9)
• lock ordering(9)
• kernel_tid_waiters(9)
• dequeue_tid_waiter(9)
• queue_qp_for_tid_wait(9)
• __trigger_tid_waiter(9)
• tid_rdma_schedule_tid_wakeup(9)
• tid_rdma_trigger_resume(9)
• kern_reserve_flow(9)
• tid_rdma_find_phys_blocks_4k(9)
• tid_flush_pages(9)
• tid_rdma_find_phys_blocks_8k(9)
• hfi1_kern_exp_rcv_setup(9)
• hfi1_kern_exp_rcv_free_flows(9)
• __trdma_clean_swqe(9)
• hfi1_schedule_tid_send(9)

🧚 sign-off 🧚

and that's the whole page, bestie!! this lil gem comes to you from ./drivers/infiniband/hw/hfi1 and it's dated October 2026 💖 go forth and read your RDMA reads the TID way. mwah. ⭐🌈👑