aboutsummaryrefslogtreecommitdiffstats

RAS TA: actually invocable (prior campaigns never were), and audited — negative

Prior campaigns never invoked RAS at all

fuzz_campaign.py recorded RAS returning 0x80000000|cmd for every command and I read that as "unsupported". It was my own harness misreading the buffer: RAS echoes the command at +0x04 with bit 31 set. The real status is at +0x08. RAS commands also come from frame[0], values 0-4 — not ta_cmd_id:

0 ENABLE FEATURES  1 DISABLE FEATURES  2 TRIGGER ERROR
3 QUERY_BLOCK_INFO 4 QUERY_SUB_BLOCK_INFO

Response region is +0x810..0x823 (matches ras_attach_response_buffer's bzero(ptr+0x810, 0x14)), plus +0x410.

asic == 49 HOLDS on this card, so the callbacks are live

ras_device_register installs the handler pointers only when dword_406060[0] == 49. Proof it is satisfied: cmd 0 returns status 0 in 1458 runs but 40978 in 18 — input-dependent, which is impossible if off_4060AC were NULL (NULL -> unconditional return 0). 40978 is emitted by ras_gfx_mode_config. So off_4060AC = ras_apply_gfx_feature is executing.

Installed and host-reachable: ras_apply_gfx_feature (cmd 0), ras_parse_trigger_request (cmd 2), ras_probe_chiplets. Permanently NULL: off_4060B8 (cmds 3/4) — it is not in the install list, which is why QUERY_BLOCK_INFO/QUERY_SUB_BLOCK_INFO always return 40976.

Campaigns run (all negative, card healthy throughout)

  • cmds 0/1/3/4, 6000 rounds, request area fuzzed: 0 hangs, 0 PSP-range pointers in the response.
  • cmd 2 TRIGGER_ERROR, 2000 rounds: all rejected at 40962/40986 — the block-type gate was never satisfied because v4 (64-bit at request+4) must be 2 or 4 with high dword 0.
  • cmd 2 with the gate satisfied, 2500 rounds: new status 40977 appeared (227x), i.e. host input now reaches the register-access layer (ras_reg_read/write{,_chiplet,64}). Still 0 hangs, 0 leaks, MEMSIZE=0x5ff0 and sos_alive checked every 200 rounds.

Refinement of FINDING-SYSDRV-CHIPLET-UNBOUNDED-INDEX.md

That doc says "all 21 call sites pass literal 0". Disassembly of ras_reg_write_chiplet @0x403b04 shows the first RPC argument is a mapped value, not a literal 0:

CMP R2,#2 / BCS .L1 ; MOVS R0,#2      a3 < 2  -> 2
.L1 ITE NE ; MOVNE R0,#0 ; MOVEQ R0,#8  a3 > 2 -> 0 , a3 == 2 -> 8

So the value handed to SYS_DRV 0x7006/0x7007 is one of {0, 2, 8} — still bounded, so the conclusion (no unbounded chiplet index from RAS) stands, but the stated reason was imprecise. The SXTBLS signed-byte index into dword_404838 is guarded by CMP R7,#2 + LS, so it cannot go negative.

Status

RAS is genuinely host-reachable and now genuinely exercised. No bug found.