lib/util/select.c

   1 /*
   2    Unix SMB/Netbios implementation.
   3    Version 3.0
   4    Samba select/poll implementation
   5    Copyright (C) Andrew Tridgell 1992-1998
   6
   7    This program is free software; you can redistribute it and/or modify
   8    it under the terms of the GNU General Public License as published by
   9    the Free Software Foundation; either version 3 of the License, or
  10    (at your option) any later version.
  11
  12    This program is distributed in the hope that it will be useful,
  13    but WITHOUT ANY WARRANTY; without even the implied warranty of
  14    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  15    GNU General Public License for more details.
  16
  17    You should have received a copy of the GNU General Public License
  18    along with this program.  If not, see <http://www.gnu.org/licenses/>.
  19 */
  20
  21 #include "includes.h"
  22 #include "system/filesys.h"
  23 #include "system/select.h"
  24 #include "lib/util/select.h"
  25
  26 /* This is here because it allows us to avoid a nasty race in signal handling.
  27    We need to guarantee that when we get a signal we get out of a select immediately
  28    but doing that involves a race condition. We can avoid the race by getting the
  29    signal handler to write to a pipe that is in the select/poll list
  30
  31    This means all Samba signal handlers should call sys_select_signal().
  32 */
  33
  34 static pid_t initialised;
  35 static int select_pipe[2];
  36 static volatile unsigned pipe_written, pipe_read;
  37
  38 /*******************************************************************
  39  Call this from all Samba signal handlers if you want to avoid a
  40  nasty signal race condition.
  41 ********************************************************************/
  42
  43 void sys_select_signal(char c)
  44 {
  45         int saved_errno = errno;
  46
  47         if (!initialised) return;
  48
  49         if (pipe_written > pipe_read+256) return;
  50
  51         if (write(select_pipe[1], &c, 1) == 1) pipe_written++;
  52
  53         errno = saved_errno;
  54 }
  55
  56 /*******************************************************************
  57  Like select() but avoids the signal race using a pipe
  58  it also guuarantees that fds on return only ever contains bits set
  59  for file descriptors that were readable.
  60 ********************************************************************/
  61
  62 int sys_select(int maxfd, fd_set *readfds, fd_set *writefds, fd_set *errorfds, struct timeval *tval)
  63 {
  64         int ret, saved_errno;
  65         fd_set *readfds2, readfds_buf;
  66
  67         if (initialised != sys_getpid()) {
  68                 if (pipe(select_pipe) == -1)
  69                 {
  70                         DEBUG(0, ("sys_select: pipe failed (%s)\n",
  71                                 strerror(errno)));
  72                         if (readfds != NULL)
  73                                 FD_ZERO(readfds);
  74                         if (writefds != NULL)
  75                                 FD_ZERO(writefds);
  76                         if (errorfds != NULL)
  77                                 FD_ZERO(errorfds);
  78                         return -1;
  79                 }
  80
  81                 /*
  82                  * These next two lines seem to fix a bug with the Linux
  83                  * 2.0.x kernel (and probably other UNIXes as well) where
  84                  * the one byte read below can block even though the
  85                  * select returned that there is data in the pipe and
  86                  * the pipe_written variable was incremented. Thanks to
  87                  * HP for finding this one. JRA.
  88                  */
  89
  90                 if(set_blocking(select_pipe[0],0)==-1)
  91                         smb_panic("select_pipe[0]: O_NONBLOCK failed");
  92                 if(set_blocking(select_pipe[1],0)==-1)
  93                         smb_panic("select_pipe[1]: O_NONBLOCK failed");
  94
  95                 initialised = sys_getpid();
  96         }
  97
  98         maxfd = MAX(select_pipe[0]+1, maxfd);
  99
 100         /* If readfds is NULL we need to provide our own set. */
 101         if (readfds) {
 102                 readfds2 = readfds;
 103         } else {
 104                 readfds2 = &readfds_buf;
 105                 FD_ZERO(readfds2);
 106         }
 107         FD_SET(select_pipe[0], readfds2);
 108
 109         errno = 0;
 110         ret = select(maxfd,readfds2,writefds,errorfds,tval);
 111
 112         if (ret <= 0) {
 113                 FD_ZERO(readfds2);
 114                 if (writefds)
 115                         FD_ZERO(writefds);
 116                 if (errorfds)
 117                         FD_ZERO(errorfds);
 118         } else if (FD_ISSET(select_pipe[0], readfds2)) {
 119                 char c;
 120                 saved_errno = errno;
 121                 if (read(select_pipe[0], &c, 1) == 1) {
 122                         pipe_read++;
 123                         /* Mark Weaver <mark-clist@npsl.co.uk> pointed out a critical
 124                            fix to ensure we don't lose signals. We must always
 125                            return -1 when the select pipe is set, otherwise if another
 126                            fd is also ready (so ret == 2) then we used to eat the
 127                            byte in the pipe and lose the signal. JRA.
 128                         */
 129                         ret = -1;
 130 #if 0
 131                         /* JRA - we can use this to debug the signal messaging... */
 132                         DEBUG(0,("select got %u signal\n", (unsigned int)c));
 133 #endif
 134                         errno = EINTR;
 135                 } else {
 136                         FD_CLR(select_pipe[0], readfds2);
 137                         ret--;
 138                         errno = saved_errno;
 139                 }
 140         }
 141
 142         return ret;
 143 }
 144
 145 /*******************************************************************
 146  Similar to sys_select() but catch EINTR and continue.
 147  This is what sys_select() used to do in Samba.
 148 ********************************************************************/
 149
 150 int sys_select_intr(int maxfd, fd_set *readfds, fd_set *writefds, fd_set *errorfds, struct timeval *tval)
 151 {
 152         int ret;
 153         fd_set *readfds2, readfds_buf, *writefds2, writefds_buf, *errorfds2, errorfds_buf;
 154         struct timeval tval2, *ptval;
 155         struct timespec end_time;
 156
 157         readfds2 = (readfds ? &readfds_buf : NULL);
 158         writefds2 = (writefds ? &writefds_buf : NULL);
 159         errorfds2 = (errorfds ? &errorfds_buf : NULL);
 160         if (tval) {
 161                 clock_gettime_mono(&end_time);
 162                 end_time.tv_sec += tval->tv_sec;
 163                 end_time.tv_nsec += tval->tv_usec *1000;
 164                 end_time.tv_sec += end_time.tv_nsec / 1000000000;
 165                 end_time.tv_nsec %= 1000000000;
 166                 errno = 0;
 167                 tval2 = *tval;
 168                 ptval = &tval2;
 169         } else {
 170                 ptval = NULL;
 171         }
 172
 173         do {
 174                 if (readfds)
 175                         readfds_buf = *readfds;
 176                 if (writefds)
 177                         writefds_buf = *writefds;
 178                 if (errorfds)
 179                         errorfds_buf = *errorfds;
 180                 if (ptval && (errno == EINTR)) {
 181                         struct timespec now_time;
 182                         int64_t tdif;
 183
 184                         clock_gettime_mono(&now_time);
 185                         tdif = nsec_time_diff(&end_time,&now_time);
 186                         if (tdif <= 0) {
 187                                 ret = 0; /* time expired. */
 188                                 break;
 189                         }
 190                         ptval->tv_sec = tdif / 1000000000;
 191                         ptval->tv_usec = (tdif % 1000000000) / 1000;
 192                 }
 193
 194                 /* We must use select and not sys_select here. If we use
 195                    sys_select we'd lose the fact a signal occurred when sys_select
 196                    read a byte from the pipe. Fix from Mark Weaver
 197                    <mark-clist@npsl.co.uk>
 198                 */
 199                 ret = select(maxfd, readfds2, writefds2, errorfds2, ptval);
 200         } while (ret == -1 && errno == EINTR);
 201
 202         if (readfds)
 203                 *readfds = readfds_buf;
 204         if (writefds)
 205                 *writefds = writefds_buf;
 206         if (errorfds)
 207                 *errorfds = errorfds_buf;
 208
 209         return ret;
 210 }
 211
 212 /*
 213  * sys_poll expects pollfd's to be a talloc'ed array.
 214  *
 215  * It expects the talloc_array_length(fds) >= num_fds+1 to give space
 216  * to the signal pipe.
 217  */
 218
 219 int sys_poll(struct pollfd *fds, int num_fds, int timeout)
 220 {
 221         int ret;
 222
 223         if (talloc_array_length(fds) < num_fds+1) {
 224                 errno = ENOSPC;
 225                 return -1;
 226         }
 227
 228         if (initialised != sys_getpid()) {
 229                 if (pipe(select_pipe) == -1)
 230                 {
 231                         int saved_errno = errno;
 232                         DEBUG(0, ("sys_poll: pipe failed (%s)\n",
 233                                 strerror(errno)));
 234                         errno = saved_errno;
 235                         return -1;
 236                 }
 237
 238                 /*
 239                  * These next two lines seem to fix a bug with the Linux
 240                  * 2.0.x kernel (and probably other UNIXes as well) where
 241                  * the one byte read below can block even though the
 242                  * select returned that there is data in the pipe and
 243                  * the pipe_written variable was incremented. Thanks to
 244                  * HP for finding this one. JRA.
 245                  */
 246
 247                 if(set_blocking(select_pipe[0],0)==-1)
 248                         smb_panic("select_pipe[0]: O_NONBLOCK failed");
 249                 if(set_blocking(select_pipe[1],0)==-1)
 250                         smb_panic("select_pipe[1]: O_NONBLOCK failed");
 251
 252                 initialised = sys_getpid();
 253         }
 254
 255         ZERO_STRUCT(fds[num_fds]);
 256         fds[num_fds].fd = select_pipe[0];
 257         fds[num_fds].events = POLLIN|POLLHUP;
 258
 259         errno = 0;
 260         ret = poll(fds, num_fds+1, timeout);
 261
 262         if ((ret >= 0) && (fds[num_fds].revents & (POLLIN|POLLHUP|POLLERR))) {
 263                 char c;
 264                 int saved_errno = errno;
 265
 266                 if (read(select_pipe[0], &c, 1) == 1) {
 267                         pipe_read += 1;
 268
 269                         /* Mark Weaver <mark-clist@npsl.co.uk> pointed out a critical
 270                            fix to ensure we don't lose signals. We must always
 271                            return -1 when the select pipe is set, otherwise if another
 272                            fd is also ready (so ret == 2) then we used to eat the
 273                            byte in the pipe and lose the signal. JRA.
 274                         */
 275                         ret = -1;
 276 #if 0
 277                         /* JRA - we can use this to debug the signal messaging... */
 278                         DEBUG(0,("select got %u signal\n", (unsigned int)c));
 279 #endif
 280                         errno = EINTR;
 281                 } else {
 282                         ret -= 1;
 283                         errno = saved_errno;
 284                 }
 285         }
 286
 287         return ret;
 288 }
 289
 290 int sys_poll_intr(struct pollfd *fds, int num_fds, int timeout)
 291 {
 292         int orig_timeout = timeout;
 293         struct timespec start;
 294         int ret;
 295
 296         clock_gettime_mono(&start);
 297
 298         while (true) {
 299                 struct timespec now;
 300                 int64_t elapsed;
 301
 302                 ret = poll(fds, num_fds, timeout);
 303                 if (ret != -1) {
 304                         break;
 305                 }
 306                 if (errno != EINTR) {
 307                         break;
 308                 }
 309                 clock_gettime_mono(&now);
 310                 elapsed = nsec_time_diff(&now, &start);
 311                 timeout = (orig_timeout - elapsed) / 1000000;
 312         };
 313         return ret;
 314 }