import os import re import subprocess import sys import time from .common import FileDownloader from ..utils import ( encodeFilename, format_bytes, ) class RtmpFD(FileDownloader): def real_download(self, filename, info_dict): def run_rtmpdump(args): start = time.time() resume_percent = None resume_downloaded_data_len = None proc = subprocess.Popen(args, stderr=subprocess.PIPE) cursor_in_new_line = True proc_stderr_closed = False while not proc_stderr_closed: # read line from stderr line = u'' while True: char = proc.stderr.read(1) if not char: proc_stderr_closed = True break if char in [b'\r', b'\n']: break line += char.decode('ascii', 'replace') if not line: # proc_stderr_closed is True continue mobj = re.search(r'([0-9]+\.[0-9]{3}) kB / [0-9]+\.[0-9]{2} sec \(([0-9]{1,2}\.[0-9])%\)', line) if mobj: downloaded_data_len = int(float(mobj.group(1))*1024) percent = float(mobj.group(2)) if not resume_percent: resume_percent = percent resume_downloaded_data_len = downloaded_data_len eta = self.calc_eta(start, time.time(), 100-resume_percent, percent-resume_percent) speed = self.calc_speed(start, time.time(), downloaded_data_len-resume_downloaded_data_len) data_len = None if percent > 0: data_len = int(downloaded_data_len * 100 / percent) data_len_str = u'~' + format_bytes(data_len) self.report_progress(percent, data_len_str, speed, eta) cursor_in_new_line = False self._hook_progress({ 'downloaded_bytes': downloaded_data_len, 'total_bytes': data_len, 'tmpfilename': tmpfilename, 'filename': filename, 'status': 'downloading', 'eta': eta, 'speed': speed, }) else: # no percent for live streams mobj = re.search(r'([0-9]+\.[0-9]{3}) kB / [0-9]+\.[0-9]{2} sec', line) if mobj: downloaded_data_len = int(float(mobj.group(1))*1024) time_now = time.time() speed = self.calc_speed(start, time_now, downloaded_data_len) self.report_progress_live_stream(downloaded_data_len, speed, time_now - start) cursor_in_new_line = False self._hook_progress({ 'downloaded_bytes': downloaded_data_len, 'tmpfilename': tmpfilename, 'filename': filename, 'status': 'downloading', 'speed': speed, }) elif self.params.get('verbose', False): if not cursor_in_new_line: self.to_screen(u'') cursor_in_new_line = True self.to_screen(u'[rtmpdump] '+line) proc.wait() if not cursor_in_new_line: self.to_screen(u'') return proc.returncode url = info_dict['url'] player_url = info_dict.get('player_url', None) page_url = info_dict.get('page_url', None) play_path = info_dict.get('play_path', None) tc_url = info_dict.get('tc_url', None) live = info_dict.get('rtmp_live', False) conn = info_dict.get('rtmp_conn', None) self.report_destination(filename) tmpfilename = self.temp_name(filename) test = self.params.get('test', False) # Check for rtmpdump first try: subprocess.call(['rtmpdump', '-h'], stdout=(open(os.path.devnull, 'w')), stderr=subprocess.STDOUT) except (OSError, IOError): self.report_error(u'RTMP download detected but "rtmpdump" could not be run') return False # Download using rtmpdump. rtmpdump returns exit code 2 when # the connection was interrumpted and resuming appears to be # possible. This is part of rtmpdump's normal usage, AFAIK. basic_args = ['rtmpdump', '--verbose', '-r', url, '-o', tmpfilename] if player_url is not None: basic_args += ['--swfVfy', player_url] if page_url is not None: basic_args += ['--pageUrl', page_url] if play_path is not None: basic_args += ['--playpath', play_path] if tc_url is not None: basic_args += ['--tcUrl', url] if test: basic_args += ['--stop', '1'] if live: basic_args += ['--live'] if conn: basic_args += ['--conn', conn] args = basic_args + [[], ['--resume', '--skip', '1']][self.params.get('continuedl', False)] if sys.platform == 'win32' and sys.version_info < (3, 0): # Windows subprocess module does not actually support Unicode # on Python 2.x # See http://stackoverflow.com/a/9951851/35070 subprocess_encoding = sys.getfilesystemencoding() args = [a.encode(subprocess_encoding, 'ignore') for a in args] else: subprocess_encoding = None if self.params.get('verbose', False): if subprocess_encoding: str_args = [ a.decode(subprocess_encoding) if isinstance(a, bytes) else a for a in args] else: str_args = args try: import pipes shell_quote = lambda args: ' '.join(map(pipes.quote, str_args)) except ImportError: shell_quote = repr self.to_screen(u'[debug] rtmpdump command line: ' + shell_quote(str_args)) retval = run_rtmpdump(args) while (retval == 2 or retval == 1) and not test: prevsize = os.path.getsize(encodeFilename(tmpfilename)) self.to_screen(u'[rtmpdump] %s bytes' % prevsize) time.sleep(5.0) # This seems to be needed retval = run_rtmpdump(basic_args + ['-e'] + [[], ['-k', '1']][retval == 1]) cursize = os.path.getsize(encodeFilename(tmpfilename)) if prevsize == cursize and retval == 1: break # Some rtmp streams seem abort after ~ 99.8%. Don't complain for those if prevsize == cursize and retval == 2 and cursize > 1024: self.to_screen(u'[rtmpdump] Could not download the whole video. This can happen for some advertisements.') retval = 0 break if retval == 0 or (test and retval == 2): fsize = os.path.getsize(encodeFilename(tmpfilename)) self.to_screen(u'[rtmpdump] %s bytes' % fsize) self.try_rename(tmpfilename, filename) self._hook_progress({ 'downloaded_bytes': fsize, 'total_bytes': fsize, 'filename': filename, 'status': 'finished', }) return True else: self.to_stderr(u"\n") self.report_error(u'rtmpdump exited with code %d' % retval) return False