#!/usr/bin/env python3 # guest_pin_stream.py - stream the CUDA client's push-buffer VA->PA pairs to # the host over vsock, continuously and without the guest agent on the hot # path. Runs in the guest as root (pagemap reads). # # Protocol (text lines to the host): # #pid the client pid just changed (host resets its state) # 0x,0x, n contiguous pages (va+i*0x1000 -> pa+i*0x1000) # # The host side is tools/guest-diag/pin_gr_stream.py. # # Usage: sudo python3 guest_pin_stream.py [host-cid] [port] [lo-hex] [hi-hex] [pa-max-hex] import os import socket import sys import time CID = int(sys.argv[1]) if len(sys.argv) > 1 else 2 PORT = int(sys.argv[2]) if len(sys.argv) > 2 else 45678 LO = int(sys.argv[3], 16) if len(sys.argv) > 3 else 0x200400000 HI = int(sys.argv[4], 16) if len(sys.argv) > 4 else 0x206000000 PAMAX = int(sys.argv[5], 16) if len(sys.argv) > 5 else 0x180000000 INTERVAL = float(os.environ.get("STREAM_INTERVAL", "0.25")) BATCH = 4096 REGION = 0x200000 SPAN = int(os.environ.get("STREAM_SPAN", "0x10000"), 16) def cdt_pid(): for d in os.listdir('/proc'): if not d.isdigit(): continue try: with open('/proc/%s/comm' % d) as f: if f.read().strip() == 'cdt': return int(d) except OSError: pass return None def scan(pid): """Present pages in the first SPAN bytes of every REGION, as {va: pa}. The pushes sit at the start of each per-channel allocation (2/4 MB aligned); scanning only the allocation heads keeps the per-loop page table work small, which is what makes the stream fast while the client is mapping and the guest is loaded. """ fd = os.open('/proc/%d/pagemap' % pid, os.O_RDONLY) out = {} try: base = LO & ~(REGION - 1) while base < HI: va = max(base, LO) & ~0xFFF end = min(base + SPAN, HI) while va < end: n = min(BATCH, (end - va) // 0x1000) if n <= 0: break os.lseek(fd, (va // 4096) * 8, 0) data = os.read(fd, n * 8) got = len(data) // 8 for i in range(got): e = int.from_bytes(data[i * 8:i * 8 + 8], 'little') if not ((e >> 63) & 1): continue pfn = e & ((1 << 55) - 1) pa = pfn << 12 if not pa or (PAMAX and pa >= PAMAX): continue out[va + i * 0x1000] = pa va += got * 0x1000 base += REGION finally: os.close(fd) return out def main(): with open('/tmp/guest_pin_stream.pid', 'w') as f: f.write('%d\n' % os.getpid()) sock = None seen = {} last_pid = None last_log = 0.0 while True: pid = cdt_pid() if pid is None: last_pid = None seen = {} time.sleep(INTERVAL) continue if pid != last_pid: last_pid = pid seen = {} if sock is not None: try: sock.sendall(("#pid %d\n" % pid).encode()) except OSError: sock.close() sock = None t0 = time.time() try: cur = scan(pid) except OSError: time.sleep(INTERVAL) continue dt = time.time() - t0 new = [(va, pa) for va, pa in cur.items() if seen.get(va) != pa] seen = cur now = time.time() if now - last_log > 1.0: last_log = now try: with open('/tmp/guest_pin_stream.log', 'a') as f: f.write('%s pid=%s present=%d new=%d scan=%.3fs\n' % (time.strftime('%T'), pid, len(cur), len(new), dt)) except OSError: pass if not new: time.sleep(INTERVAL) continue new.sort() lines = [] va0, pa0, n = new[0][0], new[0][1], 1 for va, pa in new[1:]: if va == va0 + n * 0x1000 and pa == pa0 + n * 0x1000: n += 1 continue lines.append('0x%x,0x%x,%d\n' % (va0, pa0, n)) va0, pa0, n = va, pa, 1 lines.append('0x%x,0x%x,%d\n' % (va0, pa0, n)) payload = ''.join(lines).encode() if sock is None: try: sock = socket.socket(socket.AF_VSOCK, socket.SOCK_STREAM) sock.settimeout(5) sock.connect((CID, PORT)) sock.settimeout(None) if last_pid is not None: sock.sendall(("#pid %d\n" % last_pid).encode()) except OSError: if sock: sock.close() sock = None time.sleep(INTERVAL) continue try: sock.sendall(payload) except OSError: sock.close() sock = None time.sleep(INTERVAL) if __name__ == '__main__': main()