libudev-compat/src/libudev/libudev-fs.c

/*
 * libudev-fs.c
 * 
 * All modifications to the original source file are:
 *  - Copyright (C) 2022 Aitor Cuadrado Zubizarreta <aitor_czr@gnuinos.org> 
 *
 * Original copyright and license text produced below.
 */

/*
  This file is part of libudev-compat.
   
  Copyright 2015 Jude Nelson (judecn@gmail.com)

  libudev-compat is free software; you can redistribute it and/or modify it
  under the terms of the GNU Lesser General Public License as published by
  the Free Software Foundation; either version 2.1 of the License, or
  (at your option) any later version.

  libudev-compat is distributed in the hope that it will be useful, but
  WITHOUT ANY WARRANTY; without even the implied warranty of
  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
  Lesser General Public License for more details.

  You should have received a copy of the GNU Lesser General Public License
  along with libudev-compat; If not, see <http://www.gnu.org/licenses/>.
*/

#ifndef _POSIX_C_SOURCE
#define _POSIX_C_SOURCE 200809L
#endif 

#ifndef _XOPEN_SOURCE
#define _XOPEN_SOURCE 700
#endif 

#ifndef _BSD_SOURCE
#define _BSD_SOURCE
#endif

#ifndef _GNU_SOURCE 
#define _GNU_SOURCE
#endif

#include <stdio.h>
#include <stdlib.h>
#include <stddef.h>
#include <unistd.h>
#include <stdbool.h>
#include <errno.h>
#include <string.h>
#include <poll.h>
#include <sys/socket.h>
#include <sys/un.h>
#include <netpacket/packet.h>
#include <arpa/inet.h>
#include <linux/netlink.h>
#include <linux/filter.h>
#include <sys/inotify.h>
#include <limits.h>
#include <pthread.h>
#include <sys/stat.h>
#include <sys/types.h>
#include <fcntl.h>
#include <dirent.h>
#include <sys/sendfile.h>
#include <sys/uio.h>
#include <sys/epoll.h>

#include "libudev.h"
#include "libudev-private.h"

#define UDEV_FS_WATCH_DIR_FLAGS (IN_CREATE | IN_ONESHOT)

#ifdef TEST 
#define UDEV_FS_EVENTS_DIR      "/tmp/events"
#else
#define UDEV_FS_EVENTS_DIR      "/dev/metadata/udev/events"
#endif 

// how many monitors can there be per process?
// this is probably big enough.
// email me if it isn't.
#ifndef UDEV_MAX_MONITORS
#define UDEV_MAX_MONITORS 32768
#endif


// We need to make sure that on fork, a udev_monitor listening to the underlying filesystem
// will listen to its *own* process's events directory, at all times.  To do this, we will
// track every monitor that exists, by holding a reference internally.  When the process
// forks, we'll re-initialize all monitors to point to the new process's events directory.
static struct udev_monitor* g_monitor_table[ UDEV_MAX_MONITORS ];
static int g_monitor_lock = 0;  // primitive mutex that can work across forks
static pid_t g_pid = 0;         // tells us when to re-target the monitors


// spin-lock to lock the monitor table
static void g_monitor_table_spinlock() {
   
   bool locked = false;
   while( 1 ) {
      
      locked = __sync_bool_compare_and_swap( &g_monitor_lock, 0, 1 );
      if( locked ) {
         break;
      }
   }
}

// unlock the monitor table 
static void g_monitor_table_unlock() { 
   
   g_monitor_lock = 0;
   
   // make sure spinners see it.
   __sync_synchronize();
}

// on fork(), create a new events directory for each of this process's monitors
// and point them all to them.  This way, both the parent and child can continue 
// to receive device packets.
// NOTE: can only call async-safe methods
static void udev_monitor_atfork(void) {
   
   int errsv = errno;
   int rc = 0;
   int i = 0;
   int cnt = 0;
   pid_t pid = getpid();
   struct udev_monitor* monitor = NULL;
   int socket_fds[2];
   struct epoll_event ev;
   
   write( STDERR_FILENO, "forked! begin split\n", strlen("forked! begin split\n") );
   
   memset( &ev, 0, sizeof(struct epoll_event) );
   
   // reset each monitor's inotify fd to point to a new PID-specific directory instead
   g_monitor_table_spinlock();
   
   if( g_pid != pid ) {
   
      // child; do the fork 
      for( i = 0; i < UDEV_MAX_MONITORS; i++ ) {
         
         if( g_monitor_table[i] == NULL ) {
            continue;
         }
         
         monitor = g_monitor_table[i];
         
         if( monitor->type != UDEV_MONITOR_TYPE_UDEV ) {
            continue;
         }
         
         if( monitor->inotify_fd < 0 ) {
            continue;
         }
         
         if( monitor->epoll_fd < 0 ) {
            continue;
         }
         
         if( monitor->events_wd < 0 ) {
            continue;
         }
            
         // reset the socket buffer--the parent will be said to have 
         // received intermittent events before the child was created.
         if( monitor->sock >= 0 ) {
            
            // stop watching this socket--we'll regenerate it later
            epoll_ctl( monitor->epoll_fd, EPOLL_CTL_DEL, monitor->sock, NULL );
            close( monitor->sock );
            monitor->sock = -1;
         }
         
         if( monitor->sock_fs >= 0 ) {
            close( monitor->sock_fs );
            monitor->sock_fs = -1;
         }
         
         rc = socketpair( AF_LOCAL, SOCK_RAW | SOCK_NONBLOCK | SOCK_CLOEXEC, 0, socket_fds );
         if( rc < 0 ) {
            
            // not much we can do here, except log an error 
            write( STDERR_FILENO, "Failed to generate a socketpair\n", strlen( "Failed to generate a socketpair\n" ) );

            udev_monitor_fs_shutdown( monitor );
            g_monitor_table[i] = NULL;
            continue;
         }
         
         // child's copy of the monitor has its own socketpair 
         monitor->sock = socket_fds[0];
         monitor->sock_fs = socket_fds[1];
         
         // reinstall its filter 
         udev_monitor_filter_update( monitor );
         
         // watch the child's socket 
         ev.events = EPOLLIN;
         ev.data.fd = monitor->sock;
         rc = epoll_ctl( monitor->epoll_fd, EPOLL_CTL_ADD, monitor->sock, &ev );
         if( rc < 0 ) {
            
            // not much we can do here, except log an error 
            write( STDERR_FILENO, "Failed to add monitor socket\n", strlen("Failed to add monitor socket\n") );
          
            udev_monitor_fs_shutdown( monitor );
            g_monitor_table[i] = NULL;
            continue;
         }
         
         // reset the inotify watch
         rc = inotify_rm_watch( monitor->inotify_fd, monitor->events_wd );
         monitor->events_wd = -1;
         
         if( rc < 0 ) {
            
            rc = -errno;
            if( rc == -EINVAL ) {
               
               // monitor->events_wd was invalid 
               rc = 0;
            }
            else if( rc == -EBADF ) {
               
               // monitor->inotify_fd is invalid.
               // not much we can do here, except log an error 
               write( STDERR_FILENO, "Invalid inotify handle\n", strlen("Invalid inotify handle"));
               
               udev_monitor_fs_shutdown( monitor );
               g_monitor_table[i] = NULL;
               continue;
            }
         }
         
         if( rc == 0 ) {
            
            udev_monitor_fs_events_path( "", monitor->events_dir, i );
            
            // try to create a new directory for this monitor
            rc = mkdir( monitor->events_dir, 0700 );
   
            if( rc < 0 ) {
               // failed, we have.
               // child will not get any notifications from this monitor
               rc = -errno;
               write( STDERR_FILENO, "Failed to mkdir ", strlen("Failed to mkdir ") );
               write( STDERR_FILENO, monitor->events_dir, strlen(monitor->events_dir) );
               write( STDERR_FILENO, "\n", 1 );
      
               udev_monitor_fs_shutdown( monitor );
               g_monitor_table[i] = NULL;
               return;
            }
            
            // reconnect to the new directory
            monitor->events_wd = inotify_add_watch( monitor->inotify_fd, monitor->events_dir, UDEV_FS_WATCH_DIR_FLAGS );
            if( monitor->events_wd < 0 ) {
               
               // there's not much we can safely do here, besides log an error
               write( STDERR_FILENO, "Failed to watch ", strlen( "Failed to watch " ) );
               write( STDERR_FILENO, monitor->events_dir, strlen(monitor->events_dir) );
               write( STDERR_FILENO, "\n", 1 );
            }
         }
         else {
            
            // there's not much we can safely do here, besides log an error
            rc = -errno;
            write( STDERR_FILENO, "Failed to disconnect!\n", strlen("Failed to disconnect!\n") );
         }
      }
      
      g_pid = pid;
   }
      
   g_monitor_table_unlock();
   
   write( STDERR_FILENO, "end atfork()\n", strlen("end atfork()\n") );
   
   // restore...
   errno = errsv;
}


// register a monitor in our list of monitors 
// return 0 on success
// return -ENOSPC if we're out of slots
static int udev_monitor_register( struct udev_monitor* monitor ) {
   
   g_monitor_table_spinlock();
   
   // find a free slot 
   int i = 0;
   int rc = -ENOSPC;
   for( i = 0; i < UDEV_MAX_MONITORS; i++ ) {
      
      if( g_monitor_table[i] == NULL ) {
         
         g_monitor_table[i] = monitor;
         monitor->slot = i;
         rc = 0;
         break;
      }
   }
   
   if( g_pid == 0 ) {
      
      // first monitor ever.
      // register our fork handler.
      g_pid = getpid();
      pthread_atfork( NULL, NULL, udev_monitor_atfork );
   }
   
   g_monitor_table_unlock();

   return rc;   
}


// unregister a monitor in our list of monitors 
static void udev_monitor_unregister( struct udev_monitor* monitor ) {
   
   if( monitor->slot < 0 ) {
      return;
   }
   
   g_monitor_table_spinlock();
     
   g_monitor_table[ monitor->slot ] = NULL;
      
   g_monitor_table_unlock();
      
   monitor->slot = -1;
}


// write, but mask EINTR
// return number of bytes written on success 
// return -errno on I/O error
ssize_t udev_write_uninterrupted( int fd, char const* buf, size_t len ) {
   
   ssize_t num_written = 0;
   
   if( buf == NULL ) {
      return -EINVAL;
   }
   
   while( (unsigned)num_written < len ) {
      ssize_t nw = write( fd, buf + num_written, len - num_written );
      if( nw < 0 ) {
         
         int errsv = -errno;
         if( errsv == -EINTR ) {
            continue;
         }
         
         return errsv;
      }
      if( nw == 0 ) {
         break;
      }
      
      num_written += nw;
   }
   
   return num_written;
}


// read, but mask EINTR
// return number of bytes read on success 
// return -errno on I/O error
// NOTE: must be async-safe!
ssize_t udev_read_uninterrupted( int fd, char* buf, size_t len ) {
   
   ssize_t num_read = 0;
   
   if( buf == NULL ) {
      return -EINVAL;
   }
   
   while( (unsigned)num_read < len ) {
      ssize_t nr = read( fd, buf + num_read, len - num_read );
      if( nr < 0 ) {
         
         int errsv = -errno;
         if( errsv == -EINTR ) {
            continue;
         }
         
         return errsv;
      }
      if( nr == 0 ) {
         break;
      }
      
      num_read += nr;
   }
   
   return num_read;
}


// set up a filesystem monitor
// register pthread_atfork() handlers to ensure that its children 
// get their own filesystem monitor state.
// * set up /dev/events/libudev-$PID/
// * start watching /dev/events/libudev-$PID for new files
_public_ int udev_monitor_fs_setup( struct udev_monitor* monitor ) {
   
   int rc = 0;
   struct epoll_event ev;
   
   monitor->inotify_fd = -1;
   monitor->epoll_fd = -1;
   monitor->sock = -1;
   monitor->sock_fs = -1;
   monitor->slot = -1;
   
   int socket_fd[2] = { -1, -1 };
   
   memset( &monitor->events_dir, 0, PATH_MAX+1 );
   memset( &ev, 0, sizeof(struct epoll_event) );
   
   // make sure this monitor can't disappear on us
   udev_monitor_register( monitor );
   
   // set up inotify 
   monitor->inotify_fd = inotify_init1( IN_NONBLOCK | IN_CLOEXEC );
   if( monitor->inotify_fd < 0 ) {
      
      rc = -errno;
      log_error("inotify_init rc = %d", rc );
      
      udev_monitor_fs_shutdown( monitor );
      return rc;
   }
   
   // epoll descriptor unifying inotify and event counter
   monitor->epoll_fd = epoll_create1( EPOLL_CLOEXEC );
   if( monitor->epoll_fd < 0 ) {
      
      rc = -errno;
      log_error("epoll_create rc = %d", rc );
      
      udev_monitor_fs_shutdown( monitor );
      return rc;
   }
   
   // create our monitor directory /dev/events/libudev-$PID
   udev_monitor_fs_events_path( "", monitor->events_dir, monitor->slot );
   rc = mkdir( monitor->events_dir, 0700 );
   if( rc != 0 ) {
      
      rc = -errno;
      log_error("mkdir('%s') rc = %d", monitor->events_dir, rc );
      
      udev_monitor_fs_destroy( monitor );
      return rc;
   }
   
   // begin watching /dev/events/libudev-$PID
   monitor->events_wd = inotify_add_watch( monitor->inotify_fd, monitor->events_dir, UDEV_FS_WATCH_DIR_FLAGS );
   if( monitor->events_wd < 0 ) {
      
      rc = -errno;
      log_error("inotify_add_watch('%s') rc = %d", monitor->events_dir, rc );
      
      udev_monitor_fs_destroy( monitor );
      return rc;
   }
   
   // set up local socket pair with the parent process
   // needs to be a socket (not a pipe) since we're going to attach a BPF to it.
   rc = socketpair( AF_LOCAL, SOCK_RAW | SOCK_NONBLOCK | SOCK_CLOEXEC, 0, socket_fd );
   if( rc < 0 ) {
      
      rc = -errno;
      log_error("socketpair(AF_LOCAL, SOCK_RAW|SOCK_CLOEXEC, 0) rc = %d", rc );
      
      udev_monitor_fs_destroy( monitor );
      return rc;
   }
   
   monitor->sock = socket_fd[0];                // receiving end 
   monitor->sock_fs = socket_fd[1];             // sending end
   
   // unify inotify and sock behind epoll_fd, set to poll for reading
   ev.events = EPOLLIN;
   ev.data.fd = monitor->inotify_fd;
   rc = epoll_ctl( monitor->epoll_fd, EPOLL_CTL_ADD, monitor->inotify_fd, &ev );
   if( rc != 0 ) {
      
      rc = -errno;
      log_error("epoll_ctl(%d on inotify_fd %d) rc = %d", monitor->epoll_fd, monitor->inotify_fd, rc );
      
      udev_monitor_fs_shutdown( monitor );
      return rc;
   }
   
   ev.data.fd = monitor->sock;
   rc = epoll_ctl( monitor->epoll_fd, EPOLL_CTL_ADD, monitor->sock, &ev );
   if( rc != 0 ) {
      
      rc = -errno;
      log_error("epoll_ctl(%d on inotify_fd %d) rc = %d", monitor->sock, monitor->sock, rc );
      
      udev_monitor_fs_shutdown( monitor );
      return rc;
   }
   
   monitor->pid = getpid();
   
   return rc;
}


// shut down a monitor's filesystem-specific state 
// not much we can do if any shutdown step fails, so try them all
_public_ int udev_monitor_fs_shutdown( struct udev_monitor* monitor ) {
   
   int rc = 0;
   
   // stop tracking this monitor
   udev_monitor_unregister( monitor );
   
   if( monitor->sock >= 0 ) {
      rc = shutdown( monitor->sock, SHUT_RDWR );
      if( rc < 0 ) {
         rc = -errno;
         log_error("shutdown(socket %d) rc = %d", monitor->sock, rc );
      }
   }
      
   if( monitor->sock_fs >= 0 ) {
      rc = shutdown( monitor->sock_fs, SHUT_RDWR );
      if( rc < 0 ) {
         rc = -errno;
         log_error("shutdown(socket %d) rc = %d", monitor->sock_fs, rc );
      }
   }
   
   if( monitor->sock >= 0 ) {
      rc = close( monitor->sock );
      if( rc < 0 ) {
         rc = -errno;
         log_error("close(socket %d) rc = %d", monitor->sock, rc );
      }
      else {
         monitor->sock = -1;
      }
   }
    
   if( monitor->sock_fs >= 0 ) {
      rc = close( monitor->sock_fs );
      if( rc < 0 ) {
         rc = -errno;
         log_error("close(socket %d) rc = %d", monitor->sock_fs, rc );
      }
      else {
         monitor->sock_fs = -1;
      }
   }
   
   if( monitor->epoll_fd >= 0 ) {
      rc = close( monitor->epoll_fd );
      if( rc < 0 ) {
         rc = -errno;
         log_error("close(epoll_fd %d) rc = %d", monitor->epoll_fd, rc );
      }
      else {
         monitor->epoll_fd = -1;
      }
   }
  
   if( monitor->events_wd >= 0 ) {

      if( monitor->inotify_fd >= 0 ) {
         rc = inotify_rm_watch( monitor->inotify_fd, monitor->events_wd );
         if( rc < 0 ) {
            rc = -errno;
            log_error("close(events_wd %d) rc = %d", monitor->events_wd, rc );
         }
         else {
            monitor->events_wd = -1;
         }
      }
   } 
   if( monitor->inotify_fd >= 0 ) {
      rc = close( monitor->inotify_fd );
      if( rc < 0 ) {
         rc = -errno;
         log_error("close(inotify_fd %d) rc = %d", monitor->inotify_fd, rc );
      }
      else {
         monitor->inotify_fd = -1;
      }
   }
   
   return rc;
}


// blow away all local filesystem state for a monitor
_public_ int udev_monitor_fs_destroy( struct udev_monitor* monitor ) {
   
   char pathbuf[ PATH_MAX+1 ];
   int dirfd = 0;
   int rc = 0;
   DIR* dirh = NULL;
   struct dirent entry;
   struct dirent* result = NULL;
   bool can_rmdir = true;
   
   // stop listening
   udev_monitor_fs_shutdown( monitor );
   
   // remove events dir contents 
   dirfd = open( monitor->events_dir, O_DIRECTORY | O_CLOEXEC );
   if( dirfd < 0 ) {
      
      rc = -errno;
      log_error("open('%s') rc = %d", monitor->events_dir, rc );
      return rc;
   }
   
   dirh = fdopendir( dirfd );
   if( dirh == NULL ) {
      
      // OOM
      rc = -errno;
      close( dirfd );
      return rc;
   }
   
   do {
      
      // next entry 
      rc = readdir_r( dirh, &entry, &result );
      if( rc != 0 ) {
         
         // I/O error 
         log_error("readdir_r('%s') rc = %d", monitor->events_dir, rc );
         break;
      }
      
      // skip . and ..
      if( strcmp( entry.d_name, "." ) == 0 || strcmp( entry.d_name, ".." ) == 0 ) {
         continue;
      }
      
      // generate full path 
      memset( pathbuf, 0, PATH_MAX+1 );
      
      snprintf( pathbuf, PATH_MAX, "%s/%s", monitor->events_dir, entry.d_name );
      
      // optimistically remove
      if( entry.d_type == DT_DIR ) {
         rc = rmdir( pathbuf );
      }
      else {
         rc = unlink( pathbuf );
      }
      
      if( rc != 0 ) {
         
         rc = -errno;
         log_error("remove '%s' rc = %d", pathbuf, rc );
         
         can_rmdir = false;
         rc = 0;
      }
      
   } while( result != NULL );
  
   // NOTE: closes dirfd
   closedir( dirh );
   
   if( can_rmdir ) {
      rc = rmdir( monitor->events_dir );
      if( rc != 0 ) {
         
         rc = -errno;
         log_error("rmdir('%s') rc = %d\n", monitor->events_dir, rc);
      }  
   }
   else {
      // let the caller know...
      rc = -ENOTEMPTY;
   }
   
   return rc;
}


// thread-safe and async-safe int to string for base 10
void itoa10_safe( int val, char* str ) {
   
   int rc = 0;
   int i = 0;
   int len = 0;
   int j = 0;
   bool neg = false;
   
   // sign check
   if( val < 0 ) {
      val = -val;
      neg = true;
   }
   
   // consume, lowest-order to highest-order
   if( val == 0 ) {
      str[i] = '0';
      i++;
   }
   else {
      while( val > 0 ) {
      
         int r = val % 10;
      
         str[i] = '0' + r;
         i++;
      
         val /= 10;
      }
   }
   
   if( neg ) {
      
      str[i] = '-';
      i++;
   }
   
   len = i;
   i--;
   
   // reverse order to get the number 
   while( j < i ) {
      
      char tmp = *(str + i);
      *(str + i) = *(str + j);
      *(str + j) = tmp;
      
      j++;
      i--;
   }
   
   str[len] = '\0';
}

// path to a named event for this process to consume
// pathbuf must have at least PATH_MAX bytes
// NOTE: must be async-safe, since it's used in a pthread_atfork() callback
_public_ int udev_monitor_fs_events_path( char const* name, char* pathbuf, int nonce ) {
   
   // do the equivalent of:
   //    snprintf( pathbuf, PATH_MAX, UDEV_FS_EVENTS_DIR "/libudev-%d-%d/%s", getpid(), nonce, name );
   
   char pidbuf[10];
   pidbuf[9] = 0;

   char nonce_buf[50];
   nonce_buf[49] = 0;
   
   pid_t pid = getpid();
   
   itoa10_safe( pid, pidbuf );
   itoa10_safe( nonce, nonce_buf );
   
   size_t pidbuf_len = strnlen(pidbuf, 10);
   size_t nonce_buf_len = strnlen( nonce_buf, 50 );
   size_t prefix_len = strlen( UDEV_FS_EVENTS_DIR );
   size_t dirname_prefix_len = strlen( "/libudev-" );
   int off = 0;
   
   memcpy( pathbuf, UDEV_FS_EVENTS_DIR, prefix_len );
   off += prefix_len;
   
   memcpy( pathbuf + off, "/libudev-", dirname_prefix_len );
   off += dirname_prefix_len;
   
   memcpy( pathbuf + off, pidbuf, pidbuf_len );
   off += pidbuf_len;

   memcpy( pathbuf + off, "-", 1 );
   off += 1;

   memcpy( pathbuf + off, nonce_buf, nonce_buf_len );
   off += nonce_buf_len;
   
   pathbuf[off] = '/';
   off++;
   
   memcpy( pathbuf + off, name, strlen(name) );
   off += strlen(name);
   
   pathbuf[off] = '\0';
   
   return off + strlen(name);
}

// send the file descriptor containing a serialized packet to the libudev client:
// NOTE: The file format is expected to be the same as a uevent packet:
//       * all newlines (\n) will be converted to null (\0) in the function
//         static struct udev_device *udev_monitor_receive_device_fs(struct udev_monitor *udev_monitor)
//         of libudev-minitor.c, since that's how the kernel sends it.
//       * the buffer is expected to be at most 8192 bytes long.
// return 0 on success 
// return -errno on failure
_public_ int udev_monitor_fs_push_event( int fd, struct udev_monitor* monitor ) {
   
   int rc = 0;
   off_t offset = 0;
   
   // send it along
   rc = sendfile(monitor->sock_fs, fd, &offset, 8192);
   if( rc < 0 ) {
      
      rc = -errno;
      log_error("udev_monitor_send_device rc = %d", rc );
   }
   else {
      
      rc = 0;
   }
   
   return rc;
}


// reset the oneshot inotify watch, so it will trip on the next create.
// consume pending events, if there are any, and re-watch the directory.
// if the pid has changed since last time, watch the new directory.
// return 0 on success
// return negative on error (errno)
static int udev_monitor_fs_watch_reset( struct udev_monitor* monitor ) {
   
   int rc = 0;
   bool inotify_triggerred = false;     // was inotify triggerred?
   char buf[4096] __attribute__ ((aligned(__alignof__(struct inotify_event)))); // see inotify(7)
   struct pollfd pfd[1];
   
   pfd[0].fd = monitor->inotify_fd;
   pfd[0].events = POLLIN;
   
   // reset the watch by consuming all its events (should be at most one)
   while( 1 ) {
      
      // do we have data?
      rc = poll( pfd, 1, 0 );
      if( rc <= 0 ) {
         
         // out of data, or error 
         if( rc < 0 ) {
            
            rc = -errno;
            if( rc == -EINTR ) {
               
               // shouldn't really happen since the timeout for poll is zero,
               // but you never know...
               continue;
            }
            
            log_error("poll(%d) rc = %d\n", monitor->inotify_fd, rc );
            rc = 0;
         }
         
         break;
      }
      
      // at least one event remaining
      // consume it
      rc = read( monitor->inotify_fd, buf, 4096 );
      if( rc == 0 ) {
         break;
      }
      
      if( rc < 0 ) {
         
         rc = -errno;
         
         if( rc == -EINTR ) {
            continue;
         }
         else if( rc == -EAGAIN || rc == -EWOULDBLOCK ) {
            rc = 0;
            break;
         }
      }
      
      // got one event
      inotify_triggerred = true;
   }

   
   // has the PID changed?
   // need to regenerate events path 
   if( getpid() != monitor->pid ) {

      log_trace("Switch PID from %d to %d", monitor->pid, getpid());

      udev_monitor_fs_events_path( "", monitor->events_dir, monitor->slot );
      
      rc = mkdir( monitor->events_dir, 0700 );
      if( rc != 0 ) {
         
         rc = -errno;
         if( rc != -EEXIST ) {
            
            log_error("mkdir('%s') rc = %d\n", monitor->events_dir, rc );
            return rc;
         }
         else {
            rc = 0;
         }
      }
      
      monitor->events_wd = inotify_add_watch( monitor->inotify_fd, monitor->events_dir, UDEV_FS_WATCH_DIR_FLAGS );
      if( monitor->events_wd < 0 ) {
         
         rc = -errno;
         
         log_error("inotify_add_watch('%s') rc = %d\n", monitor->events_dir, rc );
         return rc;
      }
      
      monitor->pid = getpid();
      
      // TODO: what about events that the child was supposed to receive?
      // the parent forks, receives one or more events, and the child wakes up, and will miss them 
      // if we only do the above.
      // we need to (try to) consume them here
   }

   rc = inotify_add_watch( monitor->inotify_fd, monitor->events_dir, UDEV_FS_WATCH_DIR_FLAGS );
   if( rc < 0 ) {
      
      rc = -errno;
      log_error("inotify_add_watch(%d) rc = %d", monitor->inotify_fd, rc );
   }
   
   return rc;
}

// scandir: skip . and ..
static int udev_monitor_fs_scandir_filter( const struct dirent* dent ) {
   
   if( strcmp( dent->d_name, "." ) == 0 || strcmp( dent->d_name, ".." ) == 0 ) {
      return 0;
   }
   else {
      return 1;
   }
}


// push as many available events from our filesystem buffer of events off to the monitor's socketpair.
// the order is determined lexicographically (i.e. we assume that whatever is writing events is naming them in 
// a monotonically increasing order, such as e.g. the SEQNUM field)
// return 0 on success
// return -ENODATA if there are no events
// return -EAGAIN if there are events, but we couldn't push any
// return -errno if we can't re-watch the directory
// NOTE: this method should only be used if the underlying filesystem holding the events can't help us preserve the order.
// NOTE: not thread-safe
int udev_monitor_fs_push_events( struct udev_monitor* monitor ) {
   
   char pathbuf[ PATH_MAX+1 ];
   int dirfd = -1;
   int fd = -1;
   int rc = 0;
   int num_events = 0;
   int num_valid_events = 0;
   int num_sent = 0;
   struct dirent** events = NULL;        // names of events to buffer
   int i = 0;
   
   // reset the watch on this directory, and ensure we're watching the right one.
   rc = udev_monitor_fs_watch_reset( monitor );
   if( rc < 0 ) {
      
      log_error("Failed to re-watch '%s', rc = %d", monitor->events_dir, rc );
      goto udev_monitor_fs_push_events_cleanup;
   }
   
   // find new events... 
   dirfd = open( monitor->events_dir, O_DIRECTORY | O_CLOEXEC );
   if( dirfd < 0 ) {
      
      rc = -errno;
      log_error("open('%s') rc = %d", monitor->events_dir, rc );
      goto udev_monitor_fs_push_events_cleanup;
   }
   
   num_events = scandirat( dirfd, ".", &events, udev_monitor_fs_scandir_filter, alphasort );
   
   if( num_events < 0 ) {
      
      rc = -errno;
      log_error("scandir('%s') rc = %d", monitor->events_dir, rc );
      goto udev_monitor_fs_push_events_cleanup;
   }
   
   
   if( num_events == 0 ) {
      
      // got nothing
      rc = -ENODATA;
      goto udev_monitor_fs_push_events_cleanup;
   }
   
   num_valid_events = num_events;
   
   // send them all off!
   for( i = 0; i < num_events; i++ ) {
      
      snprintf( pathbuf, PATH_MAX, "%s/%s", monitor->events_dir, events[i]->d_name );
      
      fd = open( pathbuf, O_RDONLY | O_CLOEXEC );
      if( fd < 0 ) {
         
         rc = -errno;
         log_error("cannot open event: open('%s') rc = %d", pathbuf, rc );
         
         // we consider it more important to preserve order and drop events 
         // than to try to resend later.
         unlink( pathbuf );
      }
      else {
            
         // propagate to the monitor's socket
         rc = udev_monitor_fs_push_event( fd, monitor );
         
         // garbage-collect
         close( fd );
         unlink( pathbuf );
         
         if( rc == -EBADMSG || rc == -EMSGSIZE ) {
            
            // invalid message anyway
            rc = 0;
            num_valid_events--;
            continue;
         }
            
         else if( rc < 0 ) {
            
            if( rc != -EAGAIN ) {
               
               // socket-level error
               log_error("failed to push event '%s', rc = %d", pathbuf, rc );
               break;
            }
            else {
               
               // sent as many as we could 
               rc = 0;
               break;
            }
         }
         else if( rc == 0 ) {
            
            num_sent++;
         }
      }
   }
   
udev_monitor_fs_push_events_cleanup:
 
   if( dirfd >= 0 ) { 
      close( dirfd );
   }

   if( events != NULL ) {
      for( i = 0; i < num_events; i++ ) {
         
         if( events[i] != NULL ) {
            free( events[i] );
            events[i] = NULL;
         }
      }
      
      free( events );
   }
   
   if( num_sent == 0 && num_valid_events > 0 ) {
      
      // there are pending events, but we couldn't push any.
      rc = -EAGAIN;
   }
   
   return rc;
}


#ifdef TEST 

#include "libudev.h"

int main( int argc, char** argv ) {
   
   int rc = 0;
   char pathbuf[PATH_MAX+1];
   struct udev* udev_client = NULL;
   struct udev_monitor* monitor = NULL;
   int monitor_fd = 0;
   struct udev_device* dev = NULL;
   struct pollfd pfd[1];
   int num_events = INT_MAX;
   int num_forks = 0;
   
   // usage: $0 [num events to process [num times to fork]]
   if( argc > 1 ) {
      char* tmp = NULL;
      num_events = (int)strtol( argv[1], &tmp, 10 );
      if( tmp == argv[1] || *tmp != '\0' ) {
         fprintf(stderr, "Usage: %s [number of events to process [number of times to fork]]\n", argv[0] );
         exit(1);
      }
      
      if( argc > 2 ) {
         
         num_forks = (int)strtol( argv[2], &tmp, 10 );
         if( tmp == argv[2] || *tmp != '\0' ) {
            fprintf(stderr, "Usage: %s [number of events to process [number of times to fork]]\n", argv[0] );
            exit(1);
         }
      }
   }
   
   // make sure events dir exists 
   log_trace("events directory '%s'", UDEV_FS_EVENTS_DIR);
   
   rc = mkdir( UDEV_FS_EVENTS_DIR, 0700 );
   if( rc != 0 ) {
      
      rc = -errno;
      if( rc != -EEXIST ) {
         log_error("mkdir('%s') rc = %d", UDEV_FS_EVENTS_DIR, rc );
         exit(1);
      }
   }
   
   udev_monitor_fs_events_path( "", pathbuf, 0 );
   
   printf("Watching '%s'\n", pathbuf );
   
   udev_client = udev_new();
   if( udev_client == NULL ) {
      
      // OOM
      exit(2);
   }
   
   monitor = udev_monitor_new_from_netlink( udev_client, "udev" );
   if( monitor == NULL ) {
      
      // OOM or error
      udev_unref( udev_client );
      exit(2);
   }
   
   printf("Press Ctrl-C to quit\n");
   
   monitor_fd = udev_monitor_get_fd( monitor );
   if( monitor_fd < 0 ) {
      
      rc = -errno;
      log_error("udev_monitor_get_fd rc = %d\n", rc );
      exit(3);
   }
   
   pfd[0].fd = monitor_fd;
   pfd[0].events = POLLIN;
   
   while( num_events > 0 ) {
      
      // wait for the next device 
      rc = poll( pfd, 1, -1 );
      if( rc < 0 ) {
      
         log_error("poll(%d) rc = %d\n", monitor_fd, rc );
         break;
      }
      
      // get devices 
      while( num_events > 0 ) {
         
         dev = udev_monitor_receive_device( monitor );
         if( dev == NULL ) {
            break;
         }
         
         int pid = getpid();
         struct udev_list_entry *list_entry = NULL;
         
         printf("[%d] [%d] ACTION:     '%s'\n", pid, num_events, udev_device_get_action( dev ) );
         printf("[%d] [%d] SEQNUM:      %llu\n", pid, num_events, udev_device_get_seqnum( dev ) );
         printf("[%d] [%d] USEC:        %llu\n", pid, num_events, udev_device_get_usec_since_initialized( dev ) );
         printf("[%d] [%d] DEVNODE:    '%s'\n", pid, num_events, udev_device_get_devnode( dev ) );
         printf("[%d] [%d] DEVPATH:    '%s'\n", pid, num_events, udev_device_get_devpath( dev ) );
         printf("[%d] [%d] SYSNAME:    '%s'\n", pid, num_events, udev_device_get_sysname( dev ) );
         printf("[%d] [%d] SYSPATH:    '%s'\n", pid, num_events, udev_device_get_syspath( dev ) );
         printf("[%d] [%d] SUBSYSTEM:  '%s'\n", pid, num_events, udev_device_get_subsystem( dev ) );
         printf("[%d] [%d] DEVTYPE:    '%s'\n", pid, num_events, udev_device_get_devtype( dev ) );
         printf("[%d] [%d] SYSNUM:     '%s'\n", pid, num_events, udev_device_get_sysnum( dev ) );
         printf("[%d] [%d] DRIVER:     '%s'\n", pid, num_events, udev_device_get_driver( dev ) );
         printf("[%d] [%d] DEVNUM:      %d:%d\n", pid, num_events, major( udev_device_get_devnum( dev ) ), minor( udev_device_get_devnum( dev ) ) );
         printf("[%d] [%d] IFINDEX:    '%s'\n", pid, num_events, udev_device_get_property_value( dev, "IFINDEX" ) );
         printf("[%d] [%d] DEVMODE:    '%s'\n", pid, num_events, udev_device_get_property_value( dev, "DEVMODE" ) );
         printf("[%d] [%d] DEVUID:     '%s'\n", pid, num_events, udev_device_get_property_value( dev, "DEVUID" ) );
         printf("[%d] [%d] DEVGID:     '%s'\n", pid, num_events, udev_device_get_property_value( dev, "DEVGID" ) );         
         
         list_entry = udev_device_get_devlinks_list_entry( dev );
         udev_list_entry_foreach( list_entry, udev_list_entry_get_next( list_entry )) {
          
            printf("[%d] [%d] devlink:    '%s'\n", pid, num_events, udev_list_entry_get_name( list_entry ) );
         }
         
         list_entry = udev_device_get_properties_list_entry( dev );
         udev_list_entry_foreach( list_entry, udev_list_entry_get_next( list_entry )) {
            
            printf("[%d] [%d] property:   '%s' = '%s'\n", pid, num_events, udev_list_entry_get_name( list_entry ), udev_list_entry_get_value( list_entry ) );
         }
         
         list_entry = udev_device_get_tags_list_entry( dev );
         udev_list_entry_foreach( list_entry, udev_list_entry_get_next( list_entry )) {
            
            printf("[%d] [%d] tag:        '%s'\n", pid, num_events, udev_list_entry_get_name( list_entry ) );
         }
         
         list_entry = udev_device_get_sysattr_list_entry( dev );
         udev_list_entry_foreach( list_entry, udev_list_entry_get_next( list_entry )) {
            
            printf("[%d] [%d] sysattr:    '%s'\n", pid, num_events, udev_list_entry_get_name( list_entry ) );
         }
         
         printf("\n");
         
         udev_device_unref( dev );
         
         num_events--;
      }
      
      // do our forks
      if( num_forks > 0 ) {
         
         num_forks--;
         
         int pid = fork();
         if( pid < 0 ) {
            
            rc = -errno;
            fprintf(stderr, "fork: %s\n", strerror( -rc ) );
            break;
         }
         else if( pid == 0 ) {
            
            printf("[%d]\n", getpid() );
         }
      }
   }
   
   udev_monitor_fs_destroy( monitor );
   
   exit(0);
}

#endif