drivers/i8255x: get to work on a real device #725

2026-09-26T17:48:34Z

Leency commented

2026-09-26 17:48:34 +00:00

the TX reclaim loop in int_handler multiplied sizeof.txfd by itself (mul eax) instead of by last_tx, so it looked 1024 bytes past tx_ring and never freed a descriptor
it also left txfd.status set, and transmit treats a non-zero status as busy, so once the 16-entry ring wrapped every send ended in "TX overrun": DHCP fell back to link-local, ping showed "Socket error", while the link still read as connected
reclaim only descriptors the device completed (TXFD_STATUS_C) and zero their status; transmit now checks for an unreclaimed buffer (virt_addr)

Seen on an EtherExpress Pro/100 in a Pentium II. Reproduced in QEMU (-cpu pentium2, i82557b): ping stopped at the 12th reply every time; with the fix it runs indefinitely, also under a UDP flood.

- the TX reclaim loop in int_handler multiplied sizeof.txfd by itself (mul eax) instead of by last_tx, so it looked 1024 bytes past tx_ring and never freed a descriptor - it also left txfd.status set, and transmit treats a non-zero status as busy, so once the 16-entry ring wrapped every send ended in "TX overrun": DHCP fell back to link-local, ping showed "Socket error", while the link still read as connected - reclaim only descriptors the device completed (TXFD_STATUS_C) and zero their status; transmit now checks for an unreclaimed buffer (virt_addr) Seen on an EtherExpress Pro/100 in a Pentium II. Reproduced in QEMU (-cpu pentium2, i82557b): ping stopped at the 12th reply every time; with the fix it runs indefinitely, also under a UDP flood.

CODEOWNERS rules requested review from system 2026-09-26 17:48:34 +00:00

CODEOWNERS rules requested review from network 2026-09-26 17:48:34 +00:00

Leency requested review from hidnplayr 2026-09-26 17:48:48 +00:00

Leency marked the pull request as work in progress 2026-09-26 23:57:27 +00:00

Leency changed title from ~~WIP: drivers/i8255x: reclaim sent TX descriptors, network died after 16 packets~~ to drivers/i8255x: get to work on a real device

2026-09-27 10:52:12 +00:00

hidnplayr approved these changes

2026-09-28 10:49:44 +00:00

Burer approved these changes

2026-09-29 13:09:34 +00:00

Dismissed

Burer left a comment

Must fix:

init_tx_ring doesn't clear txfd.virt_addr. After a reset with unreclaimed frames, transmit reports "TX overrun" forever. NetFree or zero it in .next_desc.
DEBUGF 2 is errors-only. Move "Command ... done" and "Restarting receiver" to level 1, and drop "Out of resources!" (now handled).

Nits:

confcmd_wait waits ~100 ms, not 1 s (udelay = Sleep(1)).
.out_of_mem still counts the dropped frame in packets_rx/bytes_rx.

Must fix: 1. `init_tx_ring` doesn't clear `txfd.virt_addr`. After a reset with unreclaimed frames, `transmit` reports "TX overrun" forever. `NetFree` or zero it in `.next_desc`. 2. `DEBUGF 2` is errors-only. Move "Command ... done" and "Restarting receiver" to level 1, and drop "Out of resources!" (now handled). Nits: - `confcmd_wait` waits ~100 ms, not 1 s (`udelay` = `Sleep(1)`). - `.out_of_mem` still counts the dropped frame in `packets_rx`/`bytes_rx`.

Burer requested changes

2026-09-29 13:10:02 +00:00

Dismissed

Burer left a comment

See comment above.

Leency referenced this issue from a commit

2026-09-29 16:10:24 +00:00

drivers/i8255x: address review of #725

CODEOWNERS rules requested review from system 2026-09-29 16:10:25 +00:00

CODEOWNERS rules requested review from network 2026-09-29 16:10:25 +00:00

Burer approved these changes

2026-09-29 18:11:06 +00:00

Burer referenced this issue from a commit

2026-09-29 18:59:51 +00:00

drivers/i8255x: address review of #725

Burer added 4 commits 2026-09-29 18:59:51 +00:00

drivers/i8255x: reclaim sent TX descriptors, network died after 16 packets 4a7f937160

- the TX reclaim loop in int_handler multiplied sizeof.txfd by itself
  (mul eax) instead of by last_tx, so it looked 1024 bytes past tx_ring
  and never freed a descriptor
- it also left txfd.status set, and transmit treats a non-zero status
  as busy, so once the 16-entry ring wrapped every send ended in
  "TX overrun": DHCP fell back to link-local, ping showed "Socket error",
  while the link still read as connected
- reclaim only descriptors the device completed (TXFD_STATUS_C) and zero
  their status; transmit now checks for an unreclaimed buffer (virt_addr)

Seen on an EtherExpress Pro/100 in a Pentium II. Reproduced in QEMU
(-cpu pentium2, i82557b): ping stopped at the 12th reply every time;
with the fix it runs indefinitely, also under a UDP flood.

Assisted-by: Claude Opus 5 <noreply@anthropic.com>

drivers/i8255x: restart the receiver when it stops, fill in the RFD 6e66706a8d

The driver keeps a single RFD and re-armed the receive unit only after
taking a frame (FR). A frame arriving before that left the RU in No
Resources, and on the RNR interrupt the driver just printed "Out of
resources!": on an EtherExpress Pro/100 in a Pentium II every DHCP
answer was dropped this way.

- handle RNR like FR, and after the RX loop re-arm the RFD whenever the
  RU is not Ready (once per interrupt, marked in reserved bit 0 of the
  saved status, so a device that never gets Ready cannot loop us)
- set the RBD address to 0xFFFFFFFF, as simplified mode wants; it held
  whatever the net buffer had
- without a new buffer, drop the frame and reuse the RFD: the buffer was
  handed to the stack while the RFD still pointed at it

QEMU never enters No Resources; the restart path was exercised with a
build that skips the normal RX_START (38 restarts, DHCP and ping fine).
On the Pentium II the log shows "Restarting receiver, RU status 08".

Assisted-by: Claude Opus 5 <noreply@anthropic.com>

drivers/i8255x: wait for commands to complete, not only to be accepted e48b560e6d

reset() gives the chip Individual Address Setup and then Configure in
the same confcmd block, waiting in between only for the SCB to accept
the command (cmd_wait). A real chip had not read the MAC yet when the
Configure data overwrote it, so our address never reached the filter:
broadcasts came in, frames sent to us (DHCP offer, ARP reply, TCP) were
dropped - "connected", yet nothing worked. QEMU executes at once and
never showed it.

- confcmd_wait: poll the C bit of confcmd (up to 1 s) after both
  commands and log the result; the Pentium II now logs
  "Command 0x4001 done, status 0xA000" and gets its DHCP lease
- transmit: CU_START is only valid while the CU is idle or suspended,
  so wait for the previous frame to leave before starting the next

Assisted-by: Claude Opus 5 <noreply@anthropic.com>

drivers/i8255x: address review of #725

Test PR / Build (es_ES) (pull_request) Successful in 2m56s

Details

Test PR / Build (en_US) (pull_request) Successful in 2m59s

Details

Test PR / Build (ru_RU) (pull_request) Successful in 3m2s

Details

fa87fa33e3

- init_tx_ring frees and clears txfd.virt_addr: after a reset with frames
  still unreclaimed, transmit took the descriptor as busy and reported
  "TX overrun" forever; init_rx_ring frees the RFD of an earlier reset
- "Command ... done" and "Restarting receiver" go to debug level 1, level
  2 is for errors; "Out of resources!" is dropped, the RU is re-armed now
- confcmd_wait: udelay is Sleep(1) = 1 ms, so 1000 rounds for the 1 s the
  comment promises (it waited 100 ms)
- a frame dropped for want of a buffer is taken back out of packets_rx /
  bytes_rx and counted in packets_rx_drop

Checked in QEMU on i82557a, i82558b, i82559c: DHCP, ping under a UDP flood.

Assisted-by: Claude Opus 5 <noreply@anthropic.com>

Burer force-pushed i8255x-tx-reclaim from 3479361723 to fa87fa33e3

2026-09-29 18:59:51 +00:00

Compare

Burer merged commit 9ec66b0eba into main

2026-09-29 19:03:25 +00:00

Burer deleted branch i8255x-tx-reclaim

2026-09-29 19:03:25 +00:00

Sign in to join this conversation.

No Reviewers

hidnplayr

Burer

KolibriOS/system

KolibriOS/network

3 Participants

Notifications

Due Date

No due date set.

Dependencies

No dependencies set.

Reference: KolibriOS/kolibrios#725