]> The Tcpdump Group git mirrors - libpcap/blobdiff - pcap-linux.c
Add support for libnl 2.x, adapted from a newer version of the iw command.
[libpcap] / pcap-linux.c
index 6fe7a24f349f0076fcd7fb806f101667a7a96b85..deabbc4a986a2232d9515207e259c7f6d70ce47d 100644 (file)
@@ -122,6 +122,7 @@ static const char rcsid[] _U_ =
 #include <errno.h>
 #include <stdio.h>
 #include <stdlib.h>
+#include <ctype.h>
 #include <unistd.h>
 #include <fcntl.h>
 #include <string.h>
@@ -135,6 +136,12 @@ static const char rcsid[] _U_ =
 #include <linux/if_ether.h>
 #include <net/if_arp.h>
 #include <poll.h>
+#include <dirent.h>
+
+#ifdef HAVE_LINUX_NET_TSTAMP_H
+#include <linux/net_tstamp.h>
+#include <linux/sockios.h>
+#endif
 
 /*
  * Got Wireless Extensions?
@@ -168,6 +175,10 @@ static const char rcsid[] _U_ =
 #include "pcap-septel.h"
 #endif /* HAVE_SEPTEL_API */
 
+#ifdef HAVE_SNF_API
+#include "pcap-snf.h"
+#endif /* HAVE_SNF_API */
+
 #ifdef PCAP_SUPPORT_USB
 #include "pcap-usb-linux.h"
 #endif
@@ -176,6 +187,10 @@ static const char rcsid[] _U_ =
 #include "pcap-bt-linux.h"
 #endif
 
+#ifdef PCAP_SUPPORT_CAN
+#include "pcap-can-linux.h"
+#endif
+
 /*
  * If PF_PACKET is defined, we can use {SOCK_RAW,SOCK_DGRAM}/PF_PACKET
  * sockets rather than SOCK_PACKET sockets.
@@ -285,7 +300,7 @@ static short int map_packet_type_to_sll_type(short int);
 static int pcap_activate_linux(pcap_t *);
 static int activate_old(pcap_t *);
 static int activate_new(pcap_t *);
-static int activate_mmap(pcap_t *);
+static int activate_mmap(pcap_t *, int *);
 static int pcap_can_set_rfmon_linux(pcap_t *);
 static int pcap_read_linux(pcap_t *, int, pcap_handler, u_char *);
 static int pcap_read_packet(pcap_t *, pcap_handler, u_char *);
@@ -305,13 +320,15 @@ union thdr {
 #define RING_GET_FRAME(h) (((union thdr **)h->buffer)[h->offset])
 
 static void destroy_ring(pcap_t *handle);
-static int create_ring(pcap_t *handle);
+static int create_ring(pcap_t *handle, int *status);
 static int prepare_tpacket_socket(pcap_t *handle);
 static void pcap_cleanup_linux_mmap(pcap_t *);
 static int pcap_read_linux_mmap(pcap_t *, int, pcap_handler , u_char *);
 static int pcap_setfilter_linux_mmap(pcap_t *, struct bpf_program *);
 static int pcap_setnonblock_mmap(pcap_t *p, int nonblock, char *errbuf);
 static int pcap_getnonblock_mmap(pcap_t *p, char *errbuf);
+static void pcap_oneshot_mmap(u_char *user, const struct pcap_pkthdr *h,
+    const u_char *bytes);
 #endif
 
 /*
@@ -333,7 +350,8 @@ static int  enter_rfmon_mode(pcap_t *handle, int sock_fd,
 static int     iface_bind_old(int fd, const char *device, char *ebuf);
 
 #ifdef SO_ATTACH_FILTER
-static int     fix_program(pcap_t *handle, struct sock_fprog *fcode);
+static int     fix_program(pcap_t *handle, struct sock_fprog *fcode,
+    int is_mapped);
 static int     fix_offset(struct bpf_insn *p);
 static int     set_kernel_filter(pcap_t *handle, struct sock_fprog *fcode);
 static int     reset_kernel_filter(pcap_t *handle);
@@ -367,12 +385,25 @@ pcap_create(const char *device, char *ebuf)
        }
 #endif /* HAVE_SEPTEL_API */
 
+#ifdef HAVE_SNF_API
+        handle = snf_create(device, ebuf);
+        if (strstr(device, "snf") || handle != NULL)
+               return handle;
+
+#endif /* HAVE_SNF_API */
+
 #ifdef PCAP_SUPPORT_BT
        if (strstr(device, "bluetooth")) {
                return bt_create(device, ebuf);
        }
 #endif
 
+#ifdef PCAP_SUPPORT_CAN
+       if (strstr(device, "can") || strstr(device, "vcan")) {
+               return can_create(device, ebuf);
+       }
+#endif
+
 #ifdef PCAP_SUPPORT_USB
        if (strstr(device, "usbmon")) {
                return usb_create(device, ebuf);
@@ -385,52 +416,73 @@ pcap_create(const char *device, char *ebuf)
 
        handle->activate_op = pcap_activate_linux;
        handle->can_set_rfmon_op = pcap_can_set_rfmon_linux;
+#if defined(HAVE_LINUX_NET_TSTAMP_H) && defined(PACKET_TIMESTAMP)
+       /*
+        * We claim that we support:
+        *
+        *      software time stamps, with no details about their precision;
+        *      hardware time stamps, synced to the host time;
+        *      hardware time stamps, not synced to the host time.
+        *
+        * XXX - we can't ask a device whether it supports
+        * hardware time stamps, so we just claim all devices do.
+        */
+       handle->tstamp_type_count = 3;
+       handle->tstamp_type_list = malloc(3 * sizeof(u_int));
+       if (handle->tstamp_type_list == NULL) {
+               free(handle);
+               return NULL;
+       }
+       handle->tstamp_type_list[0] = PCAP_TSTAMP_HOST;
+       handle->tstamp_type_list[1] = PCAP_TSTAMP_ADAPTER;
+       handle->tstamp_type_list[2] = PCAP_TSTAMP_ADAPTER_UNSYNCED;
+#endif
+
        return handle;
 }
 
 #ifdef HAVE_LIBNL
 /*
-        *
-        * If interface {if} is a mac80211 driver, the file
-        * /sys/class/net/{if}/phy80211 is a symlink to
-        * /sys/class/ieee80211/{phydev}, for some {phydev}.
-        *
-        * On Fedora 9, with a 2.6.26.3-29 kernel, my Zydas stick, at
-        * least, has a "wmaster0" device and a "wlan0" device; the
-        * latter is the one with the IP address.  Both show up in
-        * "tcpdump -D" output.  Capturing on the wmaster0 device
-        * captures with 802.11 headers.
-        *
-        * airmon-ng searches through /sys/class/net for devices named
-        * monN, starting with mon0; as soon as one *doesn't* exist,
-        * it chooses that as the monitor device name.  If the "iw"
-        * command exists, it does "iw dev {if} interface add {monif}
-        * type monitor", where {monif} is the monitor device.  It
-        * then (sigh) sleeps .1 second, and then configures the
-        * device up.  Otherwise, if /sys/class/ieee80211/{phydev}/add_iface
-        * is a file, it writes {mondev}, without a newline, to that file,
-        * and again (sigh) sleeps .1 second, and then iwconfig's that
-        * device into monitor mode and configures it up.  Otherwise,
-        * you can't do monitor mode.
-        *
-        * All these devices are "glued" together by having the
-        * /sys/class/net/{device}/phy80211 links pointing to the same
-        * place, so, given a wmaster, wlan, or mon device, you can
-        * find the other devices by looking for devices with
-        * the same phy80211 link.
-        *
-        * To turn monitor mode off, delete the monitor interface,
-        * either with "iw dev {monif} interface del" or by sending
-        * {monif}, with no NL, down /sys/class/ieee80211/{phydev}/remove_iface
-        *
-        * Note: if you try to create a monitor device named "monN", and
-        * there's already a "monN" device, it fails, as least with
-        * the netlink interface (which is what iw uses), with a return
-        * value of -ENFILE.  (Return values are negative errnos.)  We
-        * could probably use that to find an unused device.
-        *
-        * Yes, you can have multiple monitor devices for a given
-        * physical device.
+ * If interface {if} is a mac80211 driver, the file
+ * /sys/class/net/{if}/phy80211 is a symlink to
+ * /sys/class/ieee80211/{phydev}, for some {phydev}.
+ *
+ * On Fedora 9, with a 2.6.26.3-29 kernel, my Zydas stick, at
+ * least, has a "wmaster0" device and a "wlan0" device; the
+ * latter is the one with the IP address.  Both show up in
+ * "tcpdump -D" output.  Capturing on the wmaster0 device
+ * captures with 802.11 headers.
+ *
+ * airmon-ng searches through /sys/class/net for devices named
+ * monN, starting with mon0; as soon as one *doesn't* exist,
+ * it chooses that as the monitor device name.  If the "iw"
+ * command exists, it does "iw dev {if} interface add {monif}
+ * type monitor", where {monif} is the monitor device.  It
+ * then (sigh) sleeps .1 second, and then configures the
+ * device up.  Otherwise, if /sys/class/ieee80211/{phydev}/add_iface
+ * is a file, it writes {mondev}, without a newline, to that file,
+ * and again (sigh) sleeps .1 second, and then iwconfig's that
+ * device into monitor mode and configures it up.  Otherwise,
+ * you can't do monitor mode.
+ *
+ * All these devices are "glued" together by having the
+ * /sys/class/net/{device}/phy80211 links pointing to the same
+ * place, so, given a wmaster, wlan, or mon device, you can
+ * find the other devices by looking for devices with
+ * the same phy80211 link.
+ *
+ * To turn monitor mode off, delete the monitor interface,
+ * either with "iw dev {monif} interface del" or by sending
+ * {monif}, with no NL, down /sys/class/ieee80211/{phydev}/remove_iface
+ *
+ * Note: if you try to create a monitor device named "monN", and
+ * there's already a "monN" device, it fails, as least with
+ * the netlink interface (which is what iw uses), with a return
+ * value of -ENFILE.  (Return values are negative errnos.)  We
+ * could probably use that to find an unused device.
+ *
+ * Yes, you can have multiple monitor devices for a given
+ * physical device.
 */
 
 /*
@@ -475,8 +527,37 @@ get_mac80211_phydev(pcap_t *handle, const char *device, char *phydev_path,
        return 1;
 }
 
+#ifndef HAVE_LIBNL_2_x
+/* libnl 2.x compatibility code */
+
+#define nl_sock nl_handle
+
+static inline struct nl_handle *
+nl_socket_alloc(void)
+{
+       return nl_handle_alloc();
+}
+
+static inline void
+nl_socket_free(struct nl_handle *h)
+{
+       nl_handle_destroy(h);
+}
+
+static inline int
+__genl_ctrl_alloc_cache(struct nl_handle *h, struct nl_cache **cache)
+{
+       struct nl_cache *tmp = genl_ctrl_alloc_cache(h);
+       if (!tmp)
+               return -ENOMEM;
+       *cache = tmp;
+       return 0;
+}
+#define genl_ctrl_alloc_cache __genl_ctrl_alloc_cache
+#endif /* !HAVE_LIBNL_2_x */
+
 struct nl80211_state {
-       struct nl_handle *nl_handle;
+       struct nl_sock *nl_sock;
        struct nl_cache *nl_cache;
        struct genl_family *nl80211;
 };
@@ -484,23 +565,26 @@ struct nl80211_state {
 static int
 nl80211_init(pcap_t *handle, struct nl80211_state *state, const char *device)
 {
-       state->nl_handle = nl_handle_alloc();
-       if (!state->nl_handle) {
+       int err;
+
+       state->nl_sock = nl_socket_alloc();
+       if (!state->nl_sock) {
                snprintf(handle->errbuf, PCAP_ERRBUF_SIZE,
                    "%s: failed to allocate netlink handle", device);
                return PCAP_ERROR;
        }
 
-       if (genl_connect(state->nl_handle)) {
+       if (genl_connect(state->nl_sock)) {
                snprintf(handle->errbuf, PCAP_ERRBUF_SIZE,
                    "%s: failed to connect to generic netlink", device);
                goto out_handle_destroy;
        }
 
-       state->nl_cache = genl_ctrl_alloc_cache(state->nl_handle);
-       if (!state->nl_cache) {
+       err = genl_ctrl_alloc_cache(state->nl_sock, &state->nl_cache);
+       if (err < 0) {
                snprintf(handle->errbuf, PCAP_ERRBUF_SIZE,
-                   "%s: failed to allocate generic netlink cache", device);
+                   "%s: failed to allocate generic netlink cache: %s",
+                   device, strerror(-err));
                goto out_handle_destroy;
        }
 
@@ -516,7 +600,7 @@ nl80211_init(pcap_t *handle, struct nl80211_state *state, const char *device)
 out_cache_free:
        nl_cache_free(state->nl_cache);
 out_handle_destroy:
-       nl_handle_destroy(state->nl_handle);
+       nl_socket_free(state->nl_sock);
        return PCAP_ERROR;
 }
 
@@ -525,7 +609,7 @@ nl80211_cleanup(struct nl80211_state *state)
 {
        genl_family_put(state->nl80211);
        nl_cache_free(state->nl_cache);
-       nl_handle_destroy(state->nl_handle);
+       nl_socket_free(state->nl_sock);
 }
 
 static int
@@ -553,7 +637,7 @@ add_mon_if(pcap_t *handle, int sock_fd, struct nl80211_state *state,
        NLA_PUT_STRING(msg, NL80211_ATTR_IFNAME, mondevice);
        NLA_PUT_U32(msg, NL80211_ATTR_IFTYPE, NL80211_IFTYPE_MONITOR);
 
-       err = nl_send_auto_complete(state->nl_handle, msg);
+       err = nl_send_auto_complete(state->nl_sock, msg);
        if (err < 0) {
                if (err == -ENFILE) {
                        /*
@@ -574,7 +658,7 @@ add_mon_if(pcap_t *handle, int sock_fd, struct nl80211_state *state,
                        return PCAP_ERROR;
                }
        }
-       err = nl_wait_for_ack(state->nl_handle);
+       err = nl_wait_for_ack(state->nl_sock);
        if (err < 0) {
                if (err == -ENFILE) {
                        /*
@@ -633,7 +717,7 @@ del_mon_if(pcap_t *handle, int sock_fd, struct nl80211_state *state,
                    0, NL80211_CMD_DEL_INTERFACE, 0);
        NLA_PUT_U32(msg, NL80211_ATTR_IFINDEX, ifindex);
 
-       err = nl_send_auto_complete(state->nl_handle, msg);
+       err = nl_send_auto_complete(state->nl_sock, msg);
        if (err < 0) {
                if (err == -ENFILE) {
                        /*
@@ -654,7 +738,7 @@ del_mon_if(pcap_t *handle, int sock_fd, struct nl80211_state *state,
                        return PCAP_ERROR;
                }
        }
-       err = nl_wait_for_ack(state->nl_handle);
+       err = nl_wait_for_ack(state->nl_sock);
        if (err < 0) {
                if (err == -ENFILE) {
                        /*
@@ -877,6 +961,8 @@ pcap_can_set_rfmon_linux(pcap_t *handle)
        }
        if (errno == ENODEV) {
                /* The device doesn't even exist. */
+               (void)snprintf(handle->errbuf, PCAP_ERRBUF_SIZE,
+                   "SIOCGIWMODE failed: %s", pcap_strerror(errno));
                close(sock_fd);
                return PCAP_ERROR_NO_SUCH_DEVICE;
        }
@@ -885,6 +971,67 @@ pcap_can_set_rfmon_linux(pcap_t *handle)
        return 0;
 }
 
+/*
+ * Grabs the number of dropped packets by the interface from /proc/net/dev.
+ *
+ * XXX - what about /sys/class/net/{interface name}/rx_*?  There are
+ * individual devices giving, in ASCII, various rx_ and tx_ statistics.
+ *
+ * Or can we get them in binary form from netlink?
+ */
+static long int
+linux_if_drops(const char * if_name)
+{
+       char buffer[512];
+       char * bufptr;
+       FILE * file;
+       int field_to_convert = 3, if_name_sz = strlen(if_name);
+       long int dropped_pkts = 0;
+       
+       file = fopen("/proc/net/dev", "r");
+       if (!file)
+               return 0;
+
+       while (!dropped_pkts && fgets( buffer, sizeof(buffer), file ))
+       {
+               /*      search for 'bytes' -- if its in there, then
+                       that means we need to grab the fourth field. otherwise
+                       grab the third field. */
+               if (field_to_convert != 4 && strstr(buffer, "bytes"))
+               {
+                       field_to_convert = 4;
+                       continue;
+               }
+       
+               /* find iface and make sure it actually matches -- space before the name and : after it */
+               if ((bufptr = strstr(buffer, if_name)) &&
+                       (bufptr == buffer || *(bufptr-1) == ' ') &&
+                       *(bufptr + if_name_sz) == ':')
+               {
+                       bufptr = bufptr + if_name_sz + 1;
+
+                       /* grab the nth field from it */
+                       while( --field_to_convert && *bufptr != '\0')
+                       {
+                               while (*bufptr != '\0' && *(bufptr++) == ' ');
+                               while (*bufptr != '\0' && *(bufptr++) != ' ');
+                       }
+                       
+                       /* get rid of any final spaces */
+                       while (*bufptr != '\0' && *bufptr == ' ') bufptr++;
+                       
+                       if (*bufptr != '\0')
+                               dropped_pkts = strtol(bufptr, NULL, 10);
+
+                       break;
+               }
+       }
+       
+       fclose(file);
+       return dropped_pkts;
+} 
+
+
 /*
  * With older kernels promiscuous mode is kind of interesting because we
  * have to reset the interface before exiting. The problem can't really
@@ -1060,6 +1207,14 @@ pcap_activate_linux(pcap_t *handle)
                         pcap_strerror(errno) );
                return PCAP_ERROR;
        }
+       
+       /*
+        * If we're in promiscuous mode, then we probably want 
+        * to see when the interface drops packets too, so get an
+        * initial count from /proc/net/dev
+        */
+       if (handle->opt.promisc)
+               handle->md.proc_dropped = linux_if_drops(handle->md.device);
 
        /*
         * Current Linux kernels use the protocol family PF_PACKET to
@@ -1070,34 +1225,46 @@ pcap_activate_linux(pcap_t *handle)
         * to be compatible with older kernels for a while so we are
         * trying both methods with the newer method preferred.
         */
-
-       if ((status = activate_new(handle)) == 1) {
+       status = activate_new(handle);
+       if (status < 0) {
+               /*
+                * Fatal error with the new way; just fail.
+                * status has the error return; if it's PCAP_ERROR,
+                * handle->errbuf has been set appropriately.
+                */
+               goto fail;
+       }
+       if (status == 1) {
                /*
                 * Success.
                 * Try to use memory-mapped access.
                 */
-               switch (activate_mmap(handle)) {
+               switch (activate_mmap(handle, &status)) {
 
                case 1:
-                       /* we succeeded; nothing more to do */
-                       return 0;
+                       /*
+                        * We succeeded.  status has been
+                        * set to the status to return,
+                        * which might be 0, or might be
+                        * a PCAP_WARNING_ value.
+                        */
+                       return status;
 
                case 0:
                        /*
                         * Kernel doesn't support it - just continue
                         * with non-memory-mapped access.
                         */
-                       status = 0;
                        break;
 
                case -1:
                        /*
-                        * We failed to set up to use it, or kernel
-                        * supports it, but we failed to enable it;
-                        * return an error.  handle->errbuf contains
-                        * an error message.
+                        * We failed to set up to use it, or the kernel
+                        * supports it, but we failed to enable it.
+                        * status has been set to the error status to
+                        * return and, if it's PCAP_ERROR, handle->errbuf
+                        * contains the error message.
                         */
-                       status = PCAP_ERROR;
                        goto fail;
                }
        }
@@ -1111,18 +1278,12 @@ pcap_activate_linux(pcap_t *handle)
                         */
                        goto fail;
                }
-       } else {
-               /*
-                * Fatal error with the new way; just fail.
-                * status has the error return; if it's PCAP_ERROR,
-                * handle->errbuf has been set appropriately.
-                */
-               goto fail;
        }
 
        /*
         * We set up the socket, but not with memory-mapped access.
         */
+       status = 0;
        if (handle->opt.buffer_size != 0) {
                /*
                 * Set the socket buffer size to the specified value.
@@ -1258,12 +1419,12 @@ pcap_read_packet(pcap_t *handle, pcap_handler callback, u_char *userdata)
                 */
                if (handle->break_loop) {
                        /*
-                        * Yes - clear the flag that indicates that it
-                        * has, and return -2 as an indication that we
-                        * were told to break out of the loop.
+                        * Yes - clear the flag that indicates that it has,
+                        * and return PCAP_ERROR_BREAK as an indication that
+                        * we were told to break out of the loop.
                         */
                        handle->break_loop = 0;
-                       return -2;
+                       return PCAP_ERROR_BREAK;
                }
 
 #if defined(HAVE_PACKET_AUXDATA) && defined(HAVE_LINUX_TPACKET_AUXDATA_TP_VLAN_TCI)
@@ -1275,17 +1436,32 @@ pcap_read_packet(pcap_t *handle, pcap_handler callback, u_char *userdata)
                        handle->bufsize - offset, MSG_TRUNC,
                        (struct sockaddr *) &from, &fromlen);
 #endif /* defined(HAVE_PACKET_AUXDATA) && defined(HAVE_LINUX_TPACKET_AUXDATA_TP_VLAN_TCI) */
-       } while (packet_len == -1 && (errno == EINTR || errno == ENETDOWN));
+       } while (packet_len == -1 && errno == EINTR);
 
        /* Check if an error occured */
 
        if (packet_len == -1) {
-               if (errno == EAGAIN)
+               switch (errno) {
+
+               case EAGAIN:
                        return 0;       /* no packet there */
-               else {
+
+               case ENETDOWN:
+                       /*
+                        * The device on which we're capturing went away.
+                        *
+                        * XXX - we should really return
+                        * PCAP_ERROR_IFACE_NOT_UP, but pcap_dispatch()
+                        * etc. aren't defined to return that.
+                        */
+                       snprintf(handle->errbuf, PCAP_ERRBUF_SIZE,
+                               "The interface went down");
+                       return PCAP_ERROR;
+
+               default:
                        snprintf(handle->errbuf, PCAP_ERRBUF_SIZE,
                                 "recvfrom: %s", pcap_strerror(errno));
-                       return -1;
+                       return PCAP_ERROR;
                }
        }
 
@@ -1443,7 +1619,7 @@ pcap_read_packet(pcap_t *handle, pcap_handler callback, u_char *userdata)
        if (ioctl(handle->fd, SIOCGSTAMP, &pcap_header.ts) == -1) {
                snprintf(handle->errbuf, PCAP_ERRBUF_SIZE,
                         "SIOCGSTAMP: %s", pcap_strerror(errno));
-               return -1;
+               return PCAP_ERROR;
        }
        pcap_header.caplen      = caplen;
        pcap_header.len         = packet_len;
@@ -1558,6 +1734,18 @@ pcap_stats_linux(pcap_t *handle, struct pcap_stat *stats)
        socklen_t len = sizeof (struct tpacket_stats);
 #endif
 
+       long if_dropped = 0;
+       
+       /* 
+        *      To fill in ps_ifdrop, we parse /proc/net/dev for the number
+        */
+       if (handle->opt.promisc)
+       {
+               if_dropped = handle->md.proc_dropped;
+               handle->md.proc_dropped = linux_if_drops(handle->md.device);
+               handle->md.stat.ps_ifdrop += (handle->md.proc_dropped - if_dropped);
+       }
+
 #ifdef HAVE_TPACKET_STATS
        /*
         * Try to get the packet counts from the kernel.
@@ -1578,6 +1766,8 @@ pcap_stats_linux(pcap_t *handle, struct pcap_stat *stats)
                 *      dropped by the interface driver.  It counts only
                 *      packets that passed the filter.
                 *
+                *      See above for ps_ifdrop. 
+                *
                 *      Both statistics include packets not yet read from
                 *      the kernel by libpcap, and thus not yet seen by
                 *      the application.
@@ -1641,19 +1831,299 @@ pcap_stats_linux(pcap_t *handle, struct pcap_stat *stats)
         *
         *      "ps_drop" is not supported.
         *
+        *      "ps_ifdrop" is supported. It will return the number
+        *      of drops the interface reports in /proc/net/dev,
+        *      if that is available.
+        *
         *      "ps_recv" doesn't include packets not yet read from
         *      the kernel by libpcap.
         *
         * We maintain the count of packets processed by libpcap in
         * "md.packets_read", for reasons described in the comment
         * at the end of pcap_read_packet().  We have no idea how many
-        * packets were dropped.
+        * packets were dropped by the kernel buffers -- but we know 
+        * how many the interface dropped, so we can return that.
         */
+        
        stats->ps_recv = handle->md.packets_read;
        stats->ps_drop = 0;
+       stats->ps_ifdrop = handle->md.stat.ps_ifdrop;
        return 0;
 }
 
+/*
+ * Get from "/sys/class/net" all interfaces listed there; if they're
+ * already in the list of interfaces we have, that won't add another
+ * instance, but if they're not, that'll add them.
+ *
+ * We don't bother getting any addresses for them; it appears you can't
+ * use SIOCGIFADDR on Linux to get IPv6 addresses for interfaces, and,
+ * although some other types of addresses can be fetched with SIOCGIFADDR,
+ * we don't bother with them for now.
+ *
+ * We also don't fail if we couldn't open "/sys/class/net"; we just leave
+ * the list of interfaces as is, and return 0, so that we can try
+ * scanning /proc/net/dev.
+ */
+static int
+scan_sys_class_net(pcap_if_t **devlistp, char *errbuf)
+{
+       DIR *sys_class_net_d;
+       int fd;
+       struct dirent *ent;
+       char *p;
+       char name[512]; /* XXX - pick a size */
+       char *q, *saveq;
+       struct ifreq ifrflags;
+       int ret = 1;
+
+       sys_class_net_d = opendir("/sys/class/net");
+       if (sys_class_net_d == NULL && errno == ENOENT)
+               return (0);
+
+       /*
+        * Create a socket from which to fetch interface information.
+        */
+       fd = socket(AF_INET, SOCK_DGRAM, 0);
+       if (fd < 0) {
+               (void)snprintf(errbuf, PCAP_ERRBUF_SIZE,
+                   "socket: %s", pcap_strerror(errno));
+               return (-1);
+       }
+
+       for (;;) {
+               errno = 0;
+               ent = readdir(sys_class_net_d);
+               if (ent == NULL) {
+                       /*
+                        * Error or EOF; if errno != 0, it's an error.
+                        */
+                       break;
+               }
+
+               /*
+                * Ignore directories (".", "..", and any subdirectories).
+                */
+               if (ent->d_type == DT_DIR)
+                       continue;
+
+               /*
+                * Get the interface name.
+                */
+               p = &ent->d_name[0];
+               q = &name[0];
+               while (*p != '\0' && isascii(*p) && !isspace(*p)) {
+                       if (*p == ':') {
+                               /*
+                                * This could be the separator between a
+                                * name and an alias number, or it could be
+                                * the separator between a name with no
+                                * alias number and the next field.
+                                *
+                                * If there's a colon after digits, it
+                                * separates the name and the alias number,
+                                * otherwise it separates the name and the
+                                * next field.
+                                */
+                               saveq = q;
+                               while (isascii(*p) && isdigit(*p))
+                                       *q++ = *p++;
+                               if (*p != ':') {
+                                       /*
+                                        * That was the next field,
+                                        * not the alias number.
+                                        */
+                                       q = saveq;
+                               }
+                               break;
+                       } else
+                               *q++ = *p++;
+               }
+               *q = '\0';
+
+               /*
+                * Get the flags for this interface, and skip it if
+                * it's not up.
+                */
+               strncpy(ifrflags.ifr_name, name, sizeof(ifrflags.ifr_name));
+               if (ioctl(fd, SIOCGIFFLAGS, (char *)&ifrflags) < 0) {
+                       if (errno == ENXIO || errno == ENODEV)
+                               continue;
+                       (void)snprintf(errbuf, PCAP_ERRBUF_SIZE,
+                           "SIOCGIFFLAGS: %.*s: %s",
+                           (int)sizeof(ifrflags.ifr_name),
+                           ifrflags.ifr_name,
+                           pcap_strerror(errno));
+                       ret = -1;
+                       break;
+               }
+               if (!(ifrflags.ifr_flags & IFF_UP))
+                       continue;
+
+               /*
+                * Add an entry for this interface, with no addresses.
+                */
+               if (pcap_add_if(devlistp, name, ifrflags.ifr_flags, NULL,
+                   errbuf) == -1) {
+                       /*
+                        * Failure.
+                        */
+                       ret = -1;
+                       break;
+               }
+       }
+       if (ret != -1) {
+               /*
+                * Well, we didn't fail for any other reason; did we
+                * fail due to an error reading the directory?
+                */
+               if (errno != 0) {
+                       (void)snprintf(errbuf, PCAP_ERRBUF_SIZE,
+                           "Error reading /sys/class/net: %s",
+                           pcap_strerror(errno));
+                       ret = -1;
+               }
+       }
+
+       (void)close(fd);
+       (void)closedir(sys_class_net_d);
+       return (ret);
+}
+
+/*
+ * Get from "/proc/net/dev" all interfaces listed there; if they're
+ * already in the list of interfaces we have, that won't add another
+ * instance, but if they're not, that'll add them.
+ *
+ * See comments from scan_sys_class_net().
+ */
+static int
+scan_proc_net_dev(pcap_if_t **devlistp, char *errbuf)
+{
+       FILE *proc_net_f;
+       int fd;
+       char linebuf[512];
+       int linenum;
+       char *p;
+       char name[512]; /* XXX - pick a size */
+       char *q, *saveq;
+       struct ifreq ifrflags;
+       int ret = 0;
+
+       proc_net_f = fopen("/proc/net/dev", "r");
+       if (proc_net_f == NULL && errno == ENOENT)
+               return (0);
+
+       /*
+        * Create a socket from which to fetch interface information.
+        */
+       fd = socket(AF_INET, SOCK_DGRAM, 0);
+       if (fd < 0) {
+               (void)snprintf(errbuf, PCAP_ERRBUF_SIZE,
+                   "socket: %s", pcap_strerror(errno));
+               return (-1);
+       }
+
+       for (linenum = 1;
+           fgets(linebuf, sizeof linebuf, proc_net_f) != NULL; linenum++) {
+               /*
+                * Skip the first two lines - they're headers.
+                */
+               if (linenum <= 2)
+                       continue;
+
+               p = &linebuf[0];
+
+               /*
+                * Skip leading white space.
+                */
+               while (*p != '\0' && isascii(*p) && isspace(*p))
+                       p++;
+               if (*p == '\0' || *p == '\n')
+                       continue;       /* blank line */
+
+               /*
+                * Get the interface name.
+                */
+               q = &name[0];
+               while (*p != '\0' && isascii(*p) && !isspace(*p)) {
+                       if (*p == ':') {
+                               /*
+                                * This could be the separator between a
+                                * name and an alias number, or it could be
+                                * the separator between a name with no
+                                * alias number and the next field.
+                                *
+                                * If there's a colon after digits, it
+                                * separates the name and the alias number,
+                                * otherwise it separates the name and the
+                                * next field.
+                                */
+                               saveq = q;
+                               while (isascii(*p) && isdigit(*p))
+                                       *q++ = *p++;
+                               if (*p != ':') {
+                                       /*
+                                        * That was the next field,
+                                        * not the alias number.
+                                        */
+                                       q = saveq;
+                               }
+                               break;
+                       } else
+                               *q++ = *p++;
+               }
+               *q = '\0';
+
+               /*
+                * Get the flags for this interface, and skip it if
+                * it's not up.
+                */
+               strncpy(ifrflags.ifr_name, name, sizeof(ifrflags.ifr_name));
+               if (ioctl(fd, SIOCGIFFLAGS, (char *)&ifrflags) < 0) {
+                       if (errno == ENXIO)
+                               continue;
+                       (void)snprintf(errbuf, PCAP_ERRBUF_SIZE,
+                           "SIOCGIFFLAGS: %.*s: %s",
+                           (int)sizeof(ifrflags.ifr_name),
+                           ifrflags.ifr_name,
+                           pcap_strerror(errno));
+                       ret = -1;
+                       break;
+               }
+               if (!(ifrflags.ifr_flags & IFF_UP))
+                       continue;
+
+               /*
+                * Add an entry for this interface, with no addresses.
+                */
+               if (pcap_add_if(devlistp, name, ifrflags.ifr_flags, NULL,
+                   errbuf) == -1) {
+                       /*
+                        * Failure.
+                        */
+                       ret = -1;
+                       break;
+               }
+       }
+       if (ret != -1) {
+               /*
+                * Well, we didn't fail for any other reason; did we
+                * fail due to an error reading the file?
+                */
+               if (ferror(proc_net_f)) {
+                       (void)snprintf(errbuf, PCAP_ERRBUF_SIZE,
+                           "Error reading /proc/net/dev: %s",
+                           pcap_strerror(errno));
+                       ret = -1;
+               }
+       }
+
+       (void)close(fd);
+       (void)fclose(proc_net_f);
+       return (ret);
+}
+
 /*
  * Description string for the "any" device.
  */
@@ -1662,25 +2132,66 @@ static const char any_descr[] = "Pseudo-device that captures on all interfaces";
 int
 pcap_platform_finddevs(pcap_if_t **alldevsp, char *errbuf)
 {
+       int ret;
+
+       /*
+        * Read "/sys/class/net", and add to the list of interfaces all
+        * interfaces listed there that we don't already have, because,
+        * on Linux, SIOCGIFCONF reports only interfaces with IPv4 addresses,
+        * and even getifaddrs() won't return information about
+        * interfaces with no addresses, so you need to read "/sys/class/net"
+        * to get the names of the rest of the interfaces.
+        */
+       ret = scan_sys_class_net(alldevsp, errbuf);
+       if (ret == -1)
+               return (-1);    /* failed */
+       if (ret == 0) {
+               /*
+                * No /sys/class/net; try reading /proc/net/dev instead.
+                */
+               if (scan_proc_net_dev(alldevsp, errbuf) == -1)
+                       return (-1);
+       }
+
+       /*
+        * Add the "any" device.
+        */
        if (pcap_add_if(alldevsp, "any", 0, any_descr, errbuf) < 0)
                return (-1);
 
 #ifdef HAVE_DAG_API
+       /*
+        * Add DAG devices.
+        */
        if (dag_platform_finddevs(alldevsp, errbuf) < 0)
                return (-1);
 #endif /* HAVE_DAG_API */
 
 #ifdef HAVE_SEPTEL_API
+       /*
+        * Add Septel devices.
+        */
        if (septel_platform_finddevs(alldevsp, errbuf) < 0)
                return (-1);
 #endif /* HAVE_SEPTEL_API */
 
+#ifdef HAVE_SNF_API
+       if (snf_platform_finddevs(alldevsp, errbuf) < 0)
+               return (-1);
+#endif /* HAVE_SNF_API */
+
 #ifdef PCAP_SUPPORT_BT
+       /*
+        * Add Bluetooth devices.
+        */
        if (bt_platform_finddevs(alldevsp, errbuf) < 0)
                return (-1);
 #endif
 
 #ifdef PCAP_SUPPORT_USB
+       /*
+        * Add USB devices.
+        */
        if (usb_platform_finddevs(alldevsp, errbuf) < 0)
                return (-1);
 #endif
@@ -1692,7 +2203,8 @@ pcap_platform_finddevs(pcap_if_t **alldevsp, char *errbuf)
  *  Attach the given BPF code to the packet capture device.
  */
 static int
-pcap_setfilter_linux(pcap_t *handle, struct bpf_program *filter)
+pcap_setfilter_linux_common(pcap_t *handle, struct bpf_program *filter,
+    int is_mmapped)
 {
 #ifdef SO_ATTACH_FILTER
        struct sock_fprog       fcode;
@@ -1745,13 +2257,13 @@ pcap_setfilter_linux(pcap_t *handle, struct bpf_program *filter)
                 *
                 * Oh, and we also need to fix it up so that all "ret"
                 * instructions with non-zero operands have 65535 as the
-                * operand, and so that, if we're in cooked mode, all
-                * memory-reference instructions use special magic offsets
-                * in references to the link-layer header and assume that
-                * the link-layer payload begins at 0; "fix_program()"
-                * will do that.
+                * operand if we're not capturing in memory-mapped modee,
+                * and so that, if we're in cooked mode, all memory-reference
+                * instructions use special magic offsets in references to
+                * the link-layer header and assume that the link-layer
+                * payload begins at 0; "fix_program()" will do that.
                 */
-               switch (fix_program(handle, &fcode)) {
+               switch (fix_program(handle, &fcode, is_mmapped)) {
 
                case -1:
                default:
@@ -1825,6 +2337,13 @@ pcap_setfilter_linux(pcap_t *handle, struct bpf_program *filter)
        return 0;
 }
 
+static int
+pcap_setfilter_linux(pcap_t *handle, struct bpf_program *filter)
+{
+       return pcap_setfilter_linux_common(handle, filter, 0);
+}
+
+
 /*
  * Set direction flag: Which packets do we accept on a forwarding
  * single device? IN, OUT or both?
@@ -1847,7 +2366,6 @@ pcap_setdirection_linux(pcap_t *handle, pcap_direction_t d)
        return -1;
 }
 
-
 #ifdef HAVE_PF_PACKET_SOCKETS
 /*
  * Map the PACKET_ value to a LINUX_SLL_ value; we
@@ -1955,6 +2473,12 @@ static void map_arphrd_to_dlt(pcap_t *handle, int arptype, int cooked_ok)
        case ARPHRD_CHAOS:
                handle->linktype = DLT_CHAOS;
                break;
+#ifndef ARPHRD_CAN
+#define ARPHRD_CAN 280
+#endif
+       case ARPHRD_CAN:
+               handle->linktype = DLT_CAN_SOCKETCAN;
+               break;
 
 #ifndef ARPHRD_IEEE802_TR
 #define ARPHRD_IEEE802_TR 800  /* From Linux 2.4 */
@@ -2188,6 +2712,13 @@ static void map_arphrd_to_dlt(pcap_t *handle, int arptype, int cooked_ok)
                handle->linktype = DLT_RAW;
                break;
 
+#ifndef ARPHRD_IEEE802154
+#define ARPHRD_IEEE802154      804
+#endif
+       case ARPHRD_IEEE802154:
+               handle->linktype =  DLT_IEEE802_15_4_NOFCS;
+               break;
+
        default:
                handle->linktype = -1;
                break;
@@ -2489,40 +3020,97 @@ activate_new(pcap_t *handle)
 #endif
 }
 
+#ifdef HAVE_PACKET_RING
+/*
+ * Attempt to activate with memory-mapped access.
+ *
+ * On success, returns 1, and sets *status to 0 if there are no warnings
+ * or to a PCAP_WARNING_ code if there is a warning.
+ *
+ * On failure due to lack of support for memory-mapped capture, returns
+ * 0.
+ *
+ * On error, returns -1, and sets *status to the appropriate error code;
+ * if that is PCAP_ERROR, sets handle->errbuf to the appropriate message.
+ */
 static int 
-activate_mmap(pcap_t *handle)
+activate_mmap(pcap_t *handle, int *status)
 {
-#ifdef HAVE_PACKET_RING
        int ret;
 
+       /*
+        * Attempt to allocate a buffer to hold the contents of one
+        * packet, for use by the oneshot callback.
+        */
+       handle->md.oneshot_buffer = malloc(handle->snapshot);
+       if (handle->md.oneshot_buffer == NULL) {
+               snprintf(handle->errbuf, PCAP_ERRBUF_SIZE,
+                        "can't allocate oneshot buffer: %s",
+                        pcap_strerror(errno));
+               *status = PCAP_ERROR;
+               return -1;
+       }
+
        if (handle->opt.buffer_size == 0) {
                /* by default request 2M for the ring buffer */
                handle->opt.buffer_size = 2*1024*1024;
        }
        ret = prepare_tpacket_socket(handle);
-       if (ret != 1)
-               return ret;
-       ret = create_ring(handle);
-       if (ret != 1)
+       if (ret == -1) {
+               free(handle->md.oneshot_buffer);
+               *status = PCAP_ERROR;
                return ret;
+       }
+       ret = create_ring(handle, status);
+       if (ret == 0) {
+               /*
+                * We don't support memory-mapped capture; our caller
+                * will fall back on reading from the socket.
+                */
+               free(handle->md.oneshot_buffer);
+               return 0;
+       }
+       if (ret == -1) {
+               /*
+                * Error attempting to enable memory-mapped capture;
+                * fail.  create_ring() has set *status.
+                */
+               free(handle->md.oneshot_buffer);
+               return -1;
+       }
 
-       /* override some defaults and inherit the other fields from
-        * activate_new
-        * handle->offset is used to get the current position into the rx ring 
-        * handle->cc is used to store the ring size */
+       /*
+        * Success.  *status has been set either to 0 if there are no
+        * warnings or to a PCAP_WARNING_ value if there is a warning.
+        *
+        * Override some defaults and inherit the other fields from
+        * activate_new.
+        * handle->offset is used to get the current position into the rx ring.
+        * handle->cc is used to store the ring size.
+        */
        handle->read_op = pcap_read_linux_mmap;
        handle->cleanup_op = pcap_cleanup_linux_mmap;
        handle->setfilter_op = pcap_setfilter_linux_mmap;
        handle->setnonblock_op = pcap_setnonblock_mmap;
        handle->getnonblock_op = pcap_getnonblock_mmap;
+       handle->oneshot_callback = pcap_oneshot_mmap;
        handle->selectable_fd = handle->fd;
        return 1;
+}
 #else /* HAVE_PACKET_RING */
+static int 
+activate_mmap(pcap_t *handle _U_, int *status _U_)
+{
        return 0;
-#endif /* HAVE_PACKET_RING */
 }
+#endif /* HAVE_PACKET_RING */
 
 #ifdef HAVE_PACKET_RING
+/*
+ * Attempt to set the socket to version 2 of the memory-mapped header.
+ * Return 1 if we succeed or if we fail because version 2 isn't
+ * supported; return -1 on any other error, and set handle->errbuf.
+ */
 static int
 prepare_tpacket_socket(pcap_t *handle)
 {
@@ -2574,26 +3162,29 @@ prepare_tpacket_socket(pcap_t *handle)
        return 1;
 }
 
-static void
-compute_ring_block(int frame_size, unsigned *block_size, unsigned *frames_per_block)
-{
-       /* compute the minumum block size that will handle this frame. 
-        * The block has to be page size aligned. 
-        * The max block size allowed by the kernel is arch-dependent and 
-        * it's not explicitly checked here. */
-       *block_size = getpagesize();
-       while (*block_size < frame_size) 
-               *block_size <<= 1;
-
-       *frames_per_block = *block_size/frame_size;
-}
-
+/*
+ * Attempt to set up memory-mapped access.
+ *
+ * On success, returns 1, and sets *status to 0 if there are no warnings
+ * or to a PCAP_WARNING_ code if there is a warning.
+ *
+ * On failure due to lack of support for memory-mapped capture, returns
+ * 0.
+ *
+ * On error, returns -1, and sets *status to the appropriate error code;
+ * if that is PCAP_ERROR, sets handle->errbuf to the appropriate message.
+ */
 static int
-create_ring(pcap_t *handle)
+create_ring(pcap_t *handle, int *status)
 {
-       unsigned i, j, ringsize, frames_per_block;
+       unsigned i, j, frames_per_block;
        struct tpacket_req req;
 
+       /*
+        * Start out assuming no warnings or errors.
+        */
+       *status = 0;
+
        /* Note that with large snapshot (say 64K) only a few frames 
         * will be available in the ring even with pretty large ring size
         * (and a lot of memory will be unused). 
@@ -2603,20 +3194,143 @@ create_ring(pcap_t *handle)
                                          TPACKET_ALIGN(handle->md.tp_hdrlen) +
                                          sizeof(struct sockaddr_ll));
        req.tp_frame_nr = handle->opt.buffer_size/req.tp_frame_size;
-       compute_ring_block(req.tp_frame_size, &req.tp_block_size, &frames_per_block);
-       req.tp_block_nr = req.tp_frame_nr / frames_per_block;
 
-       /* req.tp_frame_nr is requested to match frames_per_block*req.tp_block_nr */
-       req.tp_frame_nr = req.tp_block_nr * frames_per_block;
+       /* compute the minumum block size that will handle this frame. 
+        * The block has to be page size aligned. 
+        * The max block size allowed by the kernel is arch-dependent and 
+        * it's not explicitly checked here. */
+       req.tp_block_size = getpagesize();
+       while (req.tp_block_size < req.tp_frame_size) 
+               req.tp_block_size <<= 1;
+
+       frames_per_block = req.tp_block_size/req.tp_frame_size;
+
+       /*
+        * PACKET_TIMESTAMP was added after linux/net_tstamp.h was,
+        * so we check for PACKET_TIMESTAMP.  We check for
+        * linux/net_tstamp.h just in case a system somehow has
+        * PACKET_TIMESTAMP but not linux/net_tstamp.h; that might
+        * be unnecessary.
+        *
+        * SIOCSHWTSTAMP was introduced in the patch that introduced
+        * linux/net_tstamp.h, so we don't bother checking whether
+        * SIOCSHWTSTAMP is defined (if your Linux system has
+        * linux/net_tstamp.h but doesn't define SIOCSHWTSTAMP, your
+        * Linux system is badly broken).
+        */
+#if defined(HAVE_LINUX_NET_TSTAMP_H) && defined(PACKET_TIMESTAMP)
+       /*
+        * If we were told to do so, ask the kernel and the driver
+        * to use hardware timestamps.
+        *
+        * Hardware timestamps are only supported with mmapped
+        * captures.
+        */
+       if (handle->opt.tstamp_type == PCAP_TSTAMP_ADAPTER ||
+           handle->opt.tstamp_type == PCAP_TSTAMP_ADAPTER_UNSYNCED) {
+               struct hwtstamp_config hwconfig;
+               struct ifreq ifr;
+               int timesource;
+
+               /*
+                * Ask for hardware time stamps on all packets,
+                * including transmitted packets.
+                */
+               memset(&hwconfig, 0, sizeof(hwconfig));
+               hwconfig.tx_type = HWTSTAMP_TX_ON;
+               hwconfig.rx_filter = HWTSTAMP_FILTER_ALL;
+
+               memset(&ifr, 0, sizeof(ifr));
+               strcpy(ifr.ifr_name, handle->opt.source);
+               ifr.ifr_data = (void *)&hwconfig;
+
+               if (ioctl(handle->fd, SIOCSHWTSTAMP, &ifr) < 0) {
+                       switch (errno) {
+
+                       case EPERM:
+                               /*
+                                * Treat this as an error, as the
+                                * user should try to run this
+                                * with the appropriate privileges -
+                                * and, if they can't, shouldn't
+                                * try requesting hardware time stamps.
+                                */
+                               *status = PCAP_ERROR_PERM_DENIED;
+                               return -1;
+
+                       case EOPNOTSUPP:
+                               /*
+                                * Treat this as a warning, as the
+                                * only way to fix the warning is to
+                                * get an adapter that supports hardware
+                                * time stamps.  We'll just fall back
+                                * on the standard host time stamps.
+                                */
+                               *status = PCAP_WARNING_TSTAMP_TYPE_NOTSUP;
+                               break;
+
+                       default:
+                               snprintf(handle->errbuf, PCAP_ERRBUF_SIZE,
+                                       "SIOCSHWTSTAMP failed: %s",
+                                       pcap_strerror(errno));
+                               *status = PCAP_ERROR;
+                               return -1;
+                       }
+               } else {
+                       /*
+                        * Well, that worked.  Now specify the type of
+                        * hardware time stamp we want for this
+                        * socket.
+                        */
+                       if (handle->opt.tstamp_type == PCAP_TSTAMP_ADAPTER) {
+                               /*
+                                * Hardware timestamp, synchronized
+                                * with the system clock.
+                                */
+                               timesource = SOF_TIMESTAMPING_SYS_HARDWARE;
+                       } else {
+                               /*
+                                * PCAP_TSTAMP_ADAPTER_UNSYNCED - hardware
+                                * timestamp, not synchronized with the
+                                * system clock.
+                                */
+                               timesource = SOF_TIMESTAMPING_RAW_HARDWARE;
+                       }
+                       if (setsockopt(handle->fd, SOL_PACKET, PACKET_TIMESTAMP,
+                               (void *)&timesource, sizeof(timesource))) {
+                               snprintf(handle->errbuf, PCAP_ERRBUF_SIZE, 
+                                       "can't set PACKET_TIMESTAMP: %s", 
+                                       pcap_strerror(errno));
+                               *status = PCAP_ERROR;
+                               return -1;
+                       }
+               }
+       }
+#endif /* HAVE_LINUX_NET_TSTAMP_H && PACKET_TIMESTAMP */
 
        /* ask the kernel to create the ring */
 retry:
+       req.tp_block_nr = req.tp_frame_nr / frames_per_block;
+
+       /* req.tp_frame_nr is requested to match frames_per_block*req.tp_block_nr */
+       req.tp_frame_nr = req.tp_block_nr * frames_per_block;
+       
        if (setsockopt(handle->fd, SOL_PACKET, PACKET_RX_RING,
                                        (void *) &req, sizeof(req))) {
-               /* try to reduce requested ring size to prevent memory failure */
                if ((errno == ENOMEM) && (req.tp_block_nr > 1)) {
-                       req.tp_frame_nr >>= 1;
-                       req.tp_block_nr = req.tp_frame_nr/frames_per_block;
+                       /*
+                        * Memory failure; try to reduce the requested ring
+                        * size.
+                        *
+                        * We used to reduce this by half -- do 5% instead.
+                        * That may result in more iterations and a longer
+                        * startup, but the user will be much happier with
+                        * the resulting buffer size.
+                        */
+                       if (req.tp_frame_nr < 20)
+                               req.tp_frame_nr -= 1;
+                       else
+                               req.tp_frame_nr -= req.tp_frame_nr/20;
                        goto retry;
                }
                if (errno == ENOPROTOOPT) {
@@ -2628,19 +3342,21 @@ retry:
                snprintf(handle->errbuf, PCAP_ERRBUF_SIZE,
                    "can't create rx ring on packet socket: %s",
                    pcap_strerror(errno));
+               *status = PCAP_ERROR;
                return -1;
        }
 
        /* memory map the rx ring */
-       ringsize = req.tp_block_nr * req.tp_block_size;
-       handle->bp = mmap(0, ringsize, PROT_READ| PROT_WRITE, MAP_SHARED, 
-                                       handle->fd, 0);
-       if (handle->bp == MAP_FAILED) {
+       handle->md.mmapbuflen = req.tp_block_nr * req.tp_block_size;
+       handle->md.mmapbuf = mmap(0, handle->md.mmapbuflen,
+           PROT_READ|PROT_WRITE, MAP_SHARED, handle->fd, 0);
+       if (handle->md.mmapbuf == MAP_FAILED) {
                snprintf(handle->errbuf, PCAP_ERRBUF_SIZE,
                    "can't mmap rx ring: %s", pcap_strerror(errno));
 
                /* clear the allocated ring on error*/
                destroy_ring(handle);
+               *status = PCAP_ERROR;
                return -1;
        }
 
@@ -2653,13 +3369,14 @@ retry:
                    pcap_strerror(errno));
 
                destroy_ring(handle);
+               *status = PCAP_ERROR;
                return -1;
        }
 
        /* fill the header ring with proper frame ptr*/
        handle->offset = 0;
        for (i=0; i<req.tp_block_nr; ++i) {
-               void *base = &handle->bp[i*req.tp_block_size];
+               void *base = &handle->md.mmapbuf[i*req.tp_block_size];
                for (j=0; j<frames_per_block; ++j, ++handle->offset) {
                        RING_GET_FRAME(handle) = base;
                        base += req.tp_frame_size;
@@ -2682,21 +3399,49 @@ destroy_ring(pcap_t *handle)
                                (void *) &req, sizeof(req));
 
        /* if ring is mapped, unmap it*/
-       if (handle->bp) {
-               /* need to re-compute the ring size */
-               unsigned frames_per_block, block_size;
-               compute_ring_block(handle->bufsize, &block_size, &frames_per_block);
-
-               /* do not perform sanity check here: we can't recover any error */
-               munmap(handle->bp, block_size * handle->cc / frames_per_block);
-               handle->bp = 0;
+       if (handle->md.mmapbuf) {
+               /* do not test for mmap failure, as we can't recover from any error */
+               munmap(handle->md.mmapbuf, handle->md.mmapbuflen);
+               handle->md.mmapbuf = NULL;
        }
 }
 
+/*
+ * Special one-shot callback, used for pcap_next() and pcap_next_ex(),
+ * for Linux mmapped capture.
+ *
+ * The problem is that pcap_next() and pcap_next_ex() expect the packet
+ * data handed to the callback to be valid after the callback returns,
+ * but pcap_read_linux_mmap() has to release that packet as soon as
+ * the callback returns (otherwise, the kernel thinks there's still
+ * at least one unprocessed packet available in the ring, so a select()
+ * will immediately return indicating that there's data to process), so,
+ * in the callback, we have to make a copy of the packet.
+ *
+ * Yes, this means that, if the capture is using the ring buffer, using
+ * pcap_next() or pcap_next_ex() requires more copies than using
+ * pcap_loop() or pcap_dispatch().  If that bothers you, don't use
+ * pcap_next() or pcap_next_ex().
+ */
+static void
+pcap_oneshot_mmap(u_char *user, const struct pcap_pkthdr *h,
+    const u_char *bytes)
+{
+       struct oneshot_userdata *sp = (struct oneshot_userdata *)user;
+
+       *sp->hdr = *h;
+       memcpy(sp->pd->md.oneshot_buffer, bytes, h->caplen);
+       *sp->pkt = sp->pd->md.oneshot_buffer;
+}
+    
 static void
 pcap_cleanup_linux_mmap( pcap_t *handle )
 {
        destroy_ring(handle);
+       if (handle->md.oneshot_buffer != NULL) {
+               free(handle->md.oneshot_buffer);
+               handle->md.oneshot_buffer = NULL;
+       }
        pcap_cleanup_linux(handle);
 }
 
@@ -2714,11 +3459,26 @@ pcap_setnonblock_mmap(pcap_t *p, int nonblock, char *errbuf)
        /* map each value to the corresponding 2's complement, to 
         * preserve the timeout value provided with pcap_set_timeout */
        if (nonblock) {
-               if (p->md.timeout > 0)
+               if (p->md.timeout >= 0) {
+                       /*
+                        * Timeout is non-negative, so we're not already
+                        * in non-blocking mode; set it to the 2's
+                        * complement, to make it negative, as an
+                        * indication that we're in non-blocking mode.
+                        */
                        p->md.timeout = p->md.timeout*-1 - 1;
-       } else 
-               if (p->md.timeout < 0)
+               }
+       } else {
+               if (p->md.timeout < 0) {
+                       /*
+                        * Timeout is negative, so we're not already
+                        * in blocking mode; reverse the previous
+                        * operation, to make the timeout non-negative
+                        * again.
+                        */
                        p->md.timeout = (p->md.timeout+1)*-1;
+               }
+       }
        return 0;
 }
 
@@ -2745,35 +3505,95 @@ pcap_get_ring_frame(pcap_t *handle, int status)
        return h.raw;
 }
 
+#ifndef POLLRDHUP
+#define POLLRDHUP 0
+#endif
+
 static int
 pcap_read_linux_mmap(pcap_t *handle, int max_packets, pcap_handler callback, 
                u_char *user)
 {
+       int timeout;
        int pkts = 0;
+       char c;
 
        /* wait for frames availability.*/
-       if ((handle->md.timeout >= 0) &&
-           !pcap_get_ring_frame(handle, TP_STATUS_USER)) {
+       if (!pcap_get_ring_frame(handle, TP_STATUS_USER)) {
                struct pollfd pollinfo;
                int ret;
 
                pollinfo.fd = handle->fd;
                pollinfo.events = POLLIN;
 
+               if (handle->md.timeout == 0)
+                       timeout = -1;   /* block forever */
+               else if (handle->md.timeout > 0)
+                       timeout = handle->md.timeout;   /* block for that amount of time */
+               else
+                       timeout = 0;    /* non-blocking mode - poll to pick up errors */
                do {
-                       /* poll() requires a negative timeout to wait forever */
-                       ret = poll(&pollinfo, 1, (handle->md.timeout > 0)?
-                                               handle->md.timeout: -1);
-                       if ((ret < 0) && (errno != EINTR)) {
+                       ret = poll(&pollinfo, 1, timeout);
+                       if (ret < 0 && errno != EINTR) {
                                snprintf(handle->errbuf, PCAP_ERRBUF_SIZE, 
-                                        "can't poll on packet socket fd %d: %d-%s",
-                                       handle->fd, errno, pcap_strerror(errno));
-                               return -1;
-                       }
+                                       "can't poll on packet socket: %s",
+                                       pcap_strerror(errno));
+                               return PCAP_ERROR;
+                       } else if (ret > 0 &&
+                           (pollinfo.revents & (POLLHUP|POLLRDHUP|POLLERR|POLLNVAL))) {
+                               /*
+                                * There's some indication other than
+                                * "you can read on this descriptor" on
+                                * the descriptor.
+                                */
+                               if (pollinfo.revents & (POLLHUP | POLLRDHUP)) {
+                                       snprintf(handle->errbuf,
+                                               PCAP_ERRBUF_SIZE,
+                                               "Hangup on packet socket");
+                                       return PCAP_ERROR;
+                               }
+                               if (pollinfo.revents & POLLERR) {
+                                       /*
+                                        * A recv() will give us the
+                                        * actual error code.
+                                        *
+                                        * XXX - make the socket non-blocking?
+                                        */
+                                       if (recv(handle->fd, &c, sizeof c,
+                                           MSG_PEEK) != -1)
+                                               continue;       /* what, no error? */
+                                       if (errno == ENETDOWN) {
+                                               /*
+                                                * The device on which we're
+                                                * capturing went away.
+                                                *
+                                                * XXX - we should really return
+                                                * PCAP_ERROR_IFACE_NOT_UP,
+                                                * but pcap_dispatch() etc.
+                                                * aren't defined to return
+                                                * that.
+                                                */
+                                               snprintf(handle->errbuf,
+                                                       PCAP_ERRBUF_SIZE,
+                                                       "The interface went down");
+                                       } else {
+                                               snprintf(handle->errbuf,
+                                                       PCAP_ERRBUF_SIZE, 
+                                                       "Error condition on packet socket: %s",
+                                                       strerror(errno));
+                                       }
+                                       return PCAP_ERROR;
+                               }
+                               if (pollinfo.revents & POLLNVAL) {
+                                       snprintf(handle->errbuf,
+                                               PCAP_ERRBUF_SIZE, 
+                                               "Invalid polling request on packet socket");
+                                       return PCAP_ERROR;
+                               }
+                       }
                        /* check for break loop condition on interrupted syscall*/
                        if (handle->break_loop) {
                                handle->break_loop = 0;
-                               return -2;
+                               return PCAP_ERROR_BREAK;
                        }
                } while (ret < 0);
        }
@@ -2940,6 +3760,18 @@ pcap_read_linux_mmap(pcap_t *handle, int max_packets, pcap_handler callback,
                }
 #endif
 
+               /*
+                * The only way to tell the kernel to cut off the
+                * packet at a snapshot length is with a filter program;
+                * if there's no filter program, the kernel won't cut
+                * the packet off.
+                *
+                * Trim the snapshot length to be no longer than the
+                * specified snapshot length.
+                */
+               if (pcaphdr.caplen > handle->snapshot)
+                       pcaphdr.caplen = handle->snapshot;
+
                /* pass the packet to the user */
                pkts++;
                callback(user, &pcaphdr, bp);
@@ -2963,7 +3795,7 @@ skip:
                /* check for break loop condition*/
                if (handle->break_loop) {
                        handle->break_loop = 0;
-                       return -2;
+                       return PCAP_ERROR_BREAK;
                }
        }
        return pkts;
@@ -2973,7 +3805,15 @@ static int
 pcap_setfilter_linux_mmap(pcap_t *handle, struct bpf_program *filter)
 {
        int n, offset;
-       int ret = pcap_setfilter_linux(handle, filter);
+       int ret;
+
+       /*
+        * Don't rewrite "ret" instructions; we don't need to, as
+        * we're not reading packets with recvmsg(), and we don't
+        * want to, as, by not rewriting them, the kernel can avoid
+        * copying extra data.
+        */
+       ret = pcap_setfilter_linux_common(handle, filter, 1);
        if (ret < 0)
                return ret;
 
@@ -4007,7 +4847,7 @@ iface_get_arptype(int fd, const char *device, char *ebuf)
 
 #ifdef SO_ATTACH_FILTER
 static int
-fix_program(pcap_t *handle, struct sock_fprog *fcode)
+fix_program(pcap_t *handle, struct sock_fprog *fcode, int is_mmapped)
 {
        size_t prog_size;
        register int i;
@@ -4040,26 +4880,33 @@ fix_program(pcap_t *handle, struct sock_fprog *fcode)
 
                case BPF_RET:
                        /*
-                        * It's a return instruction; is the snapshot
-                        * length a constant, rather than the contents
-                        * of the accumulator?
+                        * It's a return instruction; are we capturing
+                        * in memory-mapped mode?
                         */
-                       if (BPF_MODE(p->code) == BPF_K) {
+                       if (!is_mmapped) {
                                /*
-                                * Yes - if the value to be returned,
-                                * i.e. the snapshot length, is anything
-                                * other than 0, make it 65535, so that
-                                * the packet is truncated by "recvfrom()",
-                                * not by the filter.
-                                *
-                                * XXX - there's nothing we can easily do
-                                * if it's getting the value from the
-                                * accumulator; we'd have to insert
-                                * code to force non-zero values to be
-                                * 65535.
+                                * No; is the snapshot length a constant,
+                                * rather than the contents of the
+                                * accumulator?
                                 */
-                               if (p->k != 0)
-                                       p->k = 65535;
+                               if (BPF_MODE(p->code) == BPF_K) {
+                                       /*
+                                        * Yes - if the value to be returned,
+                                        * i.e. the snapshot length, is
+                                        * anything other than 0, make it
+                                        * 65535, so that the packet is
+                                        * truncated by "recvfrom()",
+                                        * not by the filter.
+                                        *
+                                        * XXX - there's nothing we can
+                                        * easily do if it's getting the
+                                        * value from the accumulator; we'd
+                                        * have to insert code to force
+                                        * non-zero values to be 65535.
+                                        */
+                                       if (p->k != 0)
+                                               p->k = 65535;
+                               }
                        }
                        break;