hl_fw_dynam...protocol_cmd(9)

📦 version 7.3.1, in nixpkgs unstable

💅 hl_fw_dynamic_send_protocol_cmd — the Kernel API Manual page, but make it make sense

okay so bestie, this is a page out of the Kernel API Manual, and it's about one function — hl_fw_dynamic_send_protocol_cmd. It's a little guy that does a big job, and honestly the way it just sends a command and waits for an ACK? iconic. let's get into it ✨

🎀 NAME 🎀

hl_fw_dynamic_send_protocol_cmd — send LKD to FW cmd and wait for ACK

like, that's literally it. that's the whole vibe. LKD talks to FW, then waits for the ACK like a girl waiting for a text back 💌

🌸 SYNOPSIS 🌸

int hl_fw_dynamic_send_protocol_cmd (struct hl_device *hdev , struct
fw_load_mgr *fw_loader , enum comms_cmd cmd , unsigned int size , bool
wait_ok , u32 timeout );

yes it's long. yes it returns an int. no we're not scared of it. 🫶

🦋 ARGUMENTS 🦋

  • hdev — pointer to the habanalabs device structure. the main character, the device itself 👑
  • fw_loader — managing structure for loading device's FW. it's the one holding the clipboard and keeping everything organized 📋
  • cmd — LKD to FW cmd code. the actual message being sent 💌
  • size — size of next FW component to be loaded (0 if not necessary). lowkey optional depending on the situation
  • wait_ok — if true also wait for OK response from FW. if you set this true bestie you're committing to the wait 🙈
  • timeout — timeout for status wait. the "how long am I willing to sit here" number ⏳

🔥 DESCRIPTION 🔥

return 0 on success, otherwise non-zero error code. periodt — 0 means slay, anything else means something went sideways and you should look at it 💀

brief: when sending protocol command we have the following steps:

  1. send clear (clear command and verify clear status register)
  2. send the actual protocol command
  3. wait for ACK on the protocol command
  4. send clear
  5. send NOOP if, in addition, the specific protocol command should wait for OK then:
  6. wait for OK
  7. send clear
  8. send NOOP

not me listing eight whole steps 😭 but like, the order MATTERS, so we do not skip, we do not shuffle, we do not freestyle. the sequence is the sequence, iykyk 💅

🧚 NOTES 🧚

  • send clear: this is necessary in order to clear the status register to avoid leftovers between command
  • NOOP command: necessary to avoid loop on the clear command by the FW

tbh these two notes are the whole reason the steps look repetitive — the clear and the NOOP are load-bearing, not filler. without them you get leftover register junk and the FW going in circles, and nobody wants that 🌈

⭐ SEE ALSO ⭐

Kernel file ./drivers/accel/habanalabs/common/firmware_if.c

the friend group (all (9) sections, all linked, all here, no cap):

that's a whole thirty entries bestie, we got the entire firmware-interface extended universe in here 🌟

okay that's the page!! mwah 💋 it lives in ./drivers/accel/habanalabs/common and it's dated October 2026, so go forth, send your protocol command, wait for that ACK, and remember: clear the status register first. bye bestie 🫶💖✨

📦 where this page comes from